Verified intelligence · Expert-reviewed from public records, platform data, primary sources, and private research beyond AI search.

MODELSFinancial servicesGLOBAL

AI models for financial services

Financial AI models should be compared by accepted output: source authority, model-risk evidence, deployment controls, benchmark caveats, workflow permission and full cost.

The colophon

Landscape reviewed: 2026-06-19
Latest revision: 2026-06-19
Kinetics: Finance model approval has shifted from choosing an LLM to proving the evidence route can survive review. AI Act consultations, GPAI enforcement and finance benchmark refreshes can change credit, insurance, filing-RAG and agent routes before renewal.
To watch: 23 June 2026: European Commission high-risk AI consultation closes; credit and insurance route classification may move ¹⁷⁰.
2 August 2026: Commission GPAI enforcement powers begin; model documentation, transparency and provider-diligence evidence may change ¹⁶⁹.
Known unknowns: Comparable public benchmarks, downstream parser-to-answer tests, longitudinal product telemetry, permissioned-corpus datasets, AML and fraud false-negative datasets, customer-outcome studies, model-change regression logs, concentration-event data and audit-export evidence remain missing for financial model routes.

Contents

Executive summary
Decision map
What changed
Model economics
Throughput and rate limits
Model lanes
Use-case route selection
Benchmarks
Source evidence path
Regulatory gates
Deployment controls
Agent controls
Eval design
Implementation plan
Risk controls
Cross-border limits
Model shortlist
Implementation cases
Operational signals
Expert watchlist
Vendor directory
Known unknowns
Source ledger
Glossary

▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒Executive summary

The best financial AI system is not the cleverest model. It is the control stack that produces a reviewable answer from approved sources under the right control regime. A strong frontier model with a stale filing, a broken table or an unapproved customer-data path is still a bad financial answer. Use frontier models where judgement, drafting, synthesis and multi-step reasoning matter. Use cheaper, private or open-weight models where throughput, sovereignty, latency or custom controls dominate. Use OCR, document intelligence, embeddings and rerankers before asking any model to answer a question over filings, KYC packs, statements, credit files or policy documents.

FinanceBench gives the warning in numbers: its original retrieval baseline incorrectly answered or refused 81% of sampled financial QA questions ²⁰. That result is old by model standards, but the lesson has aged well. Retrieval, tables, dates and evidence chains can break a finance workflow before generation begins.

Rank model lanes before ranking model names. Investment research, retail credit, AML investigation, insurance pricing, treasury analysis, month-end close and customer communications need different routes.

The procurement object is therefore the approved route: data class, source path, parser, retriever, generator, hosting route, reviewer, logs, contract terms and reapproval trigger.

Related report

AI models for banking, capital markets, and insurance

Use the regulated-finance report for banking, capital-markets and insurance route approval, customer-impact gates, securities supervision, insurance controls and demand-signal split timing.

Related report

File processing providers for financial AI

Detailed due diligence for financial OCR, document AI, parser routes, bank statements, filings, invoices, IDs, claims files, controls, pricing and parser evals.

Board-level verdict

Question	Answer	Buyer move
Which model wins?	The winning route changes by task, data class and approval gate.	Approve model lanes, not model names.
Where is the main failure?	Wrong source, lost table, stale date, weak calculation or unsupported claim.	Test source capture, retrieval, calculation and citation before model prose.
Where should frontier models lead?	Research synthesis, drafting, reasoning, workflow orchestration and exception analysis.	Pair them with verified sources, deterministic calculations and reviewer gates.
Where should cheaper or private models lead?	Classification, extraction, volume summarization, internal search and low-risk finance operations.	Use them where measured quality clears the workflow and data controls are easier.

Approve outputs by work type and risk class. Financial AI buying fails when one approval covers unlike work. The first split is not model brand. It is whether the system informs an employee, drafts a customer-facing artefact, influences a regulated decision or acts through a tool.

A filing-summary assistant can tolerate different controls from a credit-scoring model. An AML investigator's summarizer is not the same as an agent that files a suspicious-activity draft. A portfolio-research copilot is not the same as an execution system.

Build the route register around intended use. Each route should name the work, source corpus, data class, approved model lane, hosting route, tool permissions, reviewer, logging standard, evidence export and reapproval trigger.

The route register becomes the bridge between model-risk, technology-risk, compliance, legal, security and business owners.

Start with the consequence of the output.

Does the output affect credit, insurance pricing, advice, trading, surveillance or customer treatment?
Yes
Regulated decision route.
Next: Run legal, compliance, model-risk and fairness gates.
No
Continue.
Next: Check data class and tool permissions.
Does the route use non-public customer, account, trading, portfolio or employee data?
Yes
Controlled data path.
Next: Approve retention, region, training use, access logs and subprocessors.
No
Continue.
Next: Check whether the system can act.
Can the system write, submit, send, approve, trade, update records or mutate spreadsheets?
Yes
Agent route.
Next: Use tool permissions, dry-run mode, action logs and human approval.
No
Assistant route.
Next: Run task evals and review logging before scale.

Minimum approval register

Field	Why it matters	Bad entry	Good entry
Use case	Regulation follows intended use.	Finance AI.	Retail-credit adverse-action explanation draft.
Source path	The model cannot reason over missing evidence.	Documents.	SEC filings, policy KB, credit file, account events and reviewer notes.
Data class	Retention, region and access controls change.	Internal data.	Public filings plus confidential customer account data.
Model lane	Capability, cost and approval differ.	GPT.	Frontier API for reasoning; private classifier for triage; reranker for retrieval.
Reviewer	Accountability needs a named human role.	Human review.	Credit analyst signs off reason codes before customer notice.
Reapproval trigger	Model drift is a governance event.	Annual review.	Model version, prompt, data source, threshold, tool permission or vendor term change.

Financial AI moved from productivity demos to output governance. The market has stopped asking whether AI can summarize finance work. It is asking whether AI can touch regulated decisions, customer records, private data, market infrastructure and financial crime workflows without losing evidence. Three changes forced the shift. First, model capability is now strong enough to enter real workflows. Second, domain and source access matter more because financial work depends on current filings, prices, entity identities, sanctions records, policy documents and private ledgers. Third, regulators are turning AI into an inventory, governance, lifecycle and resilience question.

BloombergGPT and FinGPT made the same point from different directions: finance-specific data matters. BloombergGPT used privileged financial corpora; FinGPT argued for open financial-data pipelines and adaptation. Neither route removes the need for current sources, permissions and private evals ⁴⁹⁵⁰.

The institutional consequence is a new approval object: the model route. A bank can approve a model for analyst research and block it for credit decisions. An insurer can approve document extraction for claims support and block autonomous pricing. An asset manager can approve filing RAG and still prohibit unreviewed client communications.

The old pilot question was, 'Does the answer look good?' The production question is, 'Can we prove where it came from, who approved it, what it changed, and why it should still be trusted after the next model update?'

From pilot to production

Old question	New question	Finance implication
Which model is smartest?	Which approved route clears the task?	Model, data, tools and reviewer move together.
Can it summarize a filing?	Can it cite the right filing, table, period and calculation?	Source integrity beats prose quality.
Can employees use chat?	Which records, customers and jurisdictions enter the prompt?	Data class decides deployment.
Can the agent save time?	Can the agent act, and who approves the action?	Tool permissions become controls.
Is the vendor secure?	Can the institution exit, audit and survive concentration?	Third-party resilience is part of model selection.

Price the finished financial task, not the token. The right economic unit is not a token. It is a signed-off answer, cell, memo, alert or customer communication. Financial work has a verification tail. A model answer often needs source inspection, recalculation, variance checks, reviewer notes, policy fit, approval and record retention.

That tail can dominate model price. A cheaper model that sends reviewers to the filing five extra times loses to a more expensive model that returns the right table, period and calculation with a usable citation.

Artificial Analysis-style comparisons are useful because they put price beside intelligence, output speed, latency and context ¹⁶⁷¹⁶⁸. Financial buyers need one more column: accepted output. The unit is not a million tokens. It is a reviewed answer, reconciled cell, signed-off memo, closed alert or archived customer communication.

Silicon Valley Bank's State of Enterprise Software survey shows why that denominator matters. About 48% of respondents reported 25%+ employee-productivity improvement from AI, but half said revenue-growth attribution was hard to measure and one-quarter put the revenue lift below 10% ²⁰⁰. Treat productivity and revenue as separate economic claims.

FinanceBench and agent benchmarks move evaluation away from general language skill and toward finance-specific failure modes: finding the right evidence, doing the arithmetic, preserving dates, submitting an answer and surviving expert review.

Procurement should therefore run a cost-per-accepted-output eval. Count false answers, refusals, missing-source reports, reviewer minutes, reruns, latency, token spend, tool calls and audit-export quality.

Pricing is moving for the same reason. Silicon Valley Bank reports that nearly 40% of surveyed enterprise-software companies are primarily seat-priced today, but only 26% plan to stay there; customer demand and AI/compute economics are explicit pricing drivers ²⁰⁰. Do not let the vendor's billing unit become the buyer's ROI unit.

Token price misses finance work

Two ways to read model cost. Left: the sticker price, what the provider charges per million tokens blended at the standard 3:1 input-to-output ratio. Right: the bill for running the full Artificial Analysis Intelligence Index on that model. Data checked on 31 May 2026. Across the six bands shown, the median cheapest token price has fallen about 120x since March 2024; the median benchmark bill has fallen about 25x. The two prices come apart because models use different amounts of input, reasoning and answer tokens. Cheap tokens do not always mean a cheap finished job, and pricey tokens do not always mean a pricey one. Each line tracks the cheapest model available in a given intelligence band; lower is cheaper. Data from Artificial Analysis ¹⁶⁷; intelligence bands after a16z ¹⁶⁸.

Sticker price is only the first cost signal. Lane choice determines whether the same workload is interactive, batched, cached, flexed, provisioned or delayed, which changes both the bill and the review burden.

The best discount for filing corpora, KYC file drops, claims batches, policy archives and private eval reruns is often patience. Long extraction packets, eval reruns, classification sweeps and archive enrichment do not need a user waiting in the browser. Move that work to async or lower-priority lanes, keep deadline work on the fast path, and compare cost per financial output.

The discount lane is part of the route

Lane	How it works	Good for	Watch
OpenAI Batch API	Async jobs use a separate batch lane with a lower price and a 24-hour completion target. ¹⁷⁵	Large document extraction, eval reruns, classification, enrichment and non-urgent archive work.	Not an interactive path; queue time, file packaging, partial failures and output reconciliation become part of the route.
OpenAI Flex processing	Eligible calls use lower-priority capacity and are charged at Batch API rates for supported models. ¹⁸⁵	Long-document async work where cost matters more than immediate availability.	Treat it as a cheaper lane with lower priority, not as the path for a signing call, filing deadline or live customer answer.
Anthropic Message Batches	Bulk async requests can be priced below the standard synchronous lane and can be combined with prompt caching. ¹⁸⁷¹⁸⁸	Large non-interactive packs: extraction, summarization, evals, source tagging and overnight review queues.	Batch queue behaviour and cache checkpoint design decide whether the savings survive production.
Gemini / Vertex AI batch	Vertex batch prediction is the non-real-time lane; Google pricing separates standard, batch and cached-input treatment. ¹⁸⁹¹⁹¹	Bulk document analysis, classification and back-office model calls that do not need immediate response.	Batch mode, provisioned throughput and cached context are separate mechanisms; do not assume one discount buys the others.
Amazon Bedrock batch inference	Bedrock batch inference runs async jobs over S3 inputs, with discounted batch pricing for supported models. ¹⁹²¹⁹⁴	Cloud-controlled bulk inference where the input/output artefacts can live in the approved AWS path.	Prompt caching and batch inference are not universally combinable; model, region and feature support decide the real lane.
Groq Batch and Flex	Groq's Batch API is a discounted async lane; Flex processing trades immediacy for access to available capacity rather than a lower price. ¹⁹⁶¹⁹⁷	Throughput-sensitive work where the buyer can separate background processing from live assistant use.	Low published developer limits can still make a clever route unusable until the account, plan or enterprise path changes.
Usage tiers and quota requests	Spend tiers, usage tiers, quota requests and account-specific limits can change the ceiling before any model changes. ¹⁷⁴¹⁷⁶¹⁷⁷¹⁷⁸¹⁸⁰¹⁸²	New projects that need to absorb a sudden file drop, launch backlog or review queue.	Tier language is not capacity evidence. Check the live project, account, subscription, region and deployment type.
Provisioned or reserved capacity	Azure PTUs and reservations, Vertex Provisioned Throughput and Bedrock Provisioned Throughput buy predictability rather than a smarter model. ¹⁷⁹¹⁸¹¹⁸⁴	Recurring volume, deadline-sensitive work and routes where shared-capacity 429s would be business failures.	Commitments, idle capacity, regional availability, spillover rules and reservation terms belong in the price model.

Prompt caching is easy to misunderstand. It is not model memory and it does not cache the answer. It reuses the provider's work from the repeated beginning of a prompt: instructions, tool schemas, output schema, playbook text, matter taxonomy or source bundle. If the next request starts the same way, the provider can bill and process that repeated prefix differently. If the route prepends a timestamp, request id or one-off user sentence before the stable material, the cache may miss.

Design the prompt like a production artefact. Put the durable, reusable context first; put the unique question last; record cache hits in the financial model route register. Cache savings are real only when the route keeps a stable prefix and still produces financial output.

How to make caching work

Design rule	Why it matters	Implementation
Put stable material first	Most cache mechanisms reward an identical or explicitly marked prefix.	Start with system instructions, tool schemas, output schema, jurisdiction rules, playbook text and reusable source packs.
Put variable material last	A unique timestamp, request id, user aside or matter-specific fact at the top can break a prefix match.	Append the specific question, run id, new facts, reviewer note and requested output after the cacheable prefix.
Keep tools and schemas stable	Tool definitions, JSON schemas and message ordering can be part of the cache key.	Version schemas deliberately; avoid rebuilding tool lists, whitespace, examples or column order on every call.
Warm the cache before fan-out	Parallel extraction can waste money if every worker pays to prefill the same long context.	Send one cacheable request for the source pack or playbook, then fan out clause, entity, citation or issue tasks.
Cache source packs, not shortcuts	The cache should preserve approved context, not hide missing retrieval or stale policy.	Attach source version, route version and expiry; rebuild the prefix when authority, playbook or source set changes.
Measure cached-token economics	A cache discount that never hits is procurement fiction.	Log cache reads, writes, cached tokens, hit rate, latency and accepted-output rate beside the normal token bill.
Respect the data boundary	Caching is still provider processing under the approved route.	Cache only context that is allowed inside the approved data boundary. Caching is a billing and latency control, not permission to reuse one client's facts in another client's work.

Provider caching is not one thing

Provider	Mechanic	Buyer read
OpenAI	Automatic prompt caching applies to repeated prompt prefixes above the documented threshold; usage surfaces cached tokens, and prompt-cache keys can help route related work. ¹⁸⁶	Keep the first part of the prompt identical: instructions, tools, schema and reusable source context before matter-specific text.
Anthropic	Claude uses explicit cache-control breakpoints over tools, system and message content, with separate cache write and read economics and documented cache durations. ¹⁸⁸	Design cache checkpoints deliberately. The best target is the long shared context, not the short one-off question.
Gemini / Vertex AI	Vertex supports context caching, including explicit cached content and implicit prefix caching; pricing separates cached input from ordinary input and batch treatment. ¹⁸⁹¹⁹⁰	Check whether the route uses explicit cache objects, implicit prefix reuse or batch prediction; each has different cost and expiry behaviour.
Amazon Bedrock	Bedrock prompt caching uses cache points with time-to-live controls and separate cache pricing, while batch inference has its own feature constraints. ¹⁹²¹⁹³¹⁹⁴	Do not assume a cached interactive prompt can simply move into batch. Confirm feature support by model and region.
Groq	Groq prompt caching is automatic for exact repeated prefixes above the documented threshold, with discounted cached tokens and separate treatment in rate-limit accounting. ¹⁹⁵	The route must keep a stable prefix. This matters most when rate limits are low and large token packets would otherwise queue.

The practical question is whether the buyer can deliberately choose the cheap lane without breaking the work. A cheaper async run is useful for overnight extraction. It is dangerous for an answer needed during a signing call. A cache hit is useful when the prefix is stable. It is imaginary when every prompt is rebuilt from scratch.

What to prove before calling the route cheap

Test	Proof	Pass condition
Interactive versus async split	Tag each workflow step as user-blocking, same-day, overnight or archive work.	Only non-blocking work moves to batch, flex or lower-priority processing.
Batch economics	Run the same packet through interactive and batch lanes, including failed files, retries, queue time and output handling.	The cheaper lane still produces accepted financial output inside the work window.
Cache hit rate	Log cache reads, writes, cached tokens or provider-equivalent fields by route version.	The actual hit rate is high enough to change the finished-task bill, not just the theoretical token price.
Tier and capacity path	Capture account tier, quota, regional capacity, provisioned-capacity option and support escalation.	The buyer knows which spend, quota or reserved-capacity lever unlocks the next workload level.
Accepted-output denominator	Combine model spend, queue time, cache hit rate, reviewer minutes, failed runs and rework.	The route is cheaper per financial output, not merely cheaper per million tokens.

Cost per accepted output

Cost item	Token view	Accepted-output view	Measurement
Retrieval miss	Cheap answer.	Expensive wrong answer.	Correct-source recall and no-answer rate.
Table loss	Normal prompt.	Review fails because the key row vanished.	Table-cell preservation and source-span export.
Calculation error	Low inference cost.	Analyst recalculates or rejects.	Formula accuracy, intermediate values and reviewer minutes.
Long context	More tokens.	Better evidence, higher latency and higher cost.	Accepted answer rate per minute and dollar.
Productivity claim	Low model bill.	Savings are real only when accepted output, reviewer time or revenue lift is measured.	Accepted outputs per reviewer hour, baseline labour cost and revenue-attribution confidence.
Refusal	No hallucination.	Good if evidence is missing; broken if evidence exists.	Correct refusal, false refusal and missing-source explanation.
Audit export	Not counted.	Required for exams, disputes and model validation.	Reproducible inputs, source IDs, model version and reviewer decision.

How to price an accepted output

Workflow	Why token price misleads	Buyer test
Public filing Q&A	The answer lives in tables, notes, periods and restatements; a cheap answer is expensive if the source is wrong.	Measure correct-source recall, source-span export, no-answer behaviour and reviewer minutes.
Policy search	The model bill is small; stale retrieval, missing policy exceptions and archive gaps create the real cost.	Track current-source recall, correct refusal and audit-export completeness.
Finance productivity route	AI productivity gains may appear before revenue attribution; seats, usage and compute charges can all misstate the accepted-output economics.	Measure accepted outputs per reviewer hour, backfill avoided, reruns and confidence in revenue attribution.
Credit decision support	Customer outcome, adverse-action reasons and fairness checks move cost into validation and review.	Measure reason-code accuracy, prohibited-proxy checks, sign-off time and archive record quality.
AML case support	False negatives, entity linkage and investigation history matter more than fluent summaries.	Measure missed-link rate, investigator disposition, escalation quality and reconstructed evidence packet.

Rate limits decide whether the model can survive a live financial workload. Rate limits are production capacity, not a nuisance error. A model can be the best reasoning route and still fail a filing corpus, KYC file drop, credit queue, policy archive, claims batch or analyst research pack if the account cannot move enough tokens through it. The limiting number may be requests per minute, tokens per minute, input tokens per minute, output tokens per minute, batch-queue tokens, spend tier, project quota, regional capacity or provisioned-unit size.

This matters most when work arrives in a lump. A one-off demo does not prove that a route can process the approved workload by the production deadline. Low limits turn intelligence into waiting time, retries, partial runs and fallback risk.

Record throughput in the financial model route register. The approval question is not "is the model smart?" It is "can this approved account, region, deployment type, batch lane, provisioned capacity and fallback process the actual workload before the business deadline?"

The first proof is tier math: the same paid API can clear a large packet quickly at the top tier and stall at the entry tier. Treat the tier row as a production input, not procurement small print.

The second proof is contract capacity. Azure PTUs and reservations, Vertex Provisioned Throughput, Bedrock Provisioned Throughput, cloud quota requests, global or cross-region routing, Batch APIs, prompt caching and fallback routes all change both cost and operational risk.

Think of the approved file packet like a large print job waiting at a shared printer. The model may be ready, but the paid API tier decides how wide the paper feed is. The 750K-token packet is the scenario load for this stress test. At OpenAI's published 500K TPM entry tier, that packet waits about 1.5 minutes just to enter the lane; at the published 40M TPM top tier, it clears in about a second ¹⁷⁴. The bars show that intake queue, not model intelligence or total job time.

How long a large file packet waits

Official provider documentation checked on 31 May 2026. Bars show the minimum intake wait for one 750K-token packet, not total job time or model quality. Shorter is better. Within each provider family, the filled row is the highest public tier and the dotted row is the entry comparable public tier. Groq and Bedrock are not charted because their public docs do not expose one comparable production TPM ceiling for this route-level wait calculation. Minimum wait ignores output, retries, queueing and review.

Lowest and top tiers are different products

Provider	Named public route	Published RPM and token cap	Approval read
OpenAI	GPT-5.5 API	Lowest paid: Tier 1, 500 RPM / 500K TPM. Highest public: Tier 5, 15K RPM / 40M TPM. ¹⁷⁴	Use the live usage tier, long-context cap and shared-family cap. Do not use response-header samples as route evidence.
Anthropic	Claude Opus 4.x API	Lowest paid: Tier 1, 50 RPM / 500K ITPM / 80K OTPM. Highest public: Tier 4, 4K RPM / 10M ITPM / 800K OTPM. ¹⁷⁶	Input and output caps are separate; Opus 4.x traffic shares one family limit and custom limits require approval.
Groq	Groq API production route	No single public production TPM ceiling. Public docs expose plan/model rate-limit rows and say higher limits are available for select workloads and enterprise use cases. ¹⁷⁷	Do not approve from public base rows. Capture the live account or enterprise contract limit before using Groq as a production lane.
Azure OpenAI	gpt-5.5 GlobalStandard	Lowest nonzero paid: Tier 5, 10K RPM / 10M TPM. Highest public: Tier 6, 15K RPM / 15M TPM. Lower public tiers list 0 / 0. ¹⁷⁸	Quota is per model, deployment type, subscription and region; quota is not proof of deployable capacity.
Google Vertex AI	Gemini 2.5 Pro Standard PayGo	Lowest paid: Tier 1, 500K TPM. Highest public: Tier 3, 2M TPM. System cap: 30K RPM per model per region. ¹⁸⁰	The baseline is org-level shared capacity on the global endpoint; traffic smoothing and Provisioned Throughput still matter.
Amazon Bedrock	On-demand and Provisioned Throughput	No single comparable public default TPM row. AWS exposes model-specific token/RPM quota families and provisioned MUs. ¹⁸²	Check Service Quotas for the exact model and region; use cross-Region inference or Provisioned Throughput for the production lane.

Rows are official paid API tier or quota rows captured on 31 May 2026. The comparison rule is lowest public paid tier plus highest public paid tier where both are published; otherwise the row names the missing public production ceiling.

The wait calculation answers only one question: how hard the public token cap bites before any workaround. The buyer still has to ask which lever is actually available for the approved account, region and data class.

Capacity levers change the contract

Route	Capacity levers to check	Procurement caveat
OpenAI APIs	Usage tiers, org/project/model limits, response headers, Batch API, priority processing.	Exact limits remain account-specific.
Claude API	Spend tiers, RPM/ITPM/OTPM, prompt caching, Message Batches, Priority/custom limits.	Acceleration limits can still punish sudden spikes.
Groq API	Plan upgrade, per-model limits, cached-token treatment, Batch/Flex or enterprise workload path.	Public base rows are not a production ceiling; the buyer needs live account or enterprise limit evidence.
Azure OpenAI	Quota tiers, model/deployment/region pools, multi-region deployment, PTUs, spillover, reservations.	PTU quota does not guarantee deployable regional capacity.
Gemini / Vertex AI	Standard PayGo tiers, global endpoint, traffic smoothing, custom tier, Provisioned Throughput, caching or batch lanes.	PayGo 429s can reflect shared-resource contention, not a fixed quota breach.
Amazon Bedrock	Service quotas, quota increase request, cross-Region inference, Provisioned Throughput, commitment terms.	Model, region and account quotas decide the real production lane.

Documented capacity mechanisms are not a quality score. The useful question is which lever the buyer can actually buy, reserve, request, batch or route around for the approved workload.

Once the tier and capacity levers are visible, approval should move from provider comparison to failure design. Name the first limiter, the owner of the workaround and the proof that the fallback can scale.

How rate limits break production work

Bottleneck	Failure mode	Buyer move
Synchronous TPM, ITPM or OTPM	Large files queue behind the token budget even when the model can reason over them.	Measure tokens per accepted job; compare input, output and cached-token treatment separately.
RPM and burst shape	A filing drop, diligence batch or claims surge can hit request caps before token caps.	Run burst tests with realistic concurrency; smooth traffic and batch small tasks when latency allows.
Spend tier or usage tier	A new account or project may be stuck below the limits needed for a real matter.	Check the live account tier before procurement approval; pre-warm tiers or contract higher limits.
Batch queue	Offline work can move out of the synchronous lane, but turnaround and queue limits decide whether it fits the deadline.	Separate interactive lawyer work from overnight extraction, eval and classification runs.
Cloud quota and regional capacity	Quota by model, deployment type, region, project or subscription can block scale even after the model is approved.	Request quota early; verify region capacity and data-residency constraints before the launch window.
Provisioned or reserved capacity	Reserved units solve predictability but introduce commitments, idle capacity and capacity-allocation risk.	Size PTUs, GSUs, MUs or token reservations from real prompt and output distributions.
Fallback route	The backup may have different rate limits, data terms, quality, region, logging or source access.	Approve fallback by data class and workflow, not just by provider name.

Before production, make the route pass concrete tests. A limit that is acceptable in a demo is not approved until the live account, region, batch lane and fallback path have been exercised together.

What to prove before the route scales

Test	Proof	Pass condition
Large-file packet	Run a 750K-token packet plus expected output under target concurrency.	The route finishes inside the work window without 429 cascades, missing files or unreviewable partial output.
Deadline cluster	Replay the busiest hour of the production deadline window.	Primary and fallback routes stay inside latency and data-boundary limits.
Tier and quota evidence	Capture account tier, project limits, model limits, cloud quota, regional capacity and support escalation path.	Procurement can name the exact limiter that would fail first.
Reserved-capacity sizing	Convert request shape into provisioned units using average input, output, cache hit rate and RPM.	The buyer knows commitment cost, idle-capacity risk, spillover behaviour and scale-up constraints.
Fallback legality	Test whether fallback changes region, subprocessor, retention, source licence, audit log or model quality.	Fallback is allowed for that data class and reviewer instruction.

The finance shortlist should mix generators, retrievers, parsers and private models. A finance model stack has more than one model. The generator is visible, but the supporting models often decide whether the answer is usable. OCR sees the filing. Embeddings find candidate evidence. Rerankers choose which evidence wins. A classifier triages the alert. A frontier model reasons over the packet. A deterministic calculator checks the numbers.

Treat those as separate approval objects. A bank may approve a frontier model for analyst drafting and a different embedding model for retrieval, while keeping a self-hosted classifier for sensitive transaction triage.

Open-weight and self-hosted lanes matter when data cannot leave a controlled environment or when volume makes API economics unattractive. They also move security patching, model monitoring, red-teaming, infrastructure and validation into the buyer's house.

Cloud model platforms matter when the buyer already has IAM, logging, data-residency and procurement controls in that cloud. Direct APIs matter when the latest frontier capability beats cloud standardization. Productivity-suite routes matter for low-risk finance operations, but they still need data and recordkeeping checks.

Model lane map

Lane	Controls	Advantaged	Buyer read
Frontier reasoning models	Prompt, source packet, tool limits, output schema, human review and regression evals.	Investment research, credit memo drafting, policy interpretation, exception analysis and multi-step synthesis.	Approve for judgement support, not autonomous regulated decisions.
Open-weight and private models	Hosting, access control, patching, model card, red-team results, data lineage and monitoring.	Sensitive data, high-volume classification, sovereign routes and custom internal taxonomies.	Useful when control or unit economics beat frontier convenience.
Embeddings and rerankers	Corpus fit, multilingual fit, recall, precision, stale-source handling and retrieval logs.	Policy search, filing QA, customer-document search, knowledge bases and evidence packs.	They decide what the generator sees. Validate separately.
OCR and document intelligence	Table structure, handwriting, selection marks, page references, confidence and low-quality flags.	Statements, invoices, filings, KYC packs, claims, confirmations and scanned correspondence.	Parse before the model bakeoff.
Deterministic tools	Formula tests, data provenance, rounding policy, reconciliation logs and reviewer override.	Ratios, covenants, cash forecasts, P&L bridges, tax calculations and fee checks.	Let models propose; make tools calculate.

Data sensitivity chooses the deployment lane

Capability

Control plane

Data residency

Ops burden

Best use

Direct frontier APIUse when the newest reasoning model changes the work product.

High

Contract-led

Variable

Low

Hard synthesisResearch, drafting, analysis and tool planning.

Cloud model platformUse when IAM, logging, procurement and data controls already sit in one cloud.

High

Medium

Governed buildEnterprise RAG, managed agents and model platform workflows.

Self-hosted open weightUse when sensitive corpora or volume justify owning the stack.

Variable

High

Private scaleSensitive classification, sovereign workloads and high-volume internal routes.

Productivity-suite AIUse close to daily finance work, with narrow action permissions.

Medium

Low

Adjacent workPolicy search, spreadsheet drafting, email and presentation support.

LowMediumHighGate

The deployment lane is a control choice. Match it to data sensitivity, workflow criticality, review burden and the institution's operating model.

The same model should not receive the same powers everywhere. Finance is a bundle of workflows with different blast radii. The industry label hides too much. An asset manager summarizing earnings calls, a bank explaining an adverse credit action, an insurer pricing health cover, a broker-dealer supervising communications, a payments firm triaging fraud and a controller closing the books do not need the same model route.

Separate the workflow before comparing vendors. The right model lane depends on whether the task needs reasoning, table extraction, time-series features, identity proof, fairness evidence, market-data recency, regulator-ready logs or action permissions.

For most institutions, the first scalable portfolio contains six lanes: analyst productivity, filing and research RAG, document extraction, risk and compliance support, customer-facing controlled generation, and agentic finance operations.

Keep regulated decisioning and autonomous market action behind stricter gates until the system can prove source support, limits, audit export and human accountability.

Financial workflow controls

Workflow	Model lane	Approval	Fatal failure
Investment research	Frontier reasoning plus filing RAG, market-data tools and citation checks.	Conditional.	Fabricated metric, wrong period, stale filing or unsupported investment claim.
Credit underwriting support	Private feature models, explainable reason-code tools and controlled LLM drafting.	High-risk gate.	Unexplainable adverse-action reason or discriminatory proxy.
AML and sanctions investigation	Transaction anomaly models, entity resolution, case summarizer and investigator approval.	Conditional.	False negative, hallucinated linkage or missing audit trail.
Fraud detection	Graph, behavioural and anomaly models plus supervised alert review.	Conditional.	Silent missed fraud pattern or customer-impacting false positive without appeal path.
Insurance underwriting and claims	Document AI, risk models, claims triage and controlled explanations.	High-risk gate for covered life and health pricing.	Opaque pricing factor or unreviewed customer-impacting decision.
Finance operations	Productivity-suite AI, private reconciliation models and spreadsheet-aware checks.	Good early lane.	Hidden formula mutation, unsupported journal entry or unlogged approval.
Customer communications	Controlled generation with policy retrieval, banned-claim rules and supervision.	Conduct gate.	Misleading advice, promissory performance claim or missing record.
Market surveillance	Pattern detection, communications analysis and human compliance review.	Conditional.	Missed manipulation signal or unsupported escalation.

Acceptance tests by workflow

Workflow	Acceptance test	Minimum artefact
Research	Answer from the correct filing, table and period with calculation trace.	Citations, extracted values, formula and reviewer decision.
Credit	Produce specific, accurate and reviewable reasons without prohibited proxies.	Feature lineage, reason codes, fairness check and human sign-off.
AML	Summarize evidence without missing linked entities, transaction hops or adverse media.	Case graph, source links, confidence, investigator notes and escalation trail.
Fraud	Balance recall, false positives and customer friction under drift.	Threshold history, outcome labels, override log and drift report.
Close	Reconcile figures while preserving formulas, hidden tabs and approvals.	Workbook diff, source system IDs, exceptions and approver log.
Client communication	Generate compliant, balanced and source-supported copy.	Policy source, approved template, archive record and supervisor decision.

Public scores show where to test, not what to buy. A finance benchmark is a failure probe. FinanceBench probes open-book QA over public-company filings ²⁰. FinBen broadens the field across extraction, text analysis, QA, generation, risk management, forecasting and decision-making ²²²². TAT-QA and ConvFinQA stress tables, text and numerical reasoning ²³²⁴. Finance-agent benchmarks move closer to workflow, retrieval and tool use ²⁵.

Rogo's BigFinanceBench adds workflow-grounded rubric scoring and model-routing evidence ¹⁹⁹. Harvey LAB adds a legal-specific trajectory lens that finance evals can translate but should not copy mechanically: trace source coverage, validation, revision, retrieval and tool fan-out rather than scoring only the final answer ¹⁹⁸.

GDP.pdf adds a cross-domain professional-document stress test, including finance PDFs ²⁰². In the 6 June 2026 GDP.pdf leaderboard snapshot, the top macro row is only 30% ²⁰¹. Reducto's vendor-reported rerun says adding structured parse context moved aggregate macro accuracy from 21.7% to 30.7%, a 9.0 percentage-point lift ²⁰⁴.

None of them sees the buyer's private portfolio data, permissions, customer records, policy exceptions, current market data, examiner expectations or reviewer minutes.

Use public benchmarks to choose the private eval shape. If a model fails public filing QA, do not trust it with private credit packs. If it passes filing QA, still test stale filings, restatements, earnings-call corrections, hidden spreadsheet rows, ambiguous time periods and no-answer behaviour.

Do not combine unrelated benchmarks into a single league table. A model that performs well on extraction may still fail forecasting, advice, credit explanations or action logs.

Current PDF answers still fail often

GDP.pdf

Macro accuracy: share of professional PDF tasks where every rubric criterion passed.

1Claude Fable 5 / Mythos 530%
2GPT-5.5 (xHigh reasoning)25%
3Claude Opus 4.8 (Adaptive Max)23%
4Claude Opus 4.7 (Adaptive Max)21%
5Gemini 3.1 Pro17%

Surge AI

GDP.pdf leaderboard snapshot from 6 June 2026 ²⁰¹. Only same-source ranked rows are shown here; method papers, datasets and workflow benchmarks stay in the scope table so unlike scores are not merged.

Benchmarks see fragments of finance work

Source	Signal	Limit
FinanceBench	Open-book financial QA: 10,231 public-company questions with answers and evidence strings ²⁰. Best used to test filing retrieval, table capture, period handling, calculation and refusal behaviour.	Minimum performance standard over public filings, not a private-bank or live-market approval test.
BigFinanceBench	Workflow-grounded finance agents: Rogo-reported May 2026 benchmark with 928 questions, 15,656 rubric criteria and 36,241 weighted points across ten frontier systems ¹⁹⁹. Best used to test workflow-specific routing, derivation quality and rubric-scored financial reasoning.	Vendor-led benchmark announcement; use as workflow-routing evidence, not independent procurement truth.
GDP.pdf	Professional document QA: 100 real-world prompts and PDFs across ten domains, including finance ²⁰²; top current macro row is 30% ²⁰¹. Best used to force parser/source-path tests before comparing answer models.	Cross-domain PDF benchmark, not a regulated-finance approval test; buyer evals still need private records, source entitlements, reviewer gates and audit export.
FinBen	Holistic finance benchmark: Broad finance benchmark covering extraction, text analysis, QA, generation, risk management, forecasting and decision-making ²²²². Best used to identify task families where a model lane is weak.	Broad taxonomy, but not a substitute for a product route eval.
TAT-QA	Table and text QA: Hybrid financial reports requiring numerical reasoning over tables and text ²³. Best used to keep table extraction and numerical reasoning in the test set.	Older than current frontier models, but the table-plus-text failure remains central.
ConvFinQA	Conversational numerical QA: Multi-turn finance questions that require chained numerical reasoning ²⁴. Best used to test whether a route carries intermediate calculations across question turns.	Conversational reasoning is useful, but production finance conversations also need permissions, source age and archive records.
Finance Agent	Agentic finance workflow: Multi-step tasks requiring search, calculations, financial expertise and final answer submission ²⁵. Best used to probe tool use, search, calculations and agent recovery.	More workflow-like than QA, but still public and harness-specific.
FinanceQA	Investment-style numerical analysis: Hedge-fund, private-equity and investment-banking style numerical work ¹²³. Best used to test assumptions, hand-spread metrics and multi-step financial analysis.	Useful analyst-work proxy, but it does not validate a firm's own models, assumptions or investment policy.
FinQA	Numerical reasoning over reports: Program-style financial QA over reports ¹²⁴. Best used as a private arithmetic and formula-trace template.	Numerical reasoning helps only if the source, period, unit and row are correct.
DocFinQA	Document-level financial QA: Question answering over longer financial documents ¹²⁵. Best used to test long report packs, page references and cross-page evidence.	Document-level QA still needs parser validation and audit export before production use.
MultiHiertt	Hierarchical table-and-text reasoning: Financial QA over hierarchical tables and surrounding text ¹²⁶. Best used to catch subtotal, segment, hierarchy and note-reading failures.	Table hierarchy tests are necessary but narrower than full reporting workflows.
BloombergGPT	Finance-specialised model evaluation: Finance-tuned 50B-parameter model evaluated on financial and general tasks ⁴⁹. Best used to understand how finance corpora can change model behaviour.	Domain-model evidence, not a purchasable universal route or a substitute for licensed data controls.
FinGPT	Open financial LLM ecosystem: Data-centric framework, financial LLM pipeline and open-source ecosystem ⁵⁰⁵¹. Best used to test open financial model routes and adaptation pipelines.	Open-source flexibility moves validation, security, patching and data governance to the buyer.
FinBERT / Financial PhraseBank	Narrow finance NLP: Finance-specific sentiment model and labelled financial phrase data ⁵²⁵³. Best used to keep narrow supervised tasks separate from generative model routes.	A strong narrow classifier is not a reasoning, extraction, forecasting or agentic model.
EDGAR-CORPUS	Financial filing corpus: Public SEC filing corpus for training, retrieval and evaluation work ¹²⁸. Best used to build filing-RAG and parser regression tests.	Useful public source corpus, but currentness, corrections and production source authority still matter.
Loughran-McDonald	Financial lexical baseline: Finance-specific word lists for disclosure and sentiment analysis ¹²⁹. Best used to catch sentiment and disclosure claims that a generic model overstates.	A lexical baseline is a control, not a modern LLM benchmark.
HELM / AISI frontier evals	General model and risk evaluation: General model-evaluation and frontier-risk signals outside finance-specific tasks ¹⁴⁷¹⁴⁸. Best used to trigger re-review when frontier-model risk or capability evidence changes.	Useful for route monitoring, not a financial-work acceptance test.
MLPerf Inference	Serving and hardware benchmark: Inference-performance benchmark for serving efficiency and hardware comparison ¹⁴⁶. Best used for self-hosted economics after a model route has passed finance-quality tests.	Serving efficiency does not measure financial correctness, source support or compliance acceptability.

Each row keeps one benchmark or dataset in its own lane. Do not collapse public filing QA, professional-document QA, table reasoning, agent harnesses, domain models, corpora, lexical baselines and serving benchmarks into a single procurement ranking.

FinanceBench exposed a filing-RAG failure baseline

FinanceBench reported that GPT-4-Turbo with retrieval incorrectly answered or refused 81% of sampled questions ²⁰. Treat this as a minimum-warning signal for filing RAG, not a current universal model ranking.

Benchmark translation

Public signal	Wrong read	Correct read
High filing-QA score.	Approve for credit decisions.	Test private credit files, source recency, period handling and reviewer acceptance.
Good extraction score.	Approve for financial modelling.	Test formulas, assumptions, units, footnotes and reviewer edits.
Strong agent benchmark.	Allow autonomous finance actions.	Gate tool permissions, submission rights, logs, dry runs and human approval.
Legal-agent trajectory signal ¹⁹⁸.	Assume the legal effect sizes transfer to finance.	Log source coverage before drafting, validation, substantive revision, targeted retrieval, structured tools, grounding, tool fan-out and first-pass outputs.
Professional PDF benchmark or parser-rerun gain.	Buy the parser or model from the leaderboard.	Retest on buyer filings, statements, KYC packs, policy documents, permissions and reviewer acceptance.
Good long-context result.	Skip retrieval.	Test latency, cost, context packing, source visibility and no-answer behaviour.
Domain benchmark win.	Use one model everywhere.	Route by use case and rerun private evals after model or prompt changes.
Conversational finance QA pass.	Approve client chat.	Add source age, account-context simulation, refusal boundaries and reviewer acceptance.
Strong numerical-reasoning result.	Approve valuation or credit use.	Test assumptions, units, accounting convention, policy limits and reviewer overrides.
Finance-specialised model paper.	Treat domain pretraining as governance evidence.	Demand source rights, current data, deployment route and validation artefacts.
Public filing corpus coverage.	Treat the corpus as an official source.	Confirm against SEC APIs, amendments, deletion notices and filing timestamps.
Serving benchmark improvement.	Read throughput as model quality.	Use serving metrics only after finance-quality, security and audit tests pass.

Agent trajectory eval

Behaviour	Harvey LAB legal-agent signal	Private eval translation
Source coverage before drafting	Harvey LAB treats source coverage before drafting as a useful trajectory signal. ¹⁹⁸	Check whether the route opens the required filing, policy, market-data or customer source set before first draft.
Post-draft validation	Harvey LAB treats post-draft validation as a useful quality signal. ¹⁹⁸	Require a review step against instruction, source, policy and entitlement before acceptance.
Verifying and revising	Harvey LAB treats revise-after-check behaviour as a useful trajectory signal. ¹⁹⁸	Score whether validation changes the deliverable, not only whether the agent self-approves.
Targeted retrieval	Harvey LAB treats targeted retrieval as useful evidence-seeking behaviour. ¹⁹⁸	Test search for missing facts instead of relying only on initially opened context.
Structured analysis	Harvey LAB treats structured tool use as a quality signal when the work requires it. ¹⁹⁸	Use tools, schemas or calculations where parsing, arithmetic or reconciliation matters.
Grounding against the record	Harvey LAB treats returning to source evidence after drafting as a useful grounding signal. ¹⁹⁸	Check whether the agent returns to evidence after drafting to correct unsupported claims.
High tool parallelism	Harvey LAB treats broad tool fan-out as a risk signal when it creates noisy context. ¹⁹⁸	Measure whether broad fan-out creates extra cost, poor synthesis or context conflict.
Drafting without review	Harvey LAB treats first-pass drafting without later validation as a negative trajectory signal. ¹⁹⁸	Block first-pass outputs for consequential workflows unless a separate reviewer gate runs.

Financial evidence is structured, temporal and permissioned. A financial model cannot reason over evidence it never received. Financial corpora are awkward: filings, tables, scanned statements, call transcripts, spreadsheets, KYC files, sanctions hits, policy documents, customer records, Bloomberg-style data, PDFs with footnotes and emails with attachments.

GDP.pdf is a useful input-layer warning because it measures professional PDF answer acceptance, not just OCR neatness ²⁰¹²⁰². Reducto's vendor-reported rerun suggests structured parse context can improve final answers while lowering answer latency and reasoning tokens ²⁰⁴.

The evidence path should preserve document identity, page and row references, period, units, currency, entity, account, source timestamp, data entitlement and confidence. That information is not metadata decoration; it is how a reviewer decides whether the answer can survive.

Approve parsers, embeddings and rerankers as model components. A bad parser can remove a table. A weak embedding can miss the right policy. A reranker can choose the wrong quarter. The generator then writes a polished answer from damaged evidence.

No-answer behaviour belongs in the acceptance test. A system that refuses when the filing, schedule or permission is missing is safer than one that guesses in an examiner's voice.

How financial context gets made

Source set

The firm has more material than the model should receive.

Latest 10-Kofficial filing
Q2 10-Qperiod-specific filing
Restatement 8-Kcorrection event
Earnings-call transcriptmanagement language
Licensed market priceentitled data
Private holdingpermission-bound portfolio data
Old analyst memostale internal view
OFAC list hitofficial sanctions source

Retrieved

Search narrows the universe but can still pull stale or unauthorised material.

Latest 10-Kcandidate
Q2 10-Qcandidate
Earnings-call transcriptcandidate
Licensed market pricecandidate
Old analyst memocandidate
OFAC list hitcandidate

Reranked

The reranker should push authoritative, current and entitled sources above topical material.

Q2 10-Qperiod match
Restatement 8-Kshould be retrieved
Latest 10-Kbase disclosure
Licensed market priceentitled data
Earnings-call transcriptmanagement context
Old analyst memostale

Model context

The context window fits four chunks; the route must not crowd out the restatement or permission check.

Q2 10-Qinside context
Latest 10-Kinside context
Licensed market priceinside context
Earnings-call transcriptinside context
Restatement 8-Kmissing correction
Old analyst memooutside window

The finance risk is upstream of prose: if the correction, entitlement or current source misses retrieval, the final answer may look professional while failing review.

What the evidence path must preserve

Source	Preserve	Failure	Test
10-K or 10-Q	Filing date, period, note, table, segment, currency and page.	Revenue from the wrong period.	Ask period-specific questions with restatements and footnotes.
Professional PDF pack	Layout, tables, diagrams, footnotes, cross-references and page anchors.	Answer model reasons over a misread table or diagram.	Compare PDF-only answers against answers with structured parser context and reviewer acceptance.
Earnings call	Speaker, timestamp, prepared remarks, Q&A and guidance status.	Management aspiration treated as audited fact.	Separate guidance, actuals and analyst question context.
Spreadsheet	Formulas, hidden sheets, comments, named ranges and linked workbooks.	Visible value copied while formula logic breaks.	Diff formulas and outputs before and after AI changes.
KYC pack	Identity documents, ownership graph, sanctions hit, adverse media and reviewer notes.	Linked entity omitted.	Seed near-match names and nested ownership structures.
Policy library	Effective date, jurisdiction, business line and superseded status.	Stale policy answer.	Ask questions where old and current policies conflict.
Customer account record	Entitlement, timestamp, product, consent and audit trail.	Unauthorised data exposure.	Run permission-bound retrieval and red-team cross-account leakage.
Market-data feed	Timestamp, venue, licence, instrument identifier and adjustment logic.	Stale or unlicensed price drives the answer.	Ask the same question across delayed, real-time and unauthorised data paths.
Regulatory source	Issuer, effective date, jurisdiction, amendment and official URL.	Superseded guidance is treated as current.	Seed old and current rules with similar language.

Retrieval controls

Control	Good threshold	Reviewer signal
Recall by source type	Right document, table and period appear in candidate set.	Source IDs and missing-source warnings are visible.
Reranker precision	Correct evidence outranks similar but stale or wrong-period evidence.	Top evidence has period, entity and row support.
Permission filtering	Unauthorised records never enter retrieval candidates.	Access-denied reason appears in logs.
No-answer policy	Model refuses when required evidence is absent.	Refusal names the missing source or permission.
Regression suite	Route reruns after parser, embedding, reranker or generator changes.	Diff shows changed answers and changed evidence.

Source validation tests

Test	Breaks	Approval artefact
Restatement trap	Routes that answer from the first matching filing.	Filing ID, amendment link and answer diff.
Wrong-period table	RAG that grabs a similar table from another quarter.	Period, table, row, column and extracted value.
Currency-unit mismatch	Models that add millions, thousands and raw values.	Unit normalisation and calculation trace.
Entitlement denial	Retrieval that exposes private portfolio or licensed data to the wrong user.	Access-control decision and denied-source log.
Entity resolution conflict	KYC, LEI and issuer-name matching under similar names.	Identifier mapping and reviewer override.
Sanctions near match	Name matching that hides weak aliases or false positives.	Official list source, match score and investigator note.
Spreadsheet hidden state	Visible-value extraction that ignores hidden tabs and formulas.	Workbook diff and formula preservation report.
Standards-bound message	Free-form generation for ISO 20022, Swift or FIX fields.	Schema validation and rejected-field record.

Model approval follows intended use, jurisdiction and customer impact. Finance does not need one AI rulebook to need hard AI evidence. U.S. banking agencies revised model-risk guidance while leaving generative and agentic AI to a separate future lane ²⁶²⁷. FINRA reminded firms that existing rules still apply to generative AI ³¹. The SEC has already treated exaggerated AI claims as enforcement material ³³. MAS proposed AI inventories and lifecycle controls ³⁴. The EU AI Act names creditworthiness and life or health insurance pricing as high-risk use cases ³⁷³⁷.

The practical rule is simple: intended use decides the gate. The same general-purpose model can be low-risk in internal policy search and high-risk when deployed into credit, insurance or customer advice.

Do not let a model owner mark a route approved because the vendor has an enterprise privacy page. That page helps, but approval also needs source rights, data classification, output use, human oversight, explainability, incident handling, vendor dependency, recordkeeping and model-change control.

For agentic routes, the regulatory question expands from what the model said to what the system did.

Regulatory output gates

Gate	Trigger	Required action
High-risk AI	EU creditworthiness, credit scoring, or life and health insurance pricing for natural persons.	Classify, document, test, log, assign human oversight and maintain deployer evidence.
Model-risk governance	Bank model supports significant business lines, operations, services or functions.	Apply model identification, validation, monitoring, governance and vendor controls.
Securities conduct	Broker-dealer or adviser uses GenAI for communications, supervision, research or investor-facing workflows.	Map to existing supervision, communications, books-and-records and suitability obligations.
Credit explanation	AI contributes to adverse credit action.	Preserve specific, accurate reasons; test proxies; keep human review and audit logs.
Operational resilience	Cloud, model, data, agent or ICT provider supports important financial functions.	Assess concentration, continuity, audit rights, incident route, exit plan and third-party risk.
Marketing and public claims	Firm claims AI use, model performance, personalized advice or predictive ability.	Substantiate claims and keep evidence for what the model does in production.
Agent action	AI can update records, submit reports, alter spreadsheets, send communications or invoke tools.	Use permissions, dry-run, approval, action logs, rollback and incident review.

Jurisdictional read

Jurisdiction	What to watch	Route effect
United States banking	Revised model-risk guidance and forthcoming AI-specific treatment.	Do not treat GenAI carve-out as free use; keep governance evidence.
United States securities	FINRA technology-neutral obligations and SEC AI claim substantiation.	Archive prompts, outputs and approvals where they become business records or marketing evidence.
European Union	AI Act high-risk classification and DORA third-party ICT risk.	Separate credit, insurance, fraud, customer and operational routes by intended use.
United Kingdom	Bank of England, PRA and FCA AI monitoring, systemic-risk and third-party exposure work.	Treat common model dependency as a resilience and concentration question.
Singapore	MAS AI risk-management guidance, inventories and lifecycle controls.	Build route inventory and materiality assessment into approval flow.

The deployment path decides what the model may see and what the institution can prove. Deployment is a risk control because the host, contract and data boundary decide what the model may see and what the institution can prove. Direct APIs, cloud model platforms, self-hosted open weights, productivity suites and vertical SaaS can all run useful financial AI. They are different approval objects because they carry different evidence burdens.

Approve the route by data class, workflow consequence, source rights, reviewer evidence and operational-resilience burden before comparing model names.

The route file should name retention, training use, region, audit logs, subprocessors, incident rights, exit path, gateway controls, recovery plan and fallback owner.

Related report

Deployment options for open-source LLM models

Use the deployment report for licence, hosting, managed endpoints, cloud catalogues, serving engines, RAG, gateways, GPU economics, controls and recovery.

Finance agents need action governance before autonomy. A finance agent can be wrong and also do something. Financial agents are attractive because the work is multi-step: gather sources, fill a model, reconcile accounts, draft a case summary, update a CRM, prepare a deck, submit a review note or trigger a workflow.

That same tool access changes the risk. The control question is no longer only whether the answer is correct. It is whether the system had permission to read the data, call the tool, write the record, send the message or escalate the case.

Start agents in observe or draft mode. Promote them only after action logs, dry-run comparisons, reviewer acceptance, false-positive and false-negative tracking, rollback, separation of duties and incident routes work.

For regulated or market-facing workflows, keep final action authority with accountable staff until the institution can prove the agent's behaviour under drift, outage, permission changes and adversarial inputs.

Agent permission ladder

Level	Allowed	Finance fit	Promotion test
0 - Chat	Answer from approved sources; no tools.	Learning, policy help and low-risk Q&A.	Citation accuracy and no-answer behaviour.
1 - Search	Retrieve from approved sources and show evidence.	Research, policy search and filing QA.	Permission filtering, source recall and stale-source rejection.
2 - Draft	Prepare memos, case notes, emails, slides and spreadsheet suggestions.	Analyst and operations support.	Reviewer acceptance, formula diff and archive capture.
3 - Write behind approval	Update records or submit workflow steps only after human approval.	KYC notes, reconciliations, compliance cases and finance close.	Action log, rollback and separation-of-duties evidence.
4 - Autonomous bounded action	Act within low-risk thresholds and reversible operations.	Narrow back-office tasks after strong monitoring.	False-action rate, drift alerts, incident drills and sampled audit.

Agent failures to test

Failure	Example	Control
Permission overreach	Agent reads restricted portfolio or customer records.	Entitlement-aware retrieval and access-denied logs.
Tool hallucination	Agent assumes a database update succeeded.	Tool-response verification and idempotent writes.
Spreadsheet mutation	Agent changes formulas while preserving visible values.	Workbook diff, formula tests and approval before save.
Escalation miss	AML case summary omits linked adverse media.	False-negative suite and investigator review.
Conduct breach	Client communication includes misleading performance language.	Policy retrieval, banned-claim rules and supervisor approval.
Runaway retries	Agent calls tools repeatedly after ambiguous results.	Budget, stop rules and exception queue.

The private eval is the procurement brief. A useful eval looks like the work, including the ugly parts. Do not ask vendors to answer polished sample questions. Give them the real source mix: old and new filings, bad scans, hidden spreadsheet tabs, ambiguous periods, duplicate customers, restricted records, stale policies, near-match sanctions names, reviewer exceptions and missing evidence.

The eval should test the whole route. Parser, embedding, reranker, generator, tools, prompt, policy layer, permission filter, reviewer UI and logs all belong in scope.

Keep a holdout set and rerun it after material changes. Model version, system prompt, parser, embedding model, reranker, source corpus, threshold, tool permission, vendor term and safety-policy changes can all move output quality.

Use public benchmarks to choose categories, then let the private corpus decide procurement.

Private eval suite

Eval	Include	Pass condition
Filing QA	Public filings, restatements, guidance, segments and period traps.	Correct answer, source, values, formula and no-answer where missing.
Private-document QA	Policies, customer files, KYC packs, statements and permissions.	No cross-account leakage, correct source recall and reviewer-usable citations.
Table and spreadsheet reasoning	Footnotes, merged cells, hidden tabs, formulas and linked workbooks.	Formula integrity and calculation trace.
Regulated-decision support	Credit or insurance examples with edge cases and protected-class proxy tests.	Explainable reasons, fairness checks, human approval and audit export.
Agent action	Tool calls, failed tools, retries, writes, rollback and approval gates.	No unauthorised action and complete action log.
Adversarial and drift	Prompt injection, stale sources, changed market data and policy conflicts.	Safe refusal, source precedence and drift alert.

Eval scorecard

Dimension	Score	Evidence
Correctness	Answer accepted by domain reviewer.	Reviewer decision and correction type.
Source support	Every material claim maps to approved evidence.	Source IDs, spans, page/table references.
Calculation	Intermediate values and formulas are correct.	Formula trace and deterministic recomputation.
Permission	Only authorised data enters the route.	Access logs and denied-source tests.
Robustness	Prompt injection, stale data and missing evidence handled.	Red-team cases and refusals.
Auditability	Route can be replayed or explained.	Model version, prompt, tool calls, output and reviewer log.
Economics	Accepted output beats baseline cost.	Token cost, tool cost, latency and reviewer minutes.

Build the programme around control paths, not pilots. The quickest safe path is a small route register with hard gates. Start with a two-by-two portfolio: low-risk productivity routes and high-value evidence routes. Avoid the temptation to begin with autonomous regulated decisions. Prove source capture, review, logs and economics before adding action permissions.

Every route gets an owner, reviewer, data class, model lane, source path, acceptance test, monitoring metric and reapproval trigger.

The operating model should connect business owners, model risk, compliance, legal, security, data governance, procurement, finance and internal audit. The point is not bureaucracy. The point is to prevent the same model from being approved five times under five names with no common evidence.

Use pilots to populate the register. Retire routes that cannot produce source-supported, auditable and economically useful outputs.

Weeks 0-2Inventory

Routes

AI route register with use case, data class, owner and intended use.

Controls

Minimum approval gates by data class and customer impact.

Sources

Priority corpora and access-control map.

Weeks 3-6Bakeoff

Models

Frontier, cloud, open-weight, embedding, reranker and parser bakeoff.

Eval

Private test set with acceptance criteria and reviewer workflow.

Contracts

Retention, training-use, region, audit and exit checklist.

Weeks 7-10Controlled launch

Operations

Draft-only or human-approved routes in production.

Monitoring

Accepted output, reviewer minutes, refusal and incident dashboard.

Governance

Model-change regression and reapproval process.

Weeks 11-14Scale or stop

Scale

Promote routes that clear quality, economics and control thresholds.

Stop

Retire routes with weak source support or unfixable audit gaps.

Audit

Evidence pack for model risk, compliance and internal audit.

Buyer checklist

Item	Minimum	Owner
Route register	Every AI route has use case, data class, model lane and owner.	AI governance or model risk.
Source approval	Corpus, rights, currentness and permissions documented.	Data governance.
Eval pack	Private test set, holdout set, acceptance criteria and reviewer rubric.	Business plus model validation.
Contract review	Training use, retention, region, audit, incident and exit terms cleared.	Procurement, legal and security.
Action gate	Tool permissions, dry-run, approval, logs and rollback before writes.	Engineering and operations.
Monitoring	Accepted output, reviewer minutes, drift, refusals and incidents tracked.	Route owner.
Reapproval	Model, prompt, parser, retriever, source or tool change triggers regression.	Model risk and route owner.

Finance AI needs route controls, not a model-use policy alone. The control stack should follow the route from source to action. A generic AI policy is too blunt for finance. The useful control object is a route with named inputs, outputs, permissions, reviewers and monitoring. That route can then inherit model-risk, technology-risk, conduct, privacy, data-governance, operational-resilience and third-party controls.

The common mistake is to validate the model and forget the source chain. A route can fail because the source is stale, the retriever crossed an entitlement boundary, the parser lost a table, the model update changed refusal behaviour, or the agent gained a new write permission.

Run controls at three moments: before launch, during use and after change. Before launch, classify data and intended use. During use, log sources, outputs and actions. After change, rerun holdouts and reapprove material shifts.

For regulated customer or market-impacting work, keep the human role specific. 'Human review' is not enough. Name the credit officer, investigator, supervisor, trader, controller or adviser who owns the decision.

Control stack by layer

Layer	Control	Evidence	Owner
Inventory	Use case, model, source, data class and owner registered.	Route register and materiality score.	AI governance.
Data	Rights, retention, region, permission and lineage approved.	Data classification and access logs.	Data governance and privacy.
Source	Currentness, corpus scope, retrieval and no-answer policy tested.	Recall tests and missing-source cases.	Business and model validation.
Model	Capability, safety, drift and regression evals passed.	Private eval pack and holdout reruns.	Model risk.
Output	Citations, calculations, explanations and records preserved.	Answer packet and archive record.	Compliance and route owner.
Action	Tool permissions, approvals, rollback and separation of duties.	Action log and sampled review.	Engineering and operations.
Third party	Concentration, exit, continuity, audit and incident rights.	Vendor risk file and exit drill.	Procurement and operational risk.

Controls tighten as outputs touch money

Internal

Customer

Regulated decision

Agent action

Data classification

Medium

High

Gate

Human approval

Sample

Required

Named

Before write

Audit export

Useful

Required

Gate

Security red-team

Baseline

High

Gate

SampleHighGate

Global finance breaks simple model approval. A model can speak the language and still fail the financial system. Cross-border finance adds jurisdiction, licence, source, market convention, product taxonomy, data-transfer and reviewer problems. A general model may translate a disclosure correctly while using the wrong market convention, stale source, unauthorised data path or unsupported advice standard.

The route should separate five questions: where the data may go, which source is authoritative, which market convention governs the answer, which regulator owns the activity and which human can approve the output.

Sovereign or self-hosted models can help with data movement and concentration risk. They do not create local expertise. A PRC credit file, EU consumer-credit explanation, UK adviser note, Singapore MAS-regulated AI inventory and U.S. broker-dealer communication need different evidence.

For multinational firms, the answer is a jurisdictional route matrix, not a global ban or global approval.

Cross-border routing tests

Market	Model trap	Approval test
United States	Treating AI washing, credit explanations and broker-dealer communications as generic marketing or drafting.	Substantiation, adverse-action reason, supervision and books-and-records review.
European Union	Missing high-risk classification for covered credit and insurance use.	AI Act classification, DORA third-party route and GDPR decisioning checks.
United Kingdom	Ignoring systemic, third-party and customer-outcome framing.	FCA/PRA/BoE governance, critical-third-party and operational-resilience review.
Singapore	Treating FEAT and MAS guidance as voluntary brand material.	Inventory, materiality, lifecycle controls and FEAT assessment.
Canada	Assuming only regulatory-capital models count.	E-23 model inventory and third-party model-risk review.
Australia and Hong Kong	Using global model approval without local conduct and accountability mapping.	Local regulator expectations, data path and customer-impact assessment.

Shortlist model lanes by workflow, not by leaderboard. The shortlist should name the best lane for each job. Finance buyers need a portfolio: frontier reasoning models for hard synthesis, financial-domain models for narrow language and sentiment tasks, private models for sensitive classification, document models for source capture, embeddings and rerankers for retrieval, and deterministic tools for calculations.

Finance-specific models deserve a place on the diligence list, not a free pass. BloombergGPT shows the value of privileged domain corpora. FinGPT shows the open-source path. FinBERT shows that narrow financial sentiment can beat a broader model on a narrower job. None of them clears customer, credit, insurance, AML, trading or advice workflows by name alone.

The final shortlist should be a route table. Each lane gets a primary task, forbidden task, source dependency, data-class limit, eval suite and reapproval trigger.

If the buyer cannot name the lane's forbidden task, the approval is too broad.

Recommended model lanes

Lane	Approve for	Avoid	Private eval
Frontier reasoning	Analyst synthesis, drafting, exception analysis and multi-step question answering.	Autonomous credit, advice, trading or insurance decisions.	Filing QA, judgement rubrics, citations, refusals and reviewer minutes.
Financial-domain LLM	Financial language, sentiment, disclosure and narrow domain tasks.	Assuming domain training means current licensed data.	Currentness, corpus overlap, leakage and task-specific labels.
Open-weight private model	Sensitive classification, local summarization and high-volume low-risk tasks.	Underfunded self-hosting and validation.	Security, drift, throughput, quality and operating cost.
Embedding and reranker	Filing, policy, KYC and research retrieval.	Treating retrieval quality as a generator feature.	Recall, precision, stale-source rejection and permission filtering.
Document AI	Tables, handwriting, scans, statements, claims and forms.	Converting all source evidence into lossy text.	Table structure, page references, hidden fields and confidence flags.
Deterministic calculator/tool	Ratios, covenants, reconciliations, eligibility rules and checks.	Letting LLM prose perform arithmetic silently.	Formula trace, unit handling, rounding and exception logic.

The strongest cases sell the operating wrapper, not the model alone. Public implementation evidence points to work systems. Morgan Stanley's public OpenAI case is useful because it describes evals, workflow adoption and controlled access to a wealth-management knowledge base ³. It does not prove that every financial adviser, credit officer or trader can use the same route.

The same pattern appears elsewhere. Finance-agent templates package skills, connectors and subagents. Financial-crime vendors package detection, case management and analyst review. Reporting and close platforms package AI inside approval and audit workflows.

Case studies should be read as implementation-shape evidence. They can show where value appears: source access, review UX, action limits, workflow adoption, analyst productivity, false-positive reduction or reporting speed.

They rarely publish enough to buy from: denominator, base rate, error severity, false negatives, reviewer minutes, audit failure rate, model version and control group.

Wealth management

Morgan Stanley and OpenAI

OpenAI case study cites eval-led adoption, knowledge-base access and high advisor usage.

Workflow: Advisor knowledge retrieval and summarization.
Stack: OpenAI frontier model route over internal wealth-management content with evals and controls.
Deployment: Internal assistant for financial advisors.
Result: Faster access to internal knowledge and advisor workflow support.
Caveat: The public case does not publish full error rates, client-outcome measures or regulator-ready audit samples.

Banking, insurance and asset management

Anthropic financial services agents

Anthropic's announcement describes ready-to-run financial-services templates and governed data connectors.

Workflow: Pitchbooks, KYC screening, month-end close and finance operations.
Stack: Claude agent templates with skills, connectors and subagents.
Deployment: Plugins, cookbooks and managed-agent patterns.
Result: Reusable route patterns for repeated finance workflows.
Caveat: Template availability does not approve local data, tool permissions or reviewer gates.

Financial information and analytics

BloombergGPT

The paper reports finance-domain data scale and benchmark improvements on financial tasks.

Workflow: Financial NLP over filings, news, press and proprietary financial data.
Stack: Finance-specialized LLM trained on mixed domain and general corpora.
Deployment: Research and Bloomberg ecosystem signal rather than a generic buyer API route.
Result: Shows the value of proprietary financial corpora.
Caveat: A finance-tuned model still needs current source retrieval, licensing and private evals.

Financial crime

NICE Actimize Xceed AI agents

NICE's announcement describes agents for triage, backlog categorization, high-risk summaries and analyst-in-the-loop workflows.

Workflow: Fraud and AML alert triage, categorization and case summarization.
Stack: AI agents embedded in fraud and AML case-management workflows.
Deployment: Financial-crime SaaS platform.
Result: Targets alert-volume pressure and investigator productivity.
Caveat: Buyers still need false-negative testing, SAR accountability, model drift monitoring and audit exports.

Reporting and finance operations

Workiva and BlackLine

Vendor materials place AI inside reporting and financial-close processes.

Workflow: Reporting, close, reconciliation and controlled financial workflows.
Stack: Generative AI and automation inside finance systems of record.
Deployment: Application-layer AI with approvals and workflow context.
Result: Useful early lane for finance operations if formulas, approvals and archive records survive.
Caveat: Spreadsheet state, hidden formulas, approval rights and recordkeeping remain the acceptance tests.

Financial-services data platforms

Databricks and Snowflake

Financial-services materials emphasize governed data, sensitive workloads and resiliency.

Workflow: Governed data, analytics and AI workloads.
Stack: Lakehouse and cloud data platform routes with governance controls.
Deployment: Data and model-development infrastructure.
Result: Good route for firms that need controlled data and model pipelines.
Caveat: Platform governance does not validate the model, prompt, source retrieval or customer outcome.

Model governance needs monitoring outside the vendor deck. The dull records tell you when a finance model route has changed. A model shortlist needs records vendors do not polish into marketing: status pages, router catalogues, package registries, model hubs, issue trackers, security advisories, source-provider notices, regulator pages and archived trust pages. These sources do not rank models. They reveal context limits, price changes, supported tool parameters, release churn, dependency risk, reliability and the errors users keep hitting.

Reliability needs translation. A one-point availability gap is 0.01 x 365 days, or 3.65 days a year if spread evenly. If those hours land inside deadline-sensitive production windows, the business impact is not a small decimal. The route needs a fallback owner, a degradation mode and a retest trigger.

Use operational signals as deltas in the route register. The useful question is narrow: does this change the route, the control, the fallback, the eval, the contract or the reviewer instruction?

Worst-week downtime cannot be inferred from an aggregate uptime percentage. Incident timing, component scope, recovery cadence and customer tier decide whether a published decimal was harmless background noise or a production event.

Operational signals do not prove model quality. They identify the procurement questions that deserve renewed scrutiny.

Start with the decimal because it is the number procurement and risk teams usually see first. Use only observed status-page uptime here; contractual SLA floors are a different source class and do not belong in the same chart.

Status-page decimals hide fallback work

Status-page snapshot captured on 27 May 2026. OpenAI and Claude publish observed status-page uptime windows. Higher observed uptime is better. OpenAI is kept filled and Claude dotted for consistent route emphasis. Azure OpenAI and Gemini/Vertex SLA floors are not charted because contractual thresholds are not public aggregate uptime history.

The zero-origin percentage view explains why the numbers look reassuring. The buyer question appears only after observed uptime inputs are translated into time that someone must cover with retries, fallback routing, client communication or manual work.

Observed decimals become fallback days

Annualized downtime is arithmetic from observed status-page uptime only: (100 - availability) / 100 x 365 days. Lower is better. OpenAI is kept filled and Claude dotted for consistent route emphasis. It is not a one-year SLA result, an SLA-floor extrapolation, or a model-quality score.

Annualized days still smooth the calendar. The harder operational test is whether official incidents clustered inside a period the buyer could not move.

Clusters decide deadline risk

Official incident-page snapshot captured on 27 May 2026. OpenAI and Claude values sum status incidents in the busiest relevant seven-day cluster. Azure OpenAI and Gemini/Vertex values use official cloud incident durations because those sources expose incident history rather than a comparable observed uptime percentage. Lower clustered incident time is better; fill pattern is kept fixed to OpenAI and Claude for route continuity, not rank. The numbers are not user-specific downtime; they show why deadline-sensitive production windows need fallback design.

Reliability is only one signal family. The same route-review discipline applies when catalogues, model cards, package registries, trust pages, security advisories or source-provider records change.

Signals that reopen approval

Source	Signal	Use
Provider and cloud status pages	Published uptime windows, incident history, component degradation and recovery cadence.	Test fallback routing, degradation modes and client promises; do not treat unequal status-page windows as an SLA ranking.
Model-router catalogues	Model availability, context limits, price, tool support, structured outputs and provider routing.	Build a substitution matrix before committing sensitive workflows to one provider.
Model hubs and release notes	Downloads, licences, tags, task families, model cards, deprecations and update cadence.	Separate real route momentum from paper-only interest and stale model assumptions.
Package registries and issue trackers	SDK downloads, release churn, dependency changes, tool-call errors, rate limits and regression reports.	Design private evals around failures that users and integrators are already hitting.
Archived trust, privacy and pricing pages	Changed claims, removed limits, policy drift, data-use language and price packaging.	Ask why the route changed before renewing approval or expanding use.
Security advisories	CVEs, known exploited vulnerabilities and affected model-serving, parser, browser, notebook and agent dependencies.	Map the route's runtime stack to patch gates, compensating controls and reapproval triggers.
Regulator, procurement and source-provider records	New rules, access changes, public tenders, source-rights constraints and resilience expectations.	Attach policy, source-rights and operational-resilience changes to the route register.
Regulator and standard updates	FSB, DORA, model-risk, outsourcing, AI-security and financial-stability language.	Update inventories, materiality assessments, third-party evidence and owner attestations.
Benchmark and eval-framework releases	FinanceBench, Finance Agent, HELM, MLPerf-style or local harness changes.	Rerun only the private eval rows that map to the changed public signal.
Cloud-region and productivity-suite incidents	Region degradation, customer impact, shared provider concentration and connector downtime.	Review concentration, data-residency assumptions, fallback tests and customer-impact records.

Watch people who publish useful tests, failures and boundary conditions. The useful follows are named people publishing finance-model tests, failure analysis, market-risk methods and route-control warnings. General AI commentary and title-based market commentary are too broad for finance procurement. Monitor people who release benchmarks, datasets, validation methods, fairness tests, market-structure research or regulatory critiques that change a private eval or approval question.

If the signal is institutional, keep it in the source ledger, watch list or vendor directory. The expert watchlist is for humans whose public work gives buyers a better test.

Replace anyone whose signal turns into generic commentary, product marketing or status updates without reusable methods.

Signals to monitor

Expert	Lane	Signal	Links
Andrew W. Lo MIT Sloan	Financial economicsUSMarket behaviour, adaptive markets and financial innovation.	Useful for separating model performance from market feedback and investor-behaviour effects. New work on AI, markets, investor behaviour or systemic feedback.	Open Andrew W. Lo primary work
Marcos Lopez de Prado Cornell; MLFinLab	Financial machine learningUSML in asset management, backtesting, leakage and financial model validation.	Useful for detecting benchmark leakage, overfitting and unrealistic trading evals. Methods for financial ML validation and backtesting discipline.	Open Marcos Lopez de Prado primary work
Hilary J. Allen American University Washington College of Law	Financial stability and lawUSFintech, systemic risk and regulatory limits.	Useful counterweight to vendor optimism on systemic and public-interest risk. Legal analysis of AI, finance and systemic fragility.	Open Hilary J. Allen primary work
Sandra Wachter Oxford Internet Institute	AI governanceUKExplainability, fairness and algorithmic accountability.	Useful for credit, insurance and customer-impact explanation tests. Research on meaningful explanations and fairness evidence.	Open Sandra Wachter primary work
Daniel Schwarcz University of Minnesota Law School	Insurance and AIUSInsurance law, discrimination and algorithmic underwriting.	Useful for life, health and P&C model-route risk. Insurance AI, pricing, claims and fairness scholarship.	Open Daniel Schwarcz primary work
Maureen O'Hara Cornell Johnson	Market microstructureUSTrading, liquidity and market structure.	Useful for reading AI's market-integrity and correlation risks. Research on trading systems and market behaviour under automation.	Open Maureen O'Hara primary work
Adair Morse Berkeley Haas	Household finance and fintechUSFinancial access, fintech, consumer finance and policy design.	Useful for credit, consumer-impact and inclusion questions that procurement teams often under-test. Research or policy work on fintech, consumer outcomes and AI-enabled finance.	Open Adair Morse primary work

Use the directory to route diligence by layer. A vendor list is useful only when it preserves the layer distinction. OpenAI, Anthropic, Google, AWS and Microsoft compete at the frontier and cloud route layer. Bloomberg, LSEG, FactSet, S&P Global, AlphaSense, Moody's and PitchBook compete at the financial source and workflow layer. NICE Actimize, FICO, SAS, Feedzai, Featurespace, Mastercard and Visa compete in fraud, AML and payments. Workiva, BlackLine, Oracle and SAP compete in finance operations.

The diligence question changes by layer. Model providers need data-use, retention, version and evaluation questions. Data providers need source-rights, entitlements and currentness questions. Workflow providers need action logs, review UX and audit-export questions.

Do not compare every row as if it sells the same thing. Some vendors own the model. Some own the data. Some own the workflow. Some own the control plane. Many borrow at least one layer.

Layer labels decide which diligence questions to ask before a route reaches production.

Missing evidence should become a purchasing test. The public evidence is strongest for filings and weakest where money, permissions and actions meet. Public benchmarks expose important weaknesses, but they rarely show private customer data, live market data, portfolio entitlements, operational outages, regulator exams or final human accountability.

That gap is not an academic footnote. It tells buyers what to test before scale.

The highest-value missing evidence concerns false negatives in AML and fraud, customer-outcome effects in credit and insurance, concentration risk across common model providers, live-market agent behaviour, and audit quality for tool-using systems.

Until public evidence improves, financial institutions should assume each material route needs a private acceptance test and monitoring plan.

Gaps to turn into tests

Gap	Risk	Private test
Live market data	Agent uses stale prices or wrong corporate action.	Replay market-data changes and verify source timestamps.
Private permissions	Model sees unauthorised account, portfolio or employee data.	Permission-bound retrieval and red-team leakage tests.
AML and fraud false negatives	System misses suspicious activity or coordinated fraud.	Labelled case replay with false-negative review.
Customer outcomes	Credit or insurance route produces unfair or unexplained decisions.	Fairness, reason-code and appeal-path testing.
Agent audit export	Firm cannot reconstruct action chain after incident.	Replay tool calls, approvals, model versions and rollback events.
Third-party concentration	Common provider outage or model defect hits many functions.	Dependency map, exit drill and fallback route.
Model-change drift	A vendor upgrade silently changes output quality.	Holdout rerun after every material route change.

Primary sources, benchmarks and regulatory anchors.

Primary source list

Ref	Source	Verification	Evidence	Open source
1	OpenAI API model docs	Publisher: OpenAI API model docs; Kind: Official model documentation; Checked: 2026-06-19; Status: verified	E1: OpenAI's model catalogue positions model choice as a product and API routing decision across reasoning effort, cost, latency, multimodal input and tool use.	Open source
2	OpenAI enterprise privacy	Publisher: OpenAI enterprise privacy; Kind: Privacy and data terms; Checked: 2026-06-19; Status: verified	E2: OpenAI's enterprise privacy page describes API retention, zero data retention eligibility, compliance posture and customer control of business data.	Open source
3	OpenAI and Morgan Stanley case study	Publisher: OpenAI and Morgan Stanley case study; Kind: Case study; Checked: 2026-06-19; Status: verified	E3: OpenAI's Morgan Stanley case study emphasizes private wealth-management content, evaluation practice and domain workflow integration rather than a generic chatbot rollout.	Open source
4	Anthropic model overview	Publisher: Anthropic; Kind: Official model documentation; Checked: 2026-06-19; Status: verified	E4: Anthropic's model overview separates model families by capability, latency and cost, making finance usage a routing problem rather than a single-model mandate.	Open source
5	Anthropic - agents for financial services	Publisher: Anthropic; Kind: Vendor documentation; Checked: 2026-06-19; Status: verified	E5: Anthropic's financial-services announcement packages finance work as templates with skills, connectors and subagents for pitchbooks, KYC files and month-end close. \| E55: Anthropic's finance-agent pattern separates data connectors, domain instructions and subagents, which maps directly to approval gates for tool-using finance systems. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
6	Anthropic commercial privacy terms	Publisher: Anthropic; Kind: Privacy and data terms; Checked: 2026-06-19; Status: verified	E6: Anthropic's commercial privacy terms distinguish processor/controller posture and matter for regulated financial data paths through Claude products.	Open source
7	Google Cloud Vertex AI documentation	Publisher: Google; Kind: Cloud documentation; Checked: 2026-06-19; Status: verified	E7: Vertex AI provides a managed cloud route for Gemini and adjacent machine-learning tooling under Google Cloud control planes. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
8	Google Cloud Vertex AI data governance	Publisher: Google; Kind: Privacy and data terms; Checked: 2026-06-19; Status: verified	E8: Vertex AI data-governance documentation describes controls for customer data, model training use and enterprise deployment posture.	Open source
9	Amazon Bedrock overview	Publisher: Amazon; Kind: Cloud documentation; Checked: 2026-06-19; Status: verified	E9: Amazon Bedrock presents model choice, guardrails, knowledge bases, customization and evaluation as part of a production generative-AI platform. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
10	Microsoft Azure OpenAI data, privacy, and security	Publisher: Microsoft; Kind: Privacy and data terms; Checked: 2026-06-19; Status: verified	E10: Azure OpenAI data privacy documentation states that prompts, completions, embeddings and training data are not made available to other customers or used to train OpenAI models without permission.	Open source
11	Meta Llama model access	Publisher: Meta; Kind: Model documentation; Checked: 2026-06-19; Status: verified	E11: Meta's Llama access materials support an open-weight lane for buyers that need self-hosting, customization or lower inference-cost exposure. \| E65: Open-weight models can reduce external API dependency but move patching, red-teaming, hosting, access control and monitoring responsibilities to the buyer.	Open source
12	Mistral model overview	Publisher: Mistral model overview; Kind: Official model documentation; Checked: 2026-06-19; Status: verified	E12: Mistral's model overview gives buyers a European model-provider lane across commercial APIs and deployable model options. \| E66: European and open model-provider lanes give cross-border firms more deployment options but do not remove validation or regulatory obligations.	Open source
13	Mistral Document AI OCR	Publisher: Mistral Document AI OCR; Kind: Document AI documentation; Checked: 2026-06-19; Status: verified	E13: Mistral Document AI OCR targets complex documents and makes OCR a model-layer choice for RAG and analysis workflows.	Open source
14	Qwen3 embedding and reranker	Publisher: Qwen3 embedding and reranker; Kind: Model documentation; Checked: 2026-06-19; Status: verified	E14: Qwen3 embedding and reranker models expand the multilingual open model lane for retrieval and reranking. \| E67: Embedding and reranking models should be approved separately from generator models because retrieval failure can create a wrong financial answer before generation starts.	Open source
15	Cohere Rerank documentation	Publisher: Cohere Rerank documentation; Kind: Vendor documentation; Checked: 2026-06-19; Status: verified	E15: Cohere's Rerank documentation describes a reranking model layer that can improve retrieved-document ordering before generation. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
16	Voyage AI voyage-3.5 release	Publisher: Voyage AI voyage-3.5 release; Kind: Model documentation; Checked: 2026-06-19; Status: verified	E16: Voyage AI's embedding release frames newer embedding models across finance, law, technical documentation and multilingual retrieval tasks.	Open source
17	Azure AI Document Intelligence layout model	Publisher: Azure; Kind: Document AI documentation; Checked: 2026-06-19; Status: verified	E17: Azure AI Document Intelligence layout models extract text, tables, selection marks and layout from documents before downstream reasoning. \| E58: Document-layout extraction preserves tables and page structure before the LLM sees a source, reducing the chance that cost is wasted on damaged context.	Open source
18	Google Document AI form parser	Publisher: Google; Kind: Document AI documentation; Checked: 2026-06-19; Status: verified	E18: Google Document AI form parser supports structured extraction from forms and documents, useful for document-heavy finance workflows.	Open source
19	AWS Textract	Publisher: AWS; Kind: Document AI documentation; Checked: 2026-06-19; Status: verified	E19: AWS Textract extracts text, handwriting, forms and tables from scanned documents for document pipelines.	Open source
20	FinanceBench paper	Publisher: FinanceBench paper; Kind: Financial benchmark; Checked: 2026-06-19; Status: verified	E20: FinanceBench reports 10,231 public-company financial QA items with answers and evidence strings, built as an open-book financial QA benchmark. \| E21: FinanceBench found GPT-4-Turbo with retrieval incorrectly answered or refused 81% of sampled questions, making financial RAG quality a production issue. \| E57: FinanceBench shows that long context can improve answers but can carry enterprise latency and document-scale limits.	Open source
21	FinanceBench open-source dataset	Publisher: FinanceBench open-source dataset; Kind: Financial benchmark; Checked: 2026-06-19; Status: verified	E22: The FinanceBench repository provides the public sample, filings and evaluation artefacts needed for private reruns and route comparisons.	Open source
22	FinBen benchmark paper	Publisher: FinBen benchmark paper; Kind: Financial benchmark; Checked: 2026-06-19; Status: verified	E23: FinBen spans financial information extraction, textual analysis, question answering, generation, risk management, forecasting and decision-making tasks. \| E24: FinBen reports that LLMs handle extraction and textual analysis better than advanced reasoning, forecasting and complex decision tasks.	Open source
23	TAT-QA benchmark	Publisher: TAT-QA benchmark; Kind: Financial benchmark; Checked: 2026-06-19; Status: verified	E25: TAT-QA was built from real financial reports with hybrid tables and text, and its authors reported a large gap between model and expert-human performance.	Open source
24	ConvFinQA benchmark	Publisher: ConvFinQA benchmark; Kind: Financial benchmark; Checked: 2026-06-19; Status: verified	E26: ConvFinQA focuses on conversational financial question answering with chains of numerical reasoning over real financial context.	Open source
25	Vals AI Finance Agent benchmark	Publisher: Vals AI Finance Agent benchmark; Kind: Agent benchmark; Checked: 2026-06-19; Status: verified	E27: Vals AI's Finance Agent benchmark evaluates multi-step finance workflows that require data retrieval, calculation and evidence-bearing outputs. \| E56: The Vals finance benchmark examples require SEC search, calculations, financial expertise and judged final answers, so raw model tokens do not capture task economics. \| E72: Public agent benchmarks do not fully reveal private-data permissions, vendor connectors, reviewer minutes, false-negative rates or audit-export quality.	Open source
26	OCC Bulletin 2026-13	Publisher: OCC Bulletin 2026-13; Kind: Regulator; Checked: 2026-06-19; Status: verified	E28: OCC Bulletin 2026-13 replaced older model-risk guidance for covered banking organizations and expressly leaves generative and agentic AI outside that guidance's scope. \| E60: The revised OCC model-risk guidance still expects governance, validation, monitoring and vendor consideration for covered models even where generative AI awaits separate treatment.	Open source
27	Federal Reserve SR 26-2	Publisher: Federal Reserve SR 26-2; Kind: Regulator; Checked: 2026-06-19; Status: verified	E29: Federal Reserve SR 26-2 aligns revised model-risk guidance with the banking agencies' updated risk-based model governance posture.	Open source
28	U.S. Treasury - AI in financial services	Publisher: U.S. Treasury; Kind: Regulator; Checked: 2026-06-19; Status: verified	E30: The U.S. Treasury AI in financial services report summarizes uses across underwriting, fraud, AML/CFT, risk management, operations, customer service and cybersecurity. \| E31: Treasury highlighted third-party dependency, data quality, bias, explainability, operational risk and model governance as recurring financial-services AI concerns. \| E68: Treasury respondents reported AI use in AML/CFT and sanctions compliance, including anomaly detection, suspicious-activity support, identity verification and report creation.	Open source
29	Bank of England - AI in the financial system	Publisher: Bank of England; Kind: Regulator; Checked: 2026-06-19; Status: verified	E32: The Bank of England's Financial Stability in Focus paper treats AI adoption as a potential financial-stability issue when common models, data or vendors influence core financial decisions. \| E33: The Bank of England identified credit and insurance underwriting, market functioning, cyber, operational resilience and third-party dependency as AI risk areas. \| E69: The Bank of England warns that common weaknesses in widely used models could produce correlated mispricing or credit misallocation.	Open source
30	Bank of England and FCA - AI in UK financial services 2024	Publisher: Bank of England and FCA; Kind: Regulator; Checked: 2026-06-19; Status: verified	E34: The Bank of England and FCA 2024 survey found broad AI adoption in UK financial services, material use of foundation models and growing third-party implementation exposure.	Open source
31	FINRA Regulatory Notice 24-09	Publisher: FINRA Regulatory Notice 24-09; Kind: Regulator; Checked: 2026-06-19; Status: verified	E35: FINRA Regulatory Notice 24-09 reminds member firms that securities laws and FINRA rules remain technology neutral when firms use generative AI or LLMs. \| E61: FINRA identifies uses such as policy lookup, research summaries, SEC filings, earnings-call transcripts and surveillance reports as promising but still subject to existing obligations.	Open source
32	FINRA 2026 Annual Regulatory Oversight Report - GenAI	Publisher: FINRA 2026 Annual Regulatory Oversight Report; Kind: Regulator; Checked: 2026-06-19; Status: verified	E36: FINRA's 2026 GenAI materials flag agent risks around system access, data handling, human oversight, action tracking and guardrails.	Open source
33	SEC AI washing enforcement release	Publisher: SEC; Kind: Regulator; Checked: 2026-06-19; Status: verified	E37: The SEC's AI washing actions charged investment advisers over allegedly false and misleading claims about AI use, making model-claim substantiation a conduct risk. \| E62: SEC AI washing enforcement makes public model and product claims part of compliance evidence, not just marketing review.	Open source
34	MAS Guidelines for AI Risk Management	Publisher: MAS Guidelines for AI Risk Management; Kind: Regulator; Checked: 2026-06-19; Status: verified	E38: MAS proposed AI risk-management guidelines for financial institutions covering oversight, AI inventories, materiality assessments, lifecycle controls, evaluation, monitoring and change management. \| E59: MAS's proposed guidance calls for AI inventories and materiality assessments, which makes route registration a governance object.	Open source
35	MAS artificial intelligence programme page	Publisher: MAS artificial intelligence programme page; Kind: Regulator; Checked: 2026-06-19; Status: verified	E39: MAS's AI programme page links financial-sector adoption to shared knowledge, AI component development, workforce skills and AI risk-management guidance.	Open source
36	IOSCO - AI in capital markets	Publisher: IOSCO; Kind: Regulator; Checked: 2026-06-19; Status: verified	E40: IOSCO's AI in capital markets report identifies use cases in decision support, robo-advice, algorithmic trading, research, surveillance and AML/CFT. \| E41: IOSCO highlights model and data risk, malicious use, outsourcing, third-party dependency and human-AI interaction risks in capital markets. \| E70: IOSCO points to algorithmic trading, investment research and surveillance use cases where AI output can affect market integrity.	Open source
37	EU AI Act Service Desk - Annex III	Publisher: EU AI Act Service Desk; Kind: Regulation; Checked: 2026-06-19; Status: verified	E42: EU AI Act Annex III lists creditworthiness and credit scoring for natural persons, excluding fraud detection, as high-risk AI systems. \| E43: EU AI Act Annex III also lists risk assessment and pricing for life and health insurance as high-risk when applied to natural persons. \| E63: EU high-risk classification attaches to intended use, so a general model becomes high-risk when deployed into covered credit or insurance decisions.	Open source
38	EUR-Lex - Digital Operational Resilience Act	Publisher: EUR-Lex; Kind: Regulation; Checked: 2026-06-19; Status: verified	E44: DORA creates ICT risk-management, third-party and resilience obligations for EU financial entities that can capture AI cloud, model, data and agent infrastructure routes. \| E64: DORA's concentration and ICT third-party logic makes a shared model or cloud dependency a resilience question for financial entities.	Open source
39	EBA report on big data and advanced analytics	Publisher: EBA report on big data and advanced analytics; Kind: Regulator; Checked: 2026-06-19; Status: verified	E45: The EBA's big data and advanced analytics report treats ML and advanced analytics as part of banking transformation and emphasizes reliability of data sources and trust elements.	Open source
40	Cyber Risk Institute - Financial Services AI RMF	Publisher: Cyber Risk Institute; Kind: Risk framework; Checked: 2026-06-19; Status: verified	E46: The Cyber Risk Institute FS AI RMF aligns a financial-sector AI control framework with NIST AI RMF and adds sector-specific control objectives.	Open source
41	NIST AI Risk Management Framework	Publisher: NIST; Kind: Risk framework; Checked: 2026-06-19; Status: verified	E47: NIST AI RMF provides the govern, map, measure and manage pattern that finance teams can adapt to model-route approval.	Open source
42	NIST Generative AI Profile	Publisher: NIST; Kind: Risk framework; Checked: 2026-06-19; Status: verified	E48: NIST's Generative AI Profile adds risks and controls specific to generative AI, including confabulation, data privacy, security and misuse concerns. \| E71: NIST's Generative AI Profile names risks that still lack clean financial benchmarks, including confabulation, data leakage, misuse and insufficient human oversight.	Open source
43	CFPB circular on adverse action notices and credit models	Publisher: CFPB circular on adverse action notices and credit models; Kind: Regulator; Checked: 2026-06-19; Status: verified	E49: CFPB guidance on adverse-action notices warns creditors that complex algorithms do not remove the obligation to give specific and accurate reasons for credit decisions.	Open source
44	FSOC 2024 annual report	Publisher: FSOC 2024 annual report; Kind: Regulator; Checked: 2026-06-19; Status: verified	E50: FSOC's 2024 annual report identifies AI as a financial-stability watch item and points to cyber, fraud, model, operational and third-party risks.	Open source
45	BIS Financial Stability Institute - AI and model risk	Publisher: BIS Financial Stability Institute; Kind: Regulator; Checked: 2026-06-19; Status: monitor	E51: BIS Financial Stability Institute work on AI and model risk discusses the explainability trade-off and governance challenge of complex AI models. Limitation: verify dated release or consultation status before relying on timing.	Open source
46	OECD - AI in finance	Publisher: OECD; Kind: Policy framework; Checked: 2026-06-19; Status: monitor	E52: OECD materials on AI, machine learning and big data in finance frame AI risks around market concentration, explainability, fairness and financial stability. Limitation: verify dated release or consultation status before relying on timing.	Open source
47	OpenAI - API data controls	Publisher: OpenAI; Kind: Privacy and data terms; Checked: 2026-06-19; Status: monitor	E53: OpenAI's API data controls describe endpoint retention, abuse monitoring and training-use controls that matter for financial-data classification. Limitation: verify dated release or consultation status before relying on timing.	Open source
48	Microsoft Copilot for Finance	Publisher: Microsoft; Kind: Vendor documentation; Checked: 2026-06-19; Status: monitor	E54: Microsoft Copilot for Finance targets finance workflows inside Microsoft 365, reinforcing the productivity-suite route for lower-risk finance operations. Limitation: verify dated release or consultation status before relying on timing.	Open source
49	BloombergGPT paper	Publisher: BloombergGPT paper; Kind: Financial model paper; Checked: 2026-06-19; Status: monitor	E73: BloombergGPT describes a 50B-parameter finance model trained on a mixed corpus with 363B financial tokens and 345B general-purpose tokens. \| E74: BloombergGPT evaluated against financial, general and internal benchmarks, showing that domain data can improve financial-task performance without eliminating evaluation burden. Limitation: verify dated release or consultation status before relying on timing.	Open source
50	FinGPT paper	Publisher: FinGPT paper; Kind: Financial model paper; Checked: 2026-06-19; Status: monitor	E75: FinGPT frames finance as temporally sensitive, dynamic and low signal-to-noise, arguing for data-centric financial LLM development and adaptation. Limitation: verify dated release or consultation status before relying on timing.	Open source
51	FinGPT GitHub	Publisher: FinGPT GitHub; Kind: GitHub; Checked: 2026-06-19; Status: monitor	E76: The FinGPT repository exposes an open-source financial LLM ecosystem with training pipelines, benchmark material and model releases. Limitation: verify dated release or consultation status before relying on timing.	Open source
52	FinBERT paper	Publisher: FinBERT paper; Kind: Financial model paper; Checked: 2026-06-19; Status: monitor	E77: FinBERT demonstrates a narrower finance-specialized model lane for sentiment classification rather than broad reasoning or agentic work. Limitation: verify dated release or consultation status before relying on timing.	Open source
53	Financial PhraseBank	Publisher: Financial PhraseBank; Kind: Financial dataset; Checked: 2026-06-19; Status: monitor	E78: Financial PhraseBank provides labelled financial-sentence sentiment data, illustrating how narrow financial tasks depend on task-specific labels. Limitation: verify dated release or consultation status before relying on timing.	Open source
54	SEC EDGAR data access	Publisher: SEC; Kind: Official data source; Checked: 2026-06-19; Status: monitor	E79: SEC EDGAR data access documentation describes dissemination, corrections, deletions, indexes and public filing access rules that affect filing-RAG freshness. Limitation: verify dated release or consultation status before relying on timing.	Open source
55	SEC EDGAR APIs	Publisher: SEC; Kind: Official data source; Checked: 2026-06-19; Status: monitor	E80: SEC EDGAR APIs provide JSON submissions history and XBRL company facts, making structured public filings a first-class source path. Limitation: verify dated release or consultation status before relying on timing.	Open source
56	SEC EDGAR full-text search	Publisher: SEC; Kind: Official data source; Checked: 2026-06-19; Status: monitor	E81: SEC full-text search covers EDGAR filings submitted electronically since 2001, supporting evidence discovery but not private entitlement controls. Limitation: verify dated release or consultation status before relying on timing.	Open source
57	Federal Reserve FRED API	Publisher: Federal Reserve FRED API; Kind: Official data source; Checked: 2026-06-19; Status: monitor	E82: FRED exposes macroeconomic time series through an API, which matters for models that mix company, market and macro context. Limitation: verify dated release or consultation status before relying on timing.	Open source
58	U.S. Bureau of Economic Analysis API	Publisher: U.S. Bureau of Economic Analysis API; Kind: Official data source; Checked: 2026-06-19; Status: monitor	E83: BEA API access supports official national-account and economic data paths for macro and sector workflows. Limitation: verify dated release or consultation status before relying on timing.	Open source
59	GLEIF API	Publisher: GLEIF API; Kind: Official data source; Checked: 2026-06-19; Status: monitor	E84: GLEIF API documentation supports entity identity, LEI search and parent-relationship data for counterparty and ownership workflows. Limitation: verify dated release or consultation status before relying on timing.	Open source
60	OFAC sanctions list search	Publisher: OFAC sanctions list search; Kind: Official data source; Checked: 2026-06-19; Status: monitor	E85: OFAC sanctions list search uses approximate string matching, making sanctions screening a source and matching-control problem. Limitation: verify dated release or consultation status before relying on timing.	Open source
61	OFAC sanctions list service	Publisher: OFAC sanctions list service; Kind: Official data source; Checked: 2026-06-19; Status: monitor	E86: OFAC sanctions list service gives compliance teams a direct official route to sanctions list data rather than relying on stale copies. Limitation: verify dated release or consultation status before relying on timing.	Open source
62	FinCEN BSA E-Filing	Publisher: FinCEN BSA E-Filing; Kind: Official data source; Checked: 2026-06-19; Status: monitor	E87: FinCEN BSA E-Filing is the official filing path for BSA reports, so AI-generated case support must preserve filing accountability. Limitation: verify dated release or consultation status before relying on timing.	Open source
63	FinCEN innovation and technology	Publisher: FinCEN innovation and technology; Kind: Regulator; Checked: 2026-06-19; Status: monitor	E88: FinCEN's innovation programme shows financial-crime technology adoption is supervised through regulatory engagement, not vendor claims alone. Limitation: verify dated release or consultation status before relying on timing.	Open source
64	FATF - opportunities and challenges of new technologies for AML/CFT	Publisher: FATF; Kind: Policy framework; Checked: 2026-06-19; Status: monitor	E89: FATF's new-technology work for AML/CFT frames AI as both a compliance opportunity and a control challenge. Limitation: verify dated release or consultation status before relying on timing.	Open source
65	FATF Recommendations	Publisher: FATF Recommendations; Kind: Policy framework; Checked: 2026-06-19; Status: monitor	E90: FATF Recommendations anchor AML/CFT obligations that AI screening, monitoring and case-support systems must serve. Limitation: verify dated release or consultation status before relying on timing.	Open source
66	Wolfsberg Group financial crime publications	Publisher: Wolfsberg Group financial crime publications; Kind: Policy framework; Checked: 2026-06-19; Status: monitor	E91: Wolfsberg publications provide financial-crime control expectations that matter when AI touches KYC, transaction monitoring or investigations. Limitation: verify dated release or consultation status before relying on timing.	Open source
67	Bank of England fintech and AI work	Publisher: Bank of England fintech and AI work; Kind: Regulator; Checked: 2026-06-19; Status: monitor	E92: The Bank of England's fintech and AI work tracks AI adoption, surveys and financial-stability implications for UK-regulated firms. Limitation: verify dated release or consultation status before relying on timing.	Open source
68	FCA AI update	Publisher: FCA AI update; Kind: Regulator; Checked: 2026-06-19; Status: monitor	E93: The FCA AI page positions AI as an existing financial-services supervisory issue, including consumer, competition and market-integrity concerns. Limitation: verify dated release or consultation status before relying on timing.	Open source
69	Bank of England and FCA AI public-private forum	Publisher: Bank of England and FCA AI public-private forum; Kind: Regulator; Checked: 2026-06-19; Status: monitor	E94: The Bank of England and FCA AI Public-Private Forum provides a supervisory framework for responsible AI use in financial services. Limitation: verify dated release or consultation status before relying on timing.	Open source
70	UK critical third parties policy statement	Publisher: UK critical third parties policy statement; Kind: Regulation; Checked: 2026-06-19; Status: monitor	E95: UK critical-third-party policy makes concentration and resilience relevant when AI depends on shared cloud, model or data providers. Limitation: verify dated release or consultation status before relying on timing.	Open source
71	European Supervisory Authorities DORA technical standards	Publisher: European Supervisory Authorities DORA technical standards; Kind: Regulation; Checked: 2026-06-19; Status: monitor	E96: European Supervisory Authorities' DORA work maps ICT risk, incident reporting, testing and third-party oversight into AI infrastructure routes. Limitation: verify dated release or consultation status before relying on timing.	Open source
72	EBA machine learning for IRB models report	Publisher: EBA machine learning for IRB models report; Kind: Regulator; Checked: 2026-06-19; Status: monitor	E97: EBA work on machine learning for IRB models shows that banking model validation remains technical and use-case specific. Limitation: verify dated release or consultation status before relying on timing.	Open source
73	ESMA statement on AI in investment services	Publisher: ESMA statement on AI in investment services; Kind: Regulator; Checked: 2026-06-19; Status: monitor	E98: ESMA's AI statement for investment firms links AI use to client-interest, governance, staff competence and disclosure duties. Limitation: verify dated release or consultation status before relying on timing.	Open source
74	EIOPA AI governance principles	Publisher: EIOPA AI governance principles; Kind: Regulator; Checked: 2026-06-19; Status: monitor	E99: EIOPA AI governance principles tie insurance AI to fairness, explainability, governance and consumer-outcome controls. Limitation: verify dated release or consultation status before relying on timing.	Open source
75	ECB banking supervision technology and innovation	Publisher: ECB banking supervision technology and innovation; Kind: Regulator; Checked: 2026-06-19; Status: monitor	E100: ECB supervisory technology priorities keep digital, data and technology risk inside banking supervision rather than outside it. Limitation: verify dated release or consultation status before relying on timing.	Open source
76	Basel Committee newsletter on AI and machine learning	Publisher: Basel Committee newsletter on AI and machine learning; Kind: Regulator; Checked: 2026-06-19; Status: monitor	E101: The Basel Committee's AI and ML newsletter focused on explainability, governance, accountability and financial-stability implications. Limitation: verify dated release or consultation status before relying on timing.	Open source
77	Basel Committee speech - managing AI in banking	Publisher: Basel Committee speech; Kind: Regulator; Checked: 2026-06-19; Status: monitor	E102: The Basel Committee chair warned that AI/ML in banking has not yet been tested across a full financial cycle and could amplify future crises if unchecked. Limitation: verify dated release or consultation status before relying on timing.	Open source
78	FSB - financial stability implications of AI	Publisher: FSB; Kind: Regulator; Checked: 2026-06-19; Status: monitor	E103: The FSB's AI financial-stability work identifies third-party dependencies, market correlations, cyber risk, model governance and data risk as systemic vulnerabilities. Limitation: verify dated release or consultation status before relying on timing.	Open source
79	FSB - monitoring AI adoption and vulnerabilities	Publisher: FSB; Kind: Regulator; Checked: 2026-06-19; Status: monitor	E104: The FSB's monitoring report tracks AI adoption and vulnerabilities, including supply-chain concentration and indicators for supervisory monitoring. Limitation: verify dated release or consultation status before relying on timing.	Open source
80	IMF Global Financial Stability Report October 2024	Publisher: IMF Global Financial Stability Report October 2024; Kind: Policy framework; Checked: 2026-06-19; Status: monitor	E105: The IMF Global Financial Stability Report discusses AI implications for capital markets, including trading, advice, productivity and stability risks. Limitation: verify dated release or consultation status before relying on timing.	Open source
81	IOSCO 2021 AI and ML guidance	Publisher: IOSCO 2021 AI and ML guidance; Kind: Regulator; Checked: 2026-06-19; Status: monitor	E106: IOSCO's earlier AI and ML guidance for market intermediaries remains relevant for governance, testing, outsourcing and controls. Limitation: verify dated release or consultation status before relying on timing.	Open source
82	FS-ISAC AI risk and resilience resources	Publisher: FS-ISAC AI risk and resilience resources; Kind: Risk framework; Checked: 2026-06-19; Status: monitor	E107: FS-ISAC AI resources make cyber and operational resilience part of financial AI governance, not a separate security appendix. Limitation: verify dated release or consultation status before relying on timing.	Open source
83	OECD AI principles	Publisher: OECD AI principles; Kind: Policy framework; Checked: 2026-06-19; Status: monitor	E108: OECD AI principles provide cross-sector governance expectations that financial regulators adapt rather than ignore. Limitation: verify dated release or consultation status before relying on timing.	Open source
84	OSFI Guideline E-23 2027 letter	Publisher: OSFI Guideline E-23 2027 letter; Kind: Regulator; Checked: 2026-06-19; Status: monitor	E109: OSFI's E-23 2027 letter marks a move toward broader model-risk expectations in Canadian financial institutions. Limitation: verify dated release or consultation status before relying on timing.	Open source
85	OSFI Guideline E-23 model risk management 2027	Publisher: OSFI Guideline E-23 model risk management 2027; Kind: Regulator; Checked: 2026-06-19; Status: monitor	E110: OSFI E-23 applies enterprise-wide model-risk expectations across internally and externally sourced models, including AI and ML methods. Limitation: verify dated release or consultation status before relying on timing.	Open source
86	RBI FREE-AI committee report	Publisher: RBI FREE-AI committee report; Kind: Regulator; Checked: 2026-06-19; Status: monitor	E111: RBI's FREE-AI report frames responsible and ethical AI for the financial sector through governance, fairness, accountability and explainability principles. Limitation: verify dated release or consultation status before relying on timing.	Open source
87	HKMA AI guidance and circulars	Publisher: HKMA AI guidance and circulars; Kind: Regulator; Checked: 2026-06-19; Status: monitor	E112: HKMA AI guidance and fintech work provide an Asia-Pacific supervisory route for AI adoption in banking and finance. Limitation: verify dated release or consultation status before relying on timing.	Open source
88	ASIC AI and automated decision-making	Publisher: ASIC AI and automated decision-making; Kind: Regulator; Checked: 2026-06-19; Status: monitor	E113: ASIC's AI and automated decision-making materials make governance, accountability and consumer impact central for Australian financial firms. Limitation: verify dated release or consultation status before relying on timing.	Open source
89	APRA technology and operational resilience	Publisher: APRA technology and operational resilience; Kind: Regulator; Checked: 2026-06-19; Status: monitor	E114: APRA operational-resilience materials connect model, cloud, technology and outsourcing choices to continuity expectations. Limitation: verify dated release or consultation status before relying on timing.	Open source
90	MAS FEAT principles	Publisher: MAS FEAT principles; Kind: Policy framework; Checked: 2026-06-19; Status: monitor	E115: MAS FEAT principles anchor fairness, ethics, accountability and transparency for AI and data analytics in financial services. Limitation: verify dated release or consultation status before relying on timing.	Open source
91	MAS Veritas initiative	Publisher: MAS Veritas initiative; Kind: Policy framework; Checked: 2026-06-19; Status: monitor	E116: MAS Veritas provides assessment methodology and tooling direction for financial AI fairness, ethics, accountability and transparency. Limitation: verify dated release or consultation status before relying on timing.	Open source
92	ISO/IEC 42001 AI management system	Publisher: ISO/IEC; Kind: Standard; Checked: 2026-06-19; Status: monitor	E117: ISO/IEC 42001 gives firms an auditable AI management-system standard that can support governance evidence across model routes. Limitation: verify dated release or consultation status before relying on timing.	Open source
93	Bloomberg AI and data science	Publisher: Bloomberg AI and data science; Kind: Vendor documentation; Checked: 2026-06-19; Status: monitor	E118: Bloomberg's AI and data-science materials show financial AI value depends on proprietary financial data, workflow integration and domain expertise. Limitation: verify dated release or consultation status before relying on timing.	Open source
94	BloombergGPT press release	Publisher: BloombergGPT press release; Kind: Press release; Checked: 2026-06-19; Status: monitor	E119: BloombergGPT's press release positioned the model as finance-tuned, reinforcing the market value of high-quality financial-domain corpora. Limitation: verify dated release or consultation status before relying on timing.	Open source
95	LSEG Workspace AI	Publisher: LSEG Workspace AI; Kind: Vendor documentation; Checked: 2026-06-19; Status: monitor	E120: LSEG Workspace represents a data-terminal route where model output depends on licensed market and company data access. Limitation: verify dated release or consultation status before relying on timing.	Open source
96	FactSet Mercury	Publisher: FactSet Mercury; Kind: Vendor documentation; Checked: 2026-06-19; Status: monitor	E121: FactSet Mercury shows financial-search and workflow AI embedded inside a licensed data and analytics platform. Limitation: verify dated release or consultation status before relying on timing.	Open source
97	S&P Global Kensho	Publisher: S&P Global Kensho; Kind: Vendor documentation; Checked: 2026-06-19; Status: monitor	E122: S&P Global Kensho shows a finance-specific AI lane built around transcription, extraction, classification and domain datasets. Limitation: verify dated release or consultation status before relying on timing.	Open source
98	S&P Global Capital IQ Pro	Publisher: S&P Global Capital IQ Pro; Kind: Vendor documentation; Checked: 2026-06-19; Status: monitor	E123: Capital IQ Pro is a source-access lane for company, market and transaction data that an LLM route may need but does not own. Limitation: verify dated release or consultation status before relying on timing.	Open source
99	AlphaSense generative search	Publisher: AlphaSense generative search; Kind: Vendor documentation; Checked: 2026-06-19; Status: monitor	E124: AlphaSense generative search highlights the specialist-search route for broker research, filings, transcripts and expert-call corpora. Limitation: verify dated release or consultation status before relying on timing.	Open source
100	PitchBook AI search	Publisher: PitchBook AI search; Kind: Vendor documentation; Checked: 2026-06-19; Status: monitor	E125: PitchBook is a private-market data path where source rights and coverage affect model answers more than model fluency. Limitation: verify dated release or consultation status before relying on timing.	Open source
101	Moody's research assistant	Publisher: Moody's research assistant; Kind: Vendor documentation; Checked: 2026-06-19; Status: monitor	E126: Moody's research assistant demonstrates AI wrapped around credit research and ratings-adjacent knowledge products. Limitation: verify dated release or consultation status before relying on timing.	Open source
102	MSCI AI and analytics	Publisher: MSCI AI and analytics; Kind: Vendor documentation; Checked: 2026-06-19; Status: monitor	E127: MSCI analytics show investment and risk workflows depend on licensed indexes, models, factors and portfolio data. Limitation: verify dated release or consultation status before relying on timing.	Open source
103	Nasdaq market surveillance	Publisher: Nasdaq market surveillance; Kind: Vendor documentation; Checked: 2026-06-19; Status: monitor	E128: Nasdaq market surveillance represents a market-integrity workflow where AI support must preserve alerts, evidence and supervisory review. Limitation: verify dated release or consultation status before relying on timing.	Open source
104	NICE Actimize Xceed AI agents	Publisher: NICE Actimize Xceed AI agents; Kind: Vendor documentation; Checked: 2026-06-19; Status: monitor	E129: NICE Actimize introduced AI agents for fraud and financial crime workflows, including alert triage, backlog categorization and case summarization. Limitation: verify dated release or consultation status before relying on timing.	Open source
105	NICE Actimize Xceed AI FRAML	Publisher: NICE Actimize Xceed AI FRAML; Kind: Vendor documentation; Checked: 2026-06-19; Status: monitor	E130: NICE Actimize Xceed AI FRAML combines fraud and AML workflows, showing the convergence of detection, case management and analyst-assist agents. Limitation: verify dated release or consultation status before relying on timing.	Open source
106	SAS anti-money laundering	Publisher: SAS anti-money laundering; Kind: Vendor documentation; Checked: 2026-06-19; Status: monitor	E131: SAS AML materials show financial-crime AI is a monitoring and case-management route rather than a general chatbot use case. Limitation: verify dated release or consultation status before relying on timing.	Open source
107	FICO Falcon Fraud Manager	Publisher: FICO Falcon Fraud Manager; Kind: Vendor documentation; Checked: 2026-06-19; Status: monitor	E132: FICO Falcon Fraud Manager anchors the mature fraud-model lane where real-time scoring and drift monitoring matter. Limitation: verify dated release or consultation status before relying on timing.	Open source
108	Featurespace ARIC Risk Hub	Publisher: Featurespace ARIC Risk Hub; Kind: Vendor documentation; Checked: 2026-06-19; Status: monitor	E133: Featurespace ARIC Risk Hub represents adaptive behavioural analytics for fraud and financial-crime risk. Limitation: verify dated release or consultation status before relying on timing.	Open source
109	Feedzai financial crime management	Publisher: Feedzai financial crime management; Kind: Vendor documentation; Checked: 2026-06-19; Status: monitor	E134: Feedzai shows the platform route for financial-crime risk, combining data pipelines, models, case workflows and monitoring. Limitation: verify dated release or consultation status before relying on timing.	Open source
110	Mastercard Decision Intelligence	Publisher: Mastercard Decision Intelligence; Kind: Vendor documentation; Checked: 2026-06-19; Status: monitor	E135: Mastercard Decision Intelligence illustrates network-level fraud AI where the data and operating network are part of the model advantage. Limitation: verify dated release or consultation status before relying on timing.	Open source
111	Visa Advanced Authorization	Publisher: Visa Advanced Authorization; Kind: Vendor documentation; Checked: 2026-06-19; Status: monitor	E136: Visa authorization and fraud tooling shows payment-network AI depends on real-time transaction context and network data. Limitation: verify dated release or consultation status before relying on timing.	Open source
112	Shift Technology insurance AI	Publisher: Shift Technology insurance AI; Kind: Vendor documentation; Checked: 2026-06-19; Status: monitor	E137: Shift Technology represents the insurance fraud and claims AI lane, which differs from banking, capital-markets and finance-ops workflows. Limitation: verify dated release or consultation status before relying on timing.	Open source
113	Tractable AI insurance claims	Publisher: Tractable AI insurance claims; Kind: Vendor documentation; Checked: 2026-06-19; Status: monitor	E138: Tractable illustrates computer-vision and claims automation as a separate insurance model route from text LLMs. Limitation: verify dated release or consultation status before relying on timing.	Open source
114	Guidewire predictive analytics	Publisher: Guidewire predictive analytics; Kind: Vendor documentation; Checked: 2026-06-19; Status: monitor	E139: Guidewire predictive analytics shows insurance model routes often live inside core insurance platforms and data models. Limitation: verify dated release or consultation status before relying on timing.	Open source
115	Workiva generative AI	Publisher: Workiva generative AI; Kind: Vendor documentation; Checked: 2026-06-19; Status: monitor	E140: Workiva generative AI sits inside reporting, audit and compliance workflows where evidence and approval trails matter. Limitation: verify dated release or consultation status before relying on timing.	Open source
116	BlackLine financial close automation	Publisher: BlackLine financial close automation; Kind: Vendor documentation; Checked: 2026-06-19; Status: monitor	E141: BlackLine financial-close automation shows a finance-operations route where reconciliations, exceptions and approvals define success. Limitation: verify dated release or consultation status before relying on timing.	Open source
117	Oracle Fusion Cloud ERP AI	Publisher: Oracle Fusion Cloud ERP AI; Kind: Vendor documentation; Checked: 2026-06-19; Status: monitor	E142: Oracle application AI places finance AI inside ERP records, approvals and operating workflows. Limitation: verify dated release or consultation status before relying on timing.	Open source
118	SAP Joule finance	Publisher: SAP Joule finance; Kind: Vendor documentation; Checked: 2026-06-19; Status: monitor	E143: SAP Joule represents enterprise-app AI where model value depends on access to governed business objects and process permissions. Limitation: verify dated release or consultation status before relying on timing.	Open source
119	Databricks financial services	Publisher: Databricks; Kind: Cloud documentation; Checked: 2026-06-19; Status: monitor	E144: Databricks financial-services materials position the lakehouse as a controlled data and model-development route for finance AI. Limitation: verify dated release or consultation status before relying on timing.	Open source
120	Snowflake AI Data Cloud for Financial Services	Publisher: Snowflake AI Data Cloud for Financial Services; Kind: Cloud documentation; Checked: 2026-06-19; Status: monitor	E145: Snowflake financial-services materials emphasize governance, sensitive-data controls, resiliency and financial-services compliance standards. Limitation: verify dated release or consultation status before relying on timing.	Open source
121	Anthropic financial services marketplace	Publisher: Anthropic; Kind: GitHub; Checked: 2026-06-19; Status: monitor	E146: Anthropic's financial-services marketplace exposes reusable agent templates, showing how provider-supplied patterns can speed deployment but still need local approval. Limitation: verify dated release or consultation status before relying on timing.	Open source
122	Vals AI finance agent dataset	Publisher: Vals AI finance agent dataset; Kind: Agent benchmark; Checked: 2026-06-19; Status: monitor	E147: The Finance Agent dataset contains expert-authored questions across financial task categories, supporting agentic eval design. Limitation: verify dated release or consultation status before relying on timing.	Open source
123	FinanceQA dataset	Publisher: FinanceQA dataset; Kind: Financial benchmark; Checked: 2026-06-19; Status: monitor	E148: FinanceQA targets investment-style numerical analysis and assumption handling, filling part of the gap between filing QA and analyst work. Limitation: verify dated release or consultation status before relying on timing.	Open source
124	FinQA paper	Publisher: FinQA paper; Kind: Financial benchmark; Checked: 2026-06-19; Status: monitor	E149: FinQA focuses on numerical reasoning over financial reports and remains a useful task family for private eval design. Limitation: verify dated release or consultation status before relying on timing.	Open source
125	DocFinQA dataset	Publisher: DocFinQA dataset; Kind: Financial benchmark; Checked: 2026-06-19; Status: monitor	E150: DocFinQA adds document-level financial question answering, relevant to long reports and multi-page source packs. Limitation: verify dated release or consultation status before relying on timing.	Open source
126	MultiHiertt benchmark	Publisher: MultiHiertt benchmark; Kind: Financial benchmark; Checked: 2026-06-19; Status: monitor	E151: MultiHiertt tests hierarchical table-and-text reasoning, a common failure mode in financial statements and disclosures. Limitation: verify dated release or consultation status before relying on timing.	Open source
127	Numerical reasoning over financial reports survey	Publisher: Numerical reasoning over financial reports survey; Kind: Research paper; Checked: 2026-06-19; Status: monitor	E152: Research on numerical reasoning over financial reports shows that arithmetic, units and source structure remain benchmark gaps. Limitation: verify dated release or consultation status before relying on timing.	Open source
128	EDGAR-CORPUS	Publisher: EDGAR-CORPUS; Kind: Financial dataset; Checked: 2026-06-19; Status: monitor	E153: EDGAR-CORPUS shows how public filings can become a training or evaluation corpus, but licensing and currentness still need review. Limitation: verify dated release or consultation status before relying on timing.	Open source
129	Loughran-McDonald sentiment word lists	Publisher: Loughran-McDonald sentiment word lists; Kind: Financial dataset; Checked: 2026-06-19; Status: monitor	E154: Loughran-McDonald word lists remain a finance-specific lexical baseline for sentiment and disclosure analysis. Limitation: verify dated release or consultation status before relying on timing.	Open source
130	Nasdaq Data Link	Publisher: Nasdaq Data Link; Kind: Financial data provider; Checked: 2026-06-19; Status: monitor	E155: Nasdaq Data Link illustrates third-party market-data paths where entitlements and licensing shape what a model may retrieve. Limitation: verify dated release or consultation status before relying on timing.	Open source
131	OpenFIGI API	Publisher: OpenFIGI API; Kind: Official data source; Checked: 2026-06-19; Status: monitor	E156: OpenFIGI supports instrument identifier mapping, a necessary control for multi-source security and issuer workflows. Limitation: verify dated release or consultation status before relying on timing.	Open source
132	ISO 20022	Publisher: ISO 20022; Kind: Standard; Checked: 2026-06-19; Status: monitor	E157: ISO 20022 creates structured message semantics for payments workflows where LLMs should not invent fields or meanings. Limitation: verify dated release or consultation status before relying on timing.	Open source
133	Swift AI and data	Publisher: Swift AI and data; Kind: Vendor documentation; Checked: 2026-06-19; Status: monitor	E158: Swift data and analytics routes place AI in payment and messaging contexts with strong controls over data, identifiers and monitoring. Limitation: verify dated release or consultation status before relying on timing.	Open source
134	ISDA Common Domain Model	Publisher: ISDA Common Domain Model; Kind: Standard; Checked: 2026-06-19; Status: monitor	E159: ISDA's Common Domain Model gives derivatives workflows a machine-readable representation that can constrain AI extraction and action. Limitation: verify dated release or consultation status before relying on timing.	Open source
135	DTCC data services	Publisher: DTCC data services; Kind: Financial data provider; Checked: 2026-06-19; Status: monitor	E160: DTCC data services show post-trade and reference-data paths where source rights, timeliness and reconciliation matter. Limitation: verify dated release or consultation status before relying on timing.	Open source
136	FIX Trading Community standards	Publisher: FIX Trading Community standards; Kind: Standard; Checked: 2026-06-19; Status: monitor	E161: FIX standards constrain trading-message meaning and reduce the room for free-form generation in market workflows. Limitation: verify dated release or consultation status before relying on timing.	Open source
137	OpenAI Evals	Publisher: OpenAI Evals; Kind: Evaluation framework; Checked: 2026-06-19; Status: monitor	E162: OpenAI Evals provides an evaluation framework pattern for repeatable route tests and regression checks. Limitation: verify dated release or consultation status before relying on timing.	Open source
138	OpenAI status	Publisher: OpenAI status; Kind: Status page; Checked: 2026-06-19; Status: monitor	E163: OpenAI's status page showed APIs at 99.98% uptime for its Feb 2026-May 2026 published window on 27 May 2026; its incident feed showed the busiest API/model-relevant seven-day cluster at about 36.1 hours during 13-20 May 2026. Limitation: verify dated release or consultation status before relying on timing.	Open source
139	Claude status	Publisher: Claude status; Kind: Status page; Checked: 2026-06-19; Status: monitor	E164: Claude's status page showed Claude API at 99.07% uptime over the past 90 days on 27 May 2026; its incident feed showed the busiest model/API-relevant seven-day cluster at about 13.7 hours during 16-23 May 2026. Limitation: verify dated release or consultation status before relying on timing.	Open source
140	Google Cloud status	Publisher: Google; Kind: Status page; Checked: 2026-06-19; Status: monitor	E165: Google Cloud Service Health records a Vertex AI Gemini API incident on 27 February 2026 in which global-endpoint customers experienced increased error rates for 1 hour and 58 minutes. Limitation: verify dated release or consultation status before relying on timing.	Open source
141	AWS Health Dashboard	Publisher: AWS; Kind: Status page; Checked: 2026-06-19; Status: monitor	E166: AWS Health status matters for Bedrock, Textract and AWS-hosted financial AI routes. Limitation: verify dated release or consultation status before relying on timing.	Open source
142	Azure status history	Publisher: Azure; Kind: Status page; Checked: 2026-06-19; Status: monitor	E167: Azure status history records an Azure OpenAI GPT-5.2 availability degradation across seven regions from 23:20 UTC on 9 March to 19:32 UTC on 10 March 2026, a 20.2-hour incident window. Limitation: verify dated release or consultation status before relying on timing.	Open source
143	NVD vulnerability database	Publisher: NVD vulnerability database; Kind: Security advisory; Checked: 2026-06-19; Status: monitor	E168: NVD vulnerability records can trigger security review for model-serving, parser, vector database and agent tooling components. Limitation: verify dated release or consultation status before relying on timing.	Open source
144	GitHub security advisories	Publisher: GitHub security advisories; Kind: Security advisory; Checked: 2026-06-19; Status: monitor	E169: GitHub security advisories surface dependency vulnerabilities that may affect open-source model gateways, parsers, orchestration and eval stacks. Limitation: verify dated release or consultation status before relying on timing.	Open source
145	Hugging Face model hub	Publisher: Hugging Face; Kind: Model hub; Checked: 2026-06-19; Status: monitor	E170: Hugging Face model hub changes, model cards and downloads provide monitoring signals for open-weight financial model routes. Limitation: verify dated release or consultation status before relying on timing.	Open source
146	MLCommons MLPerf Inference	Publisher: MLCommons; Kind: Benchmark; Checked: 2026-06-19; Status: monitor	E171: MLPerf inference benchmarks provide hardware and serving-efficiency signals, useful for self-hosted economics but not finance quality. Limitation: verify dated release or consultation status before relying on timing.	Open source
147	Stanford HELM	Publisher: Stanford HELM; Kind: Evaluation framework; Checked: 2026-06-19; Status: monitor	E172: Stanford HELM gives general model-evaluation context, useful for safety and capability monitoring outside finance-specific tasks. Limitation: verify dated release or consultation status before relying on timing.	Open source
148	UK AI Security Institute evaluations	Publisher: UK AI Security Institute evaluations; Kind: Evaluation framework; Checked: 2026-06-19; Status: monitor	E173: AI Security Institute evaluations provide frontier-model risk signals that may affect high-risk finance approvals. Limitation: verify dated release or consultation status before relying on timing.	Open source
149	OWASP Top 10 for LLM Applications	Publisher: OWASP; Kind: Security framework; Checked: 2026-06-19; Status: monitor	E174: OWASP LLM guidance names application-layer risks such as prompt injection, data leakage and tool misuse that finance agents must test. Limitation: verify dated release or consultation status before relying on timing.	Open source
150	MITRE ATLAS	Publisher: MITRE ATLAS; Kind: Security framework; Checked: 2026-06-19; Status: monitor	E175: MITRE ATLAS maps adversarial tactics against AI systems, supporting red-team design for finance AI routes. Limitation: verify dated release or consultation status before relying on timing.	Open source
151	CISA AI security	Publisher: CISA; Kind: Security framework; Checked: 2026-06-19; Status: monitor	E176: CISA AI security resources give operational guidance for securing AI systems as part of financial-sector cyber resilience. Limitation: verify dated release or consultation status before relying on timing.	Open source
152	ISO/IEC 23894 AI risk management	Publisher: ISO/IEC; Kind: Standard; Checked: 2026-06-19; Status: monitor	E211: Background-only: ISO/IEC 23894 is an AI risk-management standard useful for risk terminology and control design; no rendered claim depends on it as a live financial-model fact. Limitation: verify dated release or consultation status before relying on timing.	Open source
153	Andrew W. Lo profile	Publisher: Andrew W. Lo profile; Kind: Expert profile; Checked: 2026-06-19; Status: monitor	E212: Andrew W. Lo's MIT Sloan profile supports the expert-watchlist name, MIT Sloan affiliation and market-behaviour monitoring signal. Limitation: verify dated release or consultation status before relying on timing.	Open source
154	Marcos Lopez de Prado profile	Publisher: Marcos Lopez de Prado profile; Kind: Expert profile; Checked: 2026-06-19; Status: monitor	E213: The MLFinLab profile source supports the Marcos Lopez de Prado expert-watchlist entry for financial machine-learning validation, backtesting and leakage-control monitoring. Limitation: verify dated release or consultation status before relying on timing.	Open source
155	Hilary J. Allen profile	Publisher: Hilary J. Allen profile; Kind: Expert profile; Checked: 2026-06-19; Status: monitor	E214: Hilary J. Allen's American University Washington College of Law profile supports the expert-watchlist name, affiliation and financial-stability-law focus. Limitation: verify dated release or consultation status before relying on timing.	Open source
156	Sandra Wachter profile	Publisher: Sandra Wachter profile; Kind: Expert profile; Checked: 2026-06-19; Status: monitor	E215: Sandra Wachter's Oxford Internet Institute profile supports the expert-watchlist name, affiliation and explainability, fairness and algorithmic-accountability focus. Limitation: verify dated release or consultation status before relying on timing.	Open source
157	Daniel Schwarcz profile	Publisher: Daniel Schwarcz profile; Kind: Expert profile; Checked: 2026-06-19; Status: monitor	E216: Daniel Schwarcz's University of Minnesota Law School profile supports the expert-watchlist name, affiliation and insurance-law focus. Limitation: verify dated release or consultation status before relying on timing.	Open source
158	Maureen O'Hara profile	Publisher: Maureen O'Hara profile; Kind: Expert profile; Checked: 2026-06-19; Status: monitor	E217: Maureen O'Hara's Cornell Johnson profile supports the expert-watchlist name, affiliation and market-microstructure focus. Limitation: verify dated release or consultation status before relying on timing.	Open source
159	Gary Gensler MIT profile	Publisher: Gary Gensler MIT profile; Kind: Expert profile; Checked: 2026-06-19; Status: monitor	E218: Background-only: Gary Gensler's MIT Sloan profile is retained as a potential financial-regulation and market-structure monitor, with no rendered expert-watchlist dependency. Limitation: verify dated release or consultation status before relying on timing.	Open source
160	Hui Chen MIT profile	Publisher: Hui Chen MIT profile; Kind: Expert profile; Checked: 2026-06-19; Status: monitor	E219: Background-only: Hui Chen's MIT profile is retained as a potential finance, risk and AI-economics monitor, with no rendered expert-watchlist dependency. Limitation: verify dated release or consultation status before relying on timing.	Open source
161	Adair Morse Berkeley profile	Publisher: Adair Morse Berkeley profile; Kind: Expert profile; Checked: 2026-06-19; Status: monitor	E220: Adair Morse's Berkeley Haas profile supports the expert-watchlist name, affiliation and household-finance or consumer-finance monitoring signal. Limitation: verify dated release or consultation status before relying on timing.	Open source
162	Ariel Ezrachi Oxford profile	Publisher: Ariel Ezrachi Oxford profile; Kind: Expert profile; Checked: 2026-06-19; Status: monitor	E221: Background-only: Ariel Ezrachi's Oxford profile is retained for competition, consumer and platform-governance context, with no rendered expert-watchlist dependency. Limitation: verify dated release or consultation status before relying on timing.	Open source
163	Columbia FinTech and AI research	Publisher: Columbia FinTech and AI research; Kind: Expert profile; Checked: 2026-06-19; Status: monitor	E222: Background-only: Columbia FinTech and AI research is retained as an institutional monitoring source, not as a rendered person-profile dependency. Limitation: verify dated release or consultation status before relying on timing.	Open source
164	AI4Finance research page	Publisher: AI4Finance research page; Kind: Financial model paper; Checked: 2026-06-19; Status: monitor	E223: AI4Finance is background support for open financial-model and FinGPT monitoring; rendered FinGPT claims still rely on the FinGPT paper and repository rows. Limitation: verify dated release or consultation status before relying on timing.	Open source
165	FCA Digital Sandbox	Publisher: FCA Digital Sandbox; Kind: Regulatory sandbox; Checked: 2026-06-19; Status: monitor	E224: Background-only: the FCA Digital Sandbox is a regulated testing environment signal that can inform future route-evidence monitoring. Limitation: verify dated release or consultation status before relying on timing.	Open source
166	MAS PathFin.ai	Publisher: MAS PathFin.ai; Kind: Regulatory sandbox; Checked: 2026-06-19; Status: monitor	E225: Background-only: MAS PathFin.ai is a Singapore financial-AI collaboration and monitoring signal, not a rendered claim dependency. Limitation: verify dated release or consultation status before relying on timing.	Open source
167	Artificial Analysis - model comparison and Intelligence Index cost	Publisher: Artificial Analysis; Kind: Model benchmark; Checked: 2026-06-19; Status: monitor	E177: Artificial Analysis publishes model intelligence, token pricing, speed, latency, context and the cost to run the Artificial Analysis Intelligence Index, making token price and benchmark bill separable cost measures. \| E210: Artificial Analysis checked on 31 May 2026 supplies the exact chart data: blended-price tier winners in that snapshot were Qwen3.5 0.8B non-reasoning at $0.02, Qwen3.5 0.8B reasoning at $0.02, Qwen3.5 4B non-reasoning at $0.06, Qwen3.5 9B reasoning at $0.113, MiMo-V2-Flash at $0.15 and DeepSeek V4 Pro at $0.544; Intelligence Index cost winners in that snapshot were Gemma 3n E4B at $5.7215, Granite 4.0 H Small at $4.4809, gpt-oss-20B low at $7.6768, Gemma 4 31B non-reasoning at $19.4332, DeepSeek V4 Flash high effort at $57.4318 and DeepSeek V4 Pro at $267.8209. Limitation: verify dated release or consultation status before relying on timing.	Open source
168	a16z - LLMflation: from a cost miracle to the margin wars	Publisher: a16z; Kind: Market analysis; Checked: 2026-06-19; Status: monitor	E178: a16z's LLMflation analysis introduced the cumulative intelligence-tier chart method: track the cheapest available model within each intelligence band over time, while keeping the tier framing separate from Artificial Analysis's per-model scores. Limitation: verify dated release or consultation status before relying on timing.	Open source
169	AI Act Service Desk - implementation timeline	Publisher: AI Act Service Desk; Kind: Regulatory timeline; Checked: 2026-06-19; Status: monitor	E181: The AI Act implementation timeline is a dated monitor for EU AI Act obligations, including 2 August 2026 milestones that affect model documentation, transparency and provider diligence. Limitation: verify dated release or consultation status before relying on timing.	Open source
170	European Commission - high-risk AI classification consultation	Publisher: European Commission; Kind: Regulatory consultation; Checked: 2026-06-19; Status: monitor	E180: The European Commission's high-risk AI classification consultation accepts feedback until 23 June 2026 and can affect credit and insurance route classification. Limitation: verify dated release or consultation status before relying on timing.	Open source
171	European Commission - draft AI transparency-obligations guidelines	Publisher: European Commission; Kind: Regulatory consultation; Checked: 2026-06-19; Status: monitor	E179: The European Commission's draft AI transparency-obligations consultation page set 3 June 2026 as the feedback deadline and says Article 50 transparency duties apply from 2 August 2026 for AI-system interactions and certain AI-generated or manipulated content. Limitation: verify dated release or consultation status before relying on timing.	Open source
172	Microsoft Azure OpenAI Service	Publisher: Microsoft; Kind: Cloud documentation; Checked: 2026-06-19; Status: monitor	E182: Microsoft's Azure OpenAI Service page states that Azure OpenAI is guaranteed to be available at least 99.9% of the time. Limitation: verify dated release or consultation status before relying on timing.	Open source
173	Vertex AI Platform Service Level Agreement	Publisher: Vertex AI Platform Service Level Agreement; Kind: Cloud documentation; Checked: 2026-06-19; Status: monitor	E183: The Vertex AI Platform SLA lists monthly uptime percentage commitments including at least 99.9% for covered training, deployment and batch-prediction services. Limitation: verify dated release or consultation status before relying on timing.	Open source
174	OpenAI GPT-5.5 model rate limits	Publisher: OpenAI GPT-5.5 model rate limits; Kind: Official model documentation; Checked: 2026-06-19; Status: monitor	E184: OpenAI's GPT-5.5 model page publishes hard usage-tier limits, including Tier 1 at 500 RPM and 500K TPM and Tier 5 at 15K RPM and 40M TPM; the live usage tier, long-context cap and shared-family cap still need account verification. Limitation: verify dated release or consultation status before relying on timing.	Open source
175	OpenAI Batch API	Publisher: OpenAI Batch API; Kind: Official model documentation; Checked: 2026-06-19; Status: monitor	E185: OpenAI's Batch API gives asynchronous jobs a separate pool with higher rate-limit headroom, lower cost and a 24-hour completion target, making it a different lane from interactive financial work. Limitation: verify dated release or consultation status before relying on timing.	Open source
176	Claude API rate limits	Publisher: Claude API rate limits; Kind: Official model documentation; Checked: 2026-06-19; Status: monitor	E186: Claude API limits are measured as RPM, ITPM and OTPM by model class; Claude Opus 4.x publishes Tier 1 at 50 RPM, 500K ITPM and 80K OTPM and Tier 4 at 4K RPM, 10M ITPM and 800K OTPM, with custom and Priority Tier increases handled separately. Limitation: verify dated release or consultation status before relying on timing.	Open source
177	Groq API rate limits	Publisher: Groq; Kind: Official model documentation; Checked: 2026-06-19; Status: monitor	E187: Groq's API rate-limit page defines organization-level RPM, RPD, TPM, TPD, ITPM and OTPM controls, publishes plan and model limit tables, and says higher limits are available for select workloads and enterprise use cases; it does not expose one public production TPM ceiling suitable for the tier-wait chart. Limitation: verify dated release or consultation status before relying on timing.	Open source
178	Azure OpenAI quotas and limits	Publisher: Azure; Kind: Cloud documentation; Checked: 2026-06-19; Status: monitor	E188: Azure OpenAI quotas are defined by model, deployment type, subscription and region; the public gpt-5.5 GlobalStandard rows list 0 RPM and 0 TPM through lower tiers, then 10K RPM and 10M TPM at Tier 5 and 15K RPM and 15M TPM at Tier 6. Limitation: verify dated release or consultation status before relying on timing.	Open source
179	Azure OpenAI provisioned throughput	Publisher: Azure; Kind: Cloud documentation; Checked: 2026-06-19; Status: monitor	E189: Azure provisioned throughput separates PTU quota from actual regional capacity, supports Global, Data Zone and Regional provisioned deployment types, and adds spillover plus hourly or reservation billing options. Limitation: verify dated release or consultation status before relying on timing.	Open source
180	Vertex AI Standard PayGo throughput	Publisher: Vertex AI Standard PayGo throughput; Kind: Cloud documentation; Checked: 2026-06-19; Status: monitor	E190: Vertex AI Standard PayGo uses spend-based throughput tiers: GA Gemini Pro publishes 500K TPM at Tier 1 and 2M TPM at Tier 3, Gemini Flash and Flash-Lite publish 2M TPM at Tier 1 and 10M TPM at Tier 3, and a 30K RPM per-model per-region system limit applies outside the tier-specific TPM baseline. Limitation: verify dated release or consultation status before relying on timing.	Open source
181	Vertex AI Provisioned Throughput	Publisher: Vertex AI Provisioned Throughput; Kind: Cloud documentation; Checked: 2026-06-19; Status: monitor	E191: Vertex AI Provisioned Throughput is a dedicated-capacity route with request-type controls and monitoring, distinct from shared PayGo throughput. Limitation: verify dated release or consultation status before relying on timing.	Open source
182	Amazon Bedrock quota increase mechanism	Publisher: Amazon; Kind: Cloud documentation; Checked: 2026-06-19; Status: monitor	E192: Amazon Bedrock does not publish one comparable default public TPM row for every model route; AWS documents model-specific on-demand and cross-Region InvokeModel token-per-minute and request-per-minute quota families that must be checked in Service Quotas and increased by model. Limitation: verify dated release or consultation status before relying on timing.	Open source
183	Amazon Bedrock cross-Region inference	Publisher: Amazon; Kind: Cloud documentation; Checked: 2026-06-19; Status: monitor	E193: Amazon Bedrock cross-Region inference can route on-demand requests across regions to increase throughput during quota or peak-usage constraints, while Provisioned Throughput is the separate fixed-capacity route for higher model invocation capacity. Limitation: verify dated release or consultation status before relying on timing.	Open source
184	Amazon Bedrock Provisioned Throughput	Publisher: Amazon; Kind: Cloud documentation; Checked: 2026-06-19; Status: monitor	E194: Amazon Bedrock Provisioned Throughput is purchased for higher fixed throughput, with Model Units specifying input and output tokens per minute and commitment terms ranging from no commitment to one or six months. Limitation: verify dated release or consultation status before relying on timing.	Open source
185	OpenAI Flex processing	Publisher: OpenAI Flex processing; Kind: Official model documentation; Checked: 2026-06-19; Status: monitor	E195: OpenAI Flex processing is a lower-priority processing lane for supported models and is charged at Batch API rates, making it a price and latency trade-off rather than an interactive default. Limitation: verify dated release or consultation status before relying on timing.	Open source
186	OpenAI prompt caching	Publisher: OpenAI prompt caching; Kind: Official model documentation; Checked: 2026-06-19; Status: monitor	E196: OpenAI prompt caching automatically discounts repeated prompt prefixes above the documented threshold, exposes cached-token usage and can use prompt-cache keys to improve cache routing for related requests. Limitation: verify dated release or consultation status before relying on timing.	Open source
187	Anthropic Message Batches	Publisher: Anthropic; Kind: Official model documentation; Checked: 2026-06-19; Status: monitor	E197: Anthropic Message Batches support high-volume asynchronous Claude requests with discounted pricing and can be combined with prompt caching for repeated context. Limitation: verify dated release or consultation status before relying on timing.	Open source
188	Anthropic prompt caching	Publisher: Anthropic; Kind: Official model documentation; Checked: 2026-06-19; Status: monitor	E198: Anthropic prompt caching uses explicit cache-control breakpoints across tools, system and message content, with separate cache write and read economics and documented cache durations. Limitation: verify dated release or consultation status before relying on timing.	Open source
189	Vertex AI generative AI pricing	Publisher: Vertex AI generative AI pricing; Kind: Pricing page; Checked: 2026-06-19; Status: monitor	E199: Vertex AI generative AI pricing separates standard input and output tokens from batch-mode pricing and cached-input pricing, so Gemini route cost depends on lane as well as model. Limitation: verify dated release or consultation status before relying on timing.	Open source
190	Vertex AI context caching	Publisher: Vertex AI context caching; Kind: Cloud documentation; Checked: 2026-06-19; Status: monitor	E200: Vertex AI context caching supports explicit cached content and implicit caching, with cache duration, minimum-token and exact-prefix behaviours that affect whether repeated source packets earn the cached-input price. Limitation: verify dated release or consultation status before relying on timing.	Open source
191	Vertex AI Gemini batch prediction	Publisher: Vertex AI Gemini batch prediction; Kind: Cloud documentation; Checked: 2026-06-19; Status: monitor	E201: Vertex AI Gemini batch prediction is a non-real-time route for bulk jobs; Google documents batch mode separately from provisioned throughput and online inference. Limitation: verify dated release or consultation status before relying on timing.	Open source
192	Amazon Bedrock pricing	Publisher: Amazon; Kind: Pricing page; Checked: 2026-06-19; Status: monitor	E202: Amazon Bedrock pricing separates on-demand, batch, prompt-caching and provisioned-throughput economics across supported models, so the approved AWS lane changes the bill. Limitation: verify dated release or consultation status before relying on timing.	Open source
193	Amazon Bedrock prompt caching	Publisher: Amazon; Kind: Cloud documentation; Checked: 2026-06-19; Status: monitor	E203: Amazon Bedrock prompt caching uses cache points with TTL controls and model support limits, reducing repeated-context processing where the route is eligible. Limitation: verify dated release or consultation status before relying on timing.	Open source
194	Amazon Bedrock batch inference	Publisher: Amazon; Kind: Cloud documentation; Checked: 2026-06-19; Status: monitor	E204: Amazon Bedrock batch inference runs asynchronous jobs from S3 input to S3 output for supported models, making it a separate bulk lane from synchronous inference. Limitation: verify dated release or consultation status before relying on timing.	Open source
195	Groq prompt caching	Publisher: Groq; Kind: Official model documentation; Checked: 2026-06-19; Status: monitor	E205: Groq prompt caching automatically detects repeated exact prefixes above the documented threshold, gives discounted cached-token treatment and documents rate-limit effects for cache hits. Limitation: verify dated release or consultation status before relying on timing.	Open source
196	Groq Batch API	Publisher: Groq; Kind: Official model documentation; Checked: 2026-06-19; Status: monitor	E206: Groq Batch API is an asynchronous discounted lane that does not consume standard synchronous rate limits, but completion windows and batch limits remain part of the production route. Limitation: verify dated release or consultation status before relying on timing.	Open source
197	Groq Flex processing	Publisher: Groq; Kind: Official model documentation; Checked: 2026-06-19; Status: monitor	E207: Groq Flex processing lets requests wait for available capacity to raise effective throughput without changing regular API pricing, so it solves a different problem from batch discounting. Limitation: verify dated release or consultation status before relying on timing.	Open source
198	Harvey - Legal Agent Benchmark initial results	Publisher: Harvey; Kind: Agent benchmark; Checked: 2026-06-19; Status: monitor	E208: Harvey LAB's legal-agent trace analysis reported positive all-pass associations for source coverage before drafting, post-draft validation, revise-after-check, targeted retrieval, structured analysis and grounding against the record, and negative associations for high tool fan-out and drafting without review; the effect sizes are legal-specific, but the trajectory rubric is transferable to financial private evals. Limitation: verify dated release or consultation status before relying on timing.	Open source
199	Rogo - BigFinanceBench announcement	Publisher: Rogo; Kind: Financial benchmark; Checked: 2026-06-19; Status: monitor	E209: Rogo's May 2026 BigFinanceBench announcement describes a 928-question workflow-grounded financial benchmark written by ex-finance practitioners, scored with 15,656 rubric criteria and 36,241 weighted points; Rogo reports that no top model dominates all workflows and that a workflow/source router improves rubric score by 4.5 percentage points over the best single-model baseline. Limitation: verify dated release or consultation status before relying on timing.	Open source
200	Silicon Valley Bank - State of Enterprise Software Report 2026	Publisher: Silicon Valley Bank; Kind: Market survey; Checked: 2026-06-19; Status: monitor	E226: Silicon Valley Bank's 2026 State of Enterprise Software survey of 120+ companies reports on slide 20 that about 52% of respondents say active AI is required to hit team goals or that they cannot operate without it; one in five found productivity impact hard to measure; about 48% reported 25%+ employee-productivity improvement; revenue-growth attribution was weaker, with half saying hard to measure and one-quarter reporting less than 10% lift. \| E227: Silicon Valley Bank's pricing slide says nearly 40% of respondents currently use a primarily seat-based model, but only 26% plan to remain that way in the near term as AI labour augmentation and less predictable variable costs change software pricing; customer demand is the largest pricing-decision driver at 42%, while AI/compute economics accounts for 21%. \| E228: Silicon Valley Bank reports that the five hyperscalers' combined operating cash flow rose from $116B in 2016 to $577B in 2025 while CapEx as a share of operating cash flow rose from 28% to 66%; the same slide frames rising long-term debt and lease obligations as evidence that AI infrastructure capacity is capital intensive. Limitation: verify dated release or consultation status before relying on timing.	Open source
201	Surge AI - GDP.pdf leaderboard	Publisher: Surge AI; Kind: Professional benchmark; Checked: 2026-06-19; Status: verified	E229: Surge's GDP.pdf leaderboard, last updated 6 June 2026, reports macro accuracy on professional PDF question answering; current visible rows include Claude Fable 5 / Mythos 5 at 30%, GPT-5.5 at 25%, Claude Opus 4.8 at 23%, Claude Opus 4.7 at 21%, and Gemini 3.1 Pro at 17%.	Open source
202	Surge AI - GDP.pdf benchmark blog	Publisher: Surge AI; Kind: Professional benchmark; Checked: 2026-06-19; Status: verified	E230: Surge describes GDP.pdf as 100 real-world prompts and PDFs across finance, healthcare, legal, STEM/research, engineering, construction, manufacturing/supply chain, insurance, real estate and HR, with tasks requiring parsing, understanding and synthesizing complex PDFs.	Open source
203	Surge AI - GDP.pdf dataset	Publisher: Surge AI; Kind: Dataset; Checked: 2026-06-19; Status: verified	E231: The GDP.pdf dataset is public on Hugging Face with 100 test rows, document question-answering metadata, PDF paths, prompts, domains and rubric fields.	Open source
204	Reducto - GDP.pdf parser rerun	Publisher: Reducto; Kind: Vendor research; Checked: 2026-06-19; Status: monitor	E232: Reducto's vendor-reported GDP.pdf rerun compared PDF-only inputs with PDF plus Reducto parse output over the same 100 documents and reported aggregate macro accuracy rising from 21.7% to 30.7%, micro accuracy rising from 75.6% to 77.9%, answer latency falling from 120s to 111s, reasoning tokens falling from 8.8K to 7.7K per task, and finance-domain macro accuracy rising by 3.7 percentage points. Limitation: verify dated release or consultation status before relying on timing.	Open source

Terms used in this report

Term	Plain meaning
Route	The approved path from work to accepted output: source set, data boundary, model or vendor channel, tools, reviewer, evidence record and reapproval trigger.
API	Application programming interface: a governed software connection used by an app or workflow to call a model or service.
SDK	Software development kit: packaged code and tools developers use to connect a product to an API or service.
LLM	Large language model: a model trained to read and generate text-like tokens.
Token	A chunk of text used for billing, context windows and rate limits. It is usually smaller than a word, not a page count.
RPM	Requests per minute: how many separate API calls the route can accept each minute.
RPD	Requests per day: the daily cap on separate API calls.
TPM	Tokens per minute: how much token traffic a route can process each minute, unless the provider splits input and output separately.
TPD	Tokens per day: the daily cap on token traffic.
ITPM	Input tokens per minute: how many prompt and source-document tokens may enter the model each minute.
OTPM	Output tokens per minute: how many generated tokens may leave the model each minute.
PTU	Provisioned throughput unit: reserved Azure OpenAI capacity bought for steadier throughput.
GSU	Google scale unit: provisioned Vertex AI capacity used to reserve Gemini throughput.
MU	Model Unit: Amazon Bedrock provisioned-throughput capacity reserved for a specific model route.
PayGo	Pay as you go: shared cloud capacity billed by usage rather than pre-reserved throughput.
RAG	Retrieval-augmented generation: search approved sources first, then pass selected evidence into the model.
OCR	Optical character recognition: converting scanned or image text into machine-readable text.
SLA	Service-level agreement: a contractual availability or support commitment, not observed uptime history.
JSON	JavaScript Object Notation: a structured text format commonly used for machine-readable API output.
429	HTTP 429 means too many requests: the route hit a rate limit or shared-capacity throttle.

Definitions are reader aids, not separate evidence claims.