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

The decision is control evidence. Do not buy one financial model for three different jobs. Banking, capital markets and insurance share a regulatory roof, but their model failures hurt people and markets in different ways. A banking route can affect credit, payments, fraud handling, complaints and customer treatment ⁶⁹. A capital-markets route can affect investment research, client communications, surveillance and trading-adjacent analysis ³¹⁴. An insurance route can affect underwriting, pricing, claims, fraud and actuarial judgement ⁷⁴³.

Use the broad financial-model layer for the full portfolio view. Use this control map to approve the regulated-finance use cases that sit below it. The object is still not a model name. It is the output path: source authority, data class, model lane, hosting path, reviewer, logs, action permissions and reapproval trigger ¹⁹.

The strongest public evidence says buyers should be humble. FinanceBench showed a GPT-4-Turbo retrieval route incorrectly answered or refused 81% of sampled financial questions, despite answers being grounded in public-company material ²³. That is a filing-RAG warning, not a permanent model ranking.

Keep banking, capital markets and insurance in one control map until demand justifies narrower tracks. The useful first split is a three-lane approval register. A retail-credit track, a market-surveillance track or a claims-pricing track is worth creating only when the source pack, owner and evaluation burden stop overlapping.

The finished output to measure is not a token. It is an approved credit explanation, resolved alert, cited research note, reviewed workbook, signed-off claims decision or archived customer message.

Related report

AI models for financial services

Covers the full model portfolio, open-weight route, benchmark map, source-layer economics and cross-finance controls.

Related report

File processing providers for financial AI

Use the file-processing report 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
One control map or three?	One dedicated control map for the three regulated lanes now.	Split later only when demand, source coverage and owner accountability diverge.
Which model wins?	The winning model changes by source, customer impact and action permission.	Approve model lanes by control evidence, not by leaderboard.
Where is the main failure?	Wrong source, stale data, hidden table, bad formula or unsupported action.	Test retrieval and evidence before prose.
What should procurement price?	Cost per accepted financial output.	Count review minutes, false refusals, reruns, audit export and reapproval burden.

▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒What changed

The control problem became explicit. The main change is not that financial firms discovered AI. It is that the control perimeter caught up with the difference between assistant, model and action path. OCC Bulletin 2026-13 revised banking model-risk guidance while explicitly leaving generative and agentic AI outside scope, pending further work ⁴⁵⁴⁵. That creates a practical gap: banks still need control evidence for GenAI even when the traditional model-risk text does not settle every question.

Capital-markets and broker-dealer signals moved in parallel. IOSCO maps AI across advice, trading, research, surveillance and AML/CFT, while FINRA's 2026 material points to common GenAI uses such as summarization and information extraction alongside supervision, data and guardrail concerns ⁴⁹⁵⁰.

For insurance, the control language is more customer-outcome specific. NAIC materials put governance, testing, documentation, proxy, fairness and complaint-handling questions close to underwriting and pricing routes ⁷⁸. NYDFS and EIOPA materials add governance and insurance-risk expectations for pricing and underwriting routes ⁵³⁵⁴.

The result is a three-speed market. Frontier labs and cloud providers are shipping faster workflow wrappers. Regulators are clarifying sector expectations. Buyers are left to prove that each route is still fit for its duty.

From model shortlist to control file

Old question	New question	Evidence needed
Which LLM is best?	Which route clears this duty?	Source, reviewer, model version, action limit and reapproval trigger.
Can it summarize?	Can it preserve the evidence that matters?	Source passage, table, formula, timestamp, permission and refusal log.
Can vendors claim AI?	Are claims substantiated by actual operations?	Marketing, policy, workflow, model-use and audit evidence.
Is AI a separate policy?	Which existing control owns the route?	Model risk, third-party risk, conduct, resilience and records mapping.

▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒Decision frame

Start with consequence, then choose the model lane. Financial AI splits cleanly when the buyer names the consequence of a bad output. If a model drafts an internal research note, the approval question is source quality and analyst review. If it influences credit, trading, insurance pricing or claims handling, the approval question becomes conduct, fairness, audit trail, human accountability and sometimes high-risk AI classification ³¹¹.

The Bank of England and FCA survey taxonomy is a useful cut: banks, insurance, investment and capital markets, FMIs, payments and adjacent firms have different operating risks ². The control map keeps three main lanes because they share model, source and governance questions while retaining distinct acceptance tests.

The model control should be written as a row, not narrated as a policy. Required row fields: use case, source corpus, data class, model lane, deployment channel, action permission, reviewer, log fields, fallback and reapproval trigger.

When those rows start naming different source owners, buyers and private eval packs, split the control map. Until then, a single regulated-finance control map prevents a false precision problem.

Ask what the output can change.

Can the output affect a customer, insured, investor, trade, payment, claim or credit decision?
Yes
Regulated route.
Next: Add conduct, fairness, model-risk, evidence and human-approval gates.
No
Continue.
Next: Check data class, source authority and action permissions.
Does the route use private customer, account, market, policy, claims or portfolio data?
Yes
Controlled-data path.
Next: Approve retention, region, training use, subprocessors, access logs and deletion evidence.
No
Continue.
Next: Check whether the route is source-authoritative.
Can the system write, send, update, approve, trade, decline or settle?
Yes
Action route.
Next: Require dry-run mode, tool limits, four-eye approval, audit export and rollback.
No
Assistant route.
Next: Run private evals and reviewer logging before scale.

▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒Model economics

Price the finished regulated task. Token price is a weak denominator for regulated financial work. The expensive part of a financial AI route is often not the model call. It is finding the right source, preserving the table, checking the calculation, handling permission denial, recording reviewer decisions and proving the output after a complaint, exam or dispute ²³²³.

Morgan Stanley's case study is useful because it describes an eval-led retrieval and quality-assurance process, not merely an assistant launch ¹⁴¹⁴. That is the economic pattern: spend money where it improves accepted-output rate and reduces reviewer rework.

Artificial Analysis-style comparisons are useful because they put price beside intelligence, output speed, latency and context ⁵⁶. Banking, capital-markets and insurance buyers need one more column: accepted output. The unit is not a million tokens. It is an approved answer, reviewed memo, resolved alert, compliant message or claims decision.

Silicon Valley Bank's State of Enterprise Software survey adds a market warning: productivity and revenue are not the same claim. About 48% of respondents reported 25% or greater employee-productivity improvement from AI, but half said revenue-growth attribution was hard to measure and one-quarter put revenue lift below 10% ⁹⁰. The same Silicon Valley Bank report also shows why seat price is a weak budgeting proxy: nearly 40% of respondents are primarily seat-priced today, but only 26% expect to stay there as AI labour augmentation and compute economics change pricing ⁹⁰.

Compare routes by total accepted-output cost: model usage, source licence, document processing, tool calls, latency, reruns, reviewer minutes, false negatives, false positives, audit exports and reapproval.

Token price misses regulated output

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 ⁵⁶; intelligence bands after a16z ⁵⁷.

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 credit file queues, KYC batches, surveillance runs, claims surges 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 regulated 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. ⁶⁵	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. ⁷⁵	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. ⁷⁷⁷⁸	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. ⁷⁹⁸¹	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. ⁸²⁸⁴	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. ⁸⁶⁸⁷	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. ⁶⁴⁶⁶⁶⁷⁶⁸⁷⁰⁷²	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. ⁶⁹⁷¹⁷⁴	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 banking, capital-markets and insurance route register. Cache savings are real only when the route keeps a stable prefix and still produces regulated 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. ⁷⁶	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. ⁷⁸	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. ⁷⁹⁸⁰	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. ⁸²⁸³⁸⁴	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. ⁸⁵	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 regulated 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 regulated output, not merely cheaper per million tokens.

Cost per accepted output

Cost item	Token view	Accepted-output view	Measurement
Retrieval	Tokens used.	Correct source support.	Correct source recall and unsupported-answer rate.
Calculation	Answer fluency.	Verified calculation.	Intermediate values, formula trace and reviewer acceptance.
Fraud and AML	Alert summary quality.	Resolved investigation.	False-negative rate, false-positive rate and investigator disposition.
Customer output	Message generation cost.	Approved communication.	Approved communication with source, reason, complaint flag and archive record.
Model change	New model price.	Reapproved route.	Holdout rerun cost and route reapproval burden.

▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒Throughput and rate limits

Rate limits decide whether the model can survive regulated workflow volume. Rate limits are production capacity, not a nuisance error. A model can be the best reasoning route and still fail a credit file queue, KYC batch, surveillance run, claims surge, market-research packet or regulatory response archive 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 banking, capital-markets and insurance 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 ⁶⁴. 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. ⁶⁴	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. ⁶⁶	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. ⁶⁷	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. ⁶⁸	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. ⁷⁰	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. ⁷²	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.

▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒Model lanes

Different duties need different model families. The first model decision is lane, not brand. Frontier LLMs are strongest where financial work needs reasoning, synthesis, drafting and exception handling. Finance-specific and source-provider routes matter where source rights, current market data, credit research or domain vocabulary carry the decision. Specialist risk models still belong in fraud, payments, AML, surveillance and claims workflows ²¹²².

Open-weight and self-hosted routes are attractive for data control and high-volume tasks, but they move validation, security, patching, infrastructure and monitoring onto the buyer. That can be a good trade for low-risk classification. It is a poor shortcut for regulated decisions.

The route may combine lanes: OCR or document AI, embedding and reranking, a source-provider retrieval layer, a frontier generator and a specialist risk model. The buyer should approve the combination.

Lane comparison

Lane	Best for	Approval question
Frontier general model	Reasoning, drafting, synthesis, exception analysis and tool planning.	Can it produce source-supported work under approved data terms?
Finance-specific model	Finance vocabulary, filings, sentiment, market language and domain tasks.	Is domain training current, licensed and validated for the route?
Source-provider AI	Licensed research, market data, ratings, filings, estimates and workflow search.	Do source rights, entitlements and archive records survive output use?
Specialist risk model	Fraud, AML, payment authorization, surveillance and claims-fraud signals.	What are the false negatives, false positives and escalation outcomes?
Private or open-weight model	High-volume internal classification, summarization and sensitive low-risk work.	Can the buyer operate, secure, validate and refresh it?

▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒Use-case route selection

Approve by work type, source and consequence. The control table should read like an approval register, not a vendor shortlist. Frontier models belong where synthesis, explanation, drafting and multi-step reasoning matter. Specialist models and source-provider routes belong where data rights, market freshness, fraud signals, insurance workflows or payment latency matter more.

Use lower-cost or private models where classification, extraction, summarization and triage clear the workflow with measured quality. Use the stronger route where a reviewer needs the full evidence chain or where the cost of a wrong answer is high.

The same buyer may approve five routes at once: public-source research assistant, confidential customer-data assistant, fraud triage model, document extraction model and agentic workflow. Keep them separate in contracts and evals.

Control map

Route	Best model lane	Approval gate	Do not use for
Bank policy and product Q&A	Frontier or cloud model with approved KB retrieval.	Source freshness, no-answer handling and reviewer log.	Credit decisioning or adverse-action explanation.
Credit memo support	RAG plus table extraction, deterministic calculations and analyst review.	Source, formula, reason code, fairness and model-risk evidence.	Unreviewed credit approval or decline.
AML and fraud triage	Specialist risk model plus language model for investigator narrative.	False-negative replay, typology drift and case audit trail.	Autonomous SAR, account closure or payment blocking.
Investment research	Frontier model over licensed research, filings, transcripts and market data.	Entitlements, source timestamps and analyst sign-off.	Order generation, personalized advice or client send without supervision.
Market surveillance	Specialist surveillance model plus language layer for explanation.	Adversarial replay, false-negative analysis and escalation evidence.	Unsourced market-manipulation findings.
Insurance underwriting support	Line-specific models plus document extraction and underwriter review.	State rules, fairness tests, factor lineage and appeal path.	Unreviewed pricing, denial or coverage decision.
Claims support	Document, vision and language models inside claims workflow.	Claim file, image evidence, adjuster decision and fraud review.	Autonomous settlement or denial.

▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒Evidence and benchmarks

Public benchmarks tell buyers what to test privately. Financial benchmarks are useful when they expose failure modes. They mislead when they pretend to certify a production control path. FinanceBench, BigFinanceBench, FinBen, TAT-QA, ConvFinQA, FinQA and Finance Agent each see part of the problem: source retrieval, workflow rubrics, tables, numerical reasoning, task breadth, multi-turn context or tool use ²³²⁴⁸⁹. None sees a bank's private customer permissions, an asset manager's entitlements, an insurer's state-by-state rules or a firm's reviewer policy.

Finance-specific models such as BloombergGPT and FinGPT make the domain-data point. Financial language is temporal, source-bound and full of abbreviations, conventions and tables. Domain data helps. It is not a substitute for current licensed sources and private acceptance tests ²¹²².

Use public benchmarks to seed the private eval. Include correct refusal, false refusal, source recall, formula trace, table-cell preservation, permission denial, stale-source warning, reviewer rejection and audit export.

Benchmarks do not approve regulated use

Benchmark	Signal	Limit
FinanceBench	Open-book financial QA: 10,231 public-company questions with answer and evidence strings. Best used to test retrieval, evidence strings, table capture, period handling and refusal behaviour.	Minimum standard for public filing QA, not an approval test for private banking, markets or insurance data.
BigFinanceBench	Workflow-grounded finance agents: 928 questions and 15,656 rubric criteria across ten frontier systems. Best used to add workflow-specific scoring and routing probes to private regulated-finance evals.	Vendor-led benchmark announcement; it is not a regulated banking, capital-markets or insurance approval test.
FinBen	Holistic finance benchmark: 36 datasets across 24 tasks and seven aspects. Best used as a taxonomy for routing private evals by task family.	Task coverage is broad, but regulated decisions still need permissioned-data, fairness and audit tests.
TAT-QA / FinQA / ConvFinQA	Table, formula and multi-turn reasoning: Financial questions over reports, tables and chained calculations. Best used to make calculation trace and source-cell evidence non-negotiable.	Older task sets still catch current production failures when tables, periods and intermediate calculations are lost.
Finance Agent	Tool-using finance workflow: Multi-step finance work using search, calculations and final answer submission. Best used before approving agent routes with write access or client-facing output.	Workflow-like public benchmark, but still harness-specific and not a substitute for institution data.

▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒Source layer

The model can only know the approved source. In regulated finance, source access is a model capability. A general model over stale web text is not competing with a route that has EDGAR APIs, licensed market data, broker research, portfolio analytics, credit research and entitlements. The latter may be less glamorous and more useful ²⁹³¹.

Capital-markets work makes this obvious. A useful model needs the right filing, current transcript, licensed estimate, market-data timestamp, research entitlement and internal portfolio context without leaking or inventing a source.

Do not compare model providers, data providers and workflow providers as if they sell the same thing. The route may need all three: a model, a source and an approved work surface.

Source rights are part of model capability

Layer	Controls	Advantaged routes	Buyer read
Public filings	Official APIs, accession numbers, amendments and filing dates.	EDGAR-backed RAG and filing-aware systems.	Good for public-company research and benchmark packs.
Licensed market data	Entitlements, timestamps, vendor contracts and redistribution limits.	LSEG, Bloomberg, FactSet and source-provider routes.	Required for market-sensitive work.
Credit and research sources	Subscription rights, analyst provenance, rating methodology and archive retention.	Moody's, S&P and specialist research routes.	Useful only when source rights flow through the model route.
Private records	Customer consent, account permission, policy file, claim file, transaction log and reviewer scope.	Institution-owned RAG, cloud control planes and workflow SaaS.	Highest diligence burden.

▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒Banking control path

Banking AI is a permissions and consequence problem. The bank control path should be strictest where the model changes a customer's financial position or the firm's financial-crime posture. Employee research, policy interpretation and product Q&A can move first when sources are approved and reviewers remain accountable. Credit, fraud, AML and payments need a different route: labelled cases, false negatives, reason codes, customer-treatment evidence, escalation paths and model-risk records ⁶⁹.

Payment and fraud models have their own production physics. A brilliant explanation is useless if it arrives too late, misses coordinated fraud or cannot reconstruct the decision after a dispute ³⁷⁴⁰.

Treat language models as a narrative and investigation layer over specialist risk systems until the institution can prove the route's false-negative, fairness and audit performance.

Banking acceptance tests

Work	Acceptance test	Blocked action
Credit explanation draft	Reason-code support, source evidence, fairness slice and credit analyst approval.	Autonomous decline or adverse-action notice.
AML investigation summary	Case replay, false-negative review, typology drift and investigator disposition.	Autonomous suspicious-activity filing.
Payment fraud triage	Latency, fraud loss, false positive, false negative and customer-dispute reconstruction.	Unreviewed blocking outside approved thresholds.
Customer service response	Policy source, customer record permission, complaint flag and escalation path.	Advice, promise, waiver or irreversible account change.

▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒Capital-markets control path

Keep research far from execution until controls prove otherwise. Capital-markets AI should be generous with research support and miserly with action permission. The safe early route is source-backed research over filings, transcripts, broker research, ratings, market data and internal notes. The risky route is any path that changes client advice, order logic, surveillance disposition or market-facing communication ⁴³¹.

A markets model also needs a source-time test. A filing date, transcript timestamp, estimate vintage, corporate action or market-data entitlement can make yesterday's correct answer wrong today.

Separate research generation from execution systems. Use the model to find evidence, draft analysis and prepare surveillance narratives. Keep order generation, trade approval and client send behind existing controls.

Capital-markets control tests

Route	Source test	Approval rule
Public-company research	EDGAR accession, filing date, table and amended-filing check.	Analyst sign-off before client use.
Licensed research synthesis	Entitlement, source label, timestamp and redistribution check.	Block unsupported quotes and stale estimates.
Portfolio analytics	Position permission, benchmark definition, holdings date and model assumption trace.	Portfolio manager approval before recommendation.
Surveillance narrative	Event sequence, communication record, alert threshold and comparator group.	Investigator disposition before escalation.

▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒Insurance control path

Insurance models are local, line-specific and evidence-heavy. Insurance AI fails when a general model control ignores line, state, evidence type and customer outcome. Underwriting support, pricing support, claims triage, claims fraud and service response have different acceptance tests. A system that summarizes a policy document is not approved to price a life product or deny a claim ⁷¹¹.

State adoption of the NAIC model bulletin makes the control local. A multi-state insurer needs a control map that links model use, line of business, state, consumer impact, third-party data, audit evidence and appeal path ⁸⁷.

Do not split insurance into narrower tracks until public evidence and buyer demand justify it. The same first-order controls recur across lines: factor lineage, unfair-discrimination testing, claim-file evidence, examiner review and vendor oversight.

Insurance control tests

Route	Control	Private eval
Underwriting support	Approved factors, source lineage, underwriter override and state map.	Segment tests for unfair discrimination and reason-code quality.
Claims triage	Claim-file source links, image evidence, policy terms and adjuster approval.	Replay disputed and edge-case claims.
Claims fraud	Fraud signal provenance, examiner review and appeal path.	False-negative and false-positive review by line.
Service response	Policy source, jurisdiction, customer state and escalation trigger.	Complaint-prone scenarios and refusal quality.

▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒Risks and controls

Control the model path, not the slogan. The control stack should look familiar to financial institutions: model risk, third-party risk, resilience, conduct, records and operational ownership. Do not create a special AI exception lane. Attach AI route evidence to the control frameworks that already decide whether a financial institution can use a model, vendor, cloud service, source provider, customer-data flow or automated tool ⁵⁹.

The new work is connective tissue. A model-risk owner needs the benchmark and validation evidence. A third-party-risk owner needs provider terms and exit paths. Compliance needs customer and market-conduct evidence. The business owner needs accepted-output quality.

The route register should be queryable by regulator, system owner, data class, provider, model version, source corpus, approval state, reviewer and renewal trigger.

Events that reopen approval

Scenario	Trigger	Route risk	Control
Provider model update	Model version, context, tool API, pricing or safety behaviour changes.	Approved output quality no longer holds.	Holdout rerun, reviewer sample and route reapproval.
Source provider change	Entitlement, archive, data field, identifier mapping or redistribution term changes.	Model cites unavailable or stale evidence.	Source diff, entitlement test and stale-source warning.
Cloud or model-provider outage	Status-page incident or material degradation.	Critical workflow loses availability or fails over unsafely.	DORA-style resilience mapping, fallback route and incident record.
Adversarial fraud pattern	New FINRA, payment-network or internal fraud signal.	False negatives rise before the model is retuned.	Adversarial replay, threshold review and investigator sign-off.
Cross-border classification change	EU AI Act, NAIC or local regulator update.	Previously acceptable route becomes high-risk or locally constrained.	Jurisdiction map and route owner attestation.

▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒Deployment model

Choose deployment by data and action, not preference. Deployment approval follows the data boundary, action permission and recovery burden. Direct APIs, cloud platforms, source-provider routes, vertical SaaS and self-hosted open weights are different approval objects. Match the route to data rights, regulated activity, action permission and reviewer evidence before comparing model names.

The best route can differ inside the same institution: cloud frontier models for analyst research, source-provider AI for licensed market data, specialist fraud models for payment risk and private models for low-risk classification where data residency or cost dominates.

Attach deployment evidence to each approved route: hosting path, data-retention terms, region, source retrieval controls, gateway policy, incident handling, 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.

▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒Cross-border limits

Global finance needs local control files. A global financial model control breaks when it treats jurisdiction as a language setting. Cross-border approval must split law, regulated activity, data movement, cloud region, source licence, local market convention and accountable reviewer. EU credit and insurance use cases can land in high-risk AI territory, DORA adds ICT and third-party resilience controls, and insurance obligations vary by state and line ¹¹¹².

The practical danger is silent overreach. A control approved for public UK research is not approved for EU life-insurance pricing, US credit adverse-action explanations, APAC customer records or market-data redistribution.

Use one global model policy only as the outer envelope. The approval artefact should be local enough to name regulator, data class, source licence, reviewer and escalation path.

Cross-border control tests

Test	Question	Blocked if
Regulated use	Does the route affect credit, life or health insurance, advice, trading or claims?	The route lacks local classification and control evidence.
Data movement	Where do prompts, logs, embeddings and source artefacts travel?	Region, retention, training use or subprocessor chain is unclear.
Source licence	Can licensed research or market data be used in generated output?	Entitlement and redistribution rights do not pass through the route.
Local reviewer	Who signs off the result in the relevant jurisdiction?	Accountability remains with a generic global team.

▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒Private eval design

Measure the finished job. The private eval should score what the institution will accept, not what a leaderboard can see. A good route returns the right source, refuses correctly when evidence is missing, preserves tables, performs calculations, cites current data, respects permissions, exposes uncertainty and leaves an audit trail. A poor route writes a plausible paragraph and hides the error.

Cost per accepted output should include model spend, source-provider cost, tool calls, retries, latency, reviewer minutes, false refusals, false positives, false negatives and audit-export work ²³²⁸.

Keep a holdout set by route. Rerun it after model, prompt, parser, retriever, source, data-right, tool-permission or reviewer-policy changes.

Accepted-output scorecard

Metric	Why it matters	Failure to log
Correct source recall	The model cannot reason over evidence it did not retrieve.	Wrong filing, stale policy, missing claim file or unauthorised record.
Table and formula trace	Finance errors often sit in rows, periods, units and formulas.	Lost subtotal, wrong period, hidden sheet or unexplained calculation.
Correct refusal	A safe refusal is good; a false refusal is broken work.	Evidence exists but route refuses, or route answers without evidence.
Reviewer minutes	Model cost is small if review dominates.	Reviewer must re-open sources, recalculate or rewrite unsupported claims.
Adversarial false negative	Fraud, AML, market deception and claims fraud punish missed cases.	Known bad case passes as clean.
Audit export	Disputes and exams need reconstruction.	No source span, model version, prompt, tool call, reviewer or final decision.

Agent trajectory eval

Behaviour	Harvey LAB legal-agent signal	Private eval translation
Source coverage before drafting	90%+ document coverage appeared in 52% of Harvey LAB trajectories and was associated with +0.4 all-pass points. ⁸⁸	Check whether the route opens the required source set before first draft.
Post-draft validation	Validation or review after drafting appeared in 88% and was associated with +0.8 points. ⁸⁸	Require a review step against instruction, source and policy before acceptance.
Verifying and revising	Revise-after-check appeared in 71% and was associated with +1.5 points. ⁸⁸	Score whether validation changes the deliverable, not only whether the agent self-approves.
Targeted retrieval	Targeted retrieval appeared in 60% and was associated with +0.3 points. ⁸⁸	Test search for missing facts instead of relying only on initially opened context.
Structured analysis	Code or shell analysis appeared in 36% and was associated with +0.3 points. ⁸⁸	Use tools, schemas or calculations where parsing, arithmetic or reconciliation matters.
Grounding against the record	Reopening source evidence after drafting appeared in 19% and was associated with +0.3 points. ⁸⁸	Check whether the agent returns to evidence after drafting to correct unsupported claims.
High tool parallelism	Five-plus tool calls in one turn appeared in 35% and was associated with -0.5 points. ⁸⁸	Measure whether broad fan-out creates noisy context, extra cost or poor synthesis.
Drafting without review	Drafting without later validation or revision appeared in 11% and was associated with -1.2 points. ⁸⁸	Block first-pass outputs for consequential workflows unless a separate reviewer gate runs.

▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒Opportunities

Use models where the duty stays reviewable. The best early opportunities make expert work faster without hiding the expert. Banking can move in policy Q&A, call-centre assist, complaint triage, credit memo support, fraud investigation summaries and compliance evidence packs before it moves to autonomous customer decisions. Capital-markets teams can move in research synthesis, source-backed drafting, surveillance narratives and portfolio risk explanation before execution. Insurers can move in claims-file summarization, policy comparison, underwriter assist and fraud investigation before pricing or denial ¹³¹⁵.

The common feature is reviewability. The model should reduce time to a better human decision, not become an unreviewed decision-maker with a pleasant writing style.

A useful pilot produces an issue register, source links, reviewer outcomes and rejected-output examples. A demo that produces only polished text is unfinished.

Best first routes

Lane	Good first route	Why
Banking	Credit memo support and fraud investigation narrative.	Analysts and investigators remain accountable while the route preserves evidence.
Capital markets	Filing, transcript and research synthesis.	Source-backed drafting helps analysts without touching execution.
Insurance	Claims-file summarization and underwriter assist.	The model organizes evidence while adjusters and underwriters decide.
Cross-lane	Compliance evidence packs and policy Q&A.	The output is reviewable and can be grounded in approved documents.

▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒Operational signals

Signals reopen approval; they do not prove fitness. Demand signals are useful only when they change the next diligence question. 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 banking, capital-markets or insurance 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.
Sector regulator updates	FSB, EU AI Act, DORA, OCC, FINRA, IOSCO, NAIC, NYDFS or EIOPA language.	Reopen classification, supervision, evidence, model-risk, conduct and resilience rows.
Concentration and outsourcing signals	Common-provider exposure, critical third-party dependency, shared cloud routes and supply-chain fragility.	Run fallback and dependency-map review before the route touches customer, market or claims outcomes.
Distinct buyer demand	One lane gains a separate owner, budget, source pack, public tender or private eval need.	Consider splitting banking, capital markets or insurance into a narrower report and approval pack.

▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒Expert watchlist

Follow people who change the test, not people who summarize the market. The useful watchlist combines financial-model builders, benchmark authors, regulators and scholars of model risk, fairness and market structure. Useful signals change private evals: new finance benchmarks, validation methods, model-risk guidance, insurance fairness work, financial-crime typologies and market-structure research.

Avoid generic AI commentary. The buyer needs people whose public work changes a route test, a source assumption or a control requirement.

Replace names when the signal becomes product marketing or general market commentary without reusable tests.

People to monitor

Expert	Lane	Signal	Links
Bertie Vidgen Patronus AI; FinanceBench author	Financial benchmarkUK/USAFinancial QA, evidence strings and model-evaluation failure modes.	Useful for filing-RAG and source-supported answer tests. FinanceBench updates, retrieval failure analysis and evaluation-method changes.	Open Bertie Vidgen primary work
Hongyang Yang AI4Finance Foundation	Open financial modelsUSFinGPT, open financial LLMs, data-centric finance AI and financial agent tooling.	Useful for tracking open-source financial-model and financial-data methods. FinGPT releases, datasets, benchmarks and model cards.	Open Hongyang Yang primary work
Andrew W. Lo MIT Sloan	Financial economicsUSFinancial markets, adaptive markets, risk and financial innovation.	Useful for separating model output quality from market feedback and behaviour. Research on AI, finance, market behaviour and systemic feedback.	Open Andrew W. Lo primary work
Marcos Lopez de Prado Cornell; MLFinLab	Financial machine learningUSBacktesting, leakage, overfitting and validation in financial ML.	Useful for catching trading and markets evals that look good because they leak. Validation methods for financial ML and backtesting discipline.	Open Marcos Lopez de Prado primary work
Hilary J. Allen American University Washington College of Law	Financial stabilityUSFintech, systemic risk and financial regulatory limits.	Useful for counterweighting vendor claims about systemic safety. Legal scholarship on AI, finance and systemic fragility.	Open Hilary J. Allen primary work
Daniel Schwarcz University of Minnesota Law School	Insurance lawUSInsurance discrimination, regulation and algorithmic underwriting.	Useful for insurance pricing, underwriting and claims fairness tests. Insurance AI scholarship and regulator-facing work.	Open Daniel Schwarcz primary work
Sandra Wachter Oxford Internet Institute	AI governanceUKExplainability, fairness and algorithmic accountability.	Useful for credit and insurance explanation evidence. Research on meaningful explanations and fairness testing.	Open Sandra Wachter primary work
Maureen O'Hara Cornell Johnson	Market microstructureUSTrading, liquidity, market structure and automation.	Useful for model routes near markets, surveillance and execution. Research on automation, market integrity and liquidity effects.	Open Maureen O'Hara primary work

▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒Vendor directory

Use the directory to route diligence by layer. A vendor directory is useful only if it preserves the layer distinction. The directory should preserve four layers: frontier and cloud routes, financial source and workflow routes, fraud and payments routes, and insurance workflow routes. Each layer carries a different proof burden.

The diligence question changes by layer. Model providers need data-use, retention, region and model-change questions. Source providers need entitlement and currentness questions. Workflow providers need review UX, audit exports and action-gate 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.

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

▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒Known unknowns

Missing public evidence should become private tests. The public evidence is strongest for filings and weakest where permissioned data, money movement and customer outcomes meet. Regulated-finance buyers should assume every material route needs private acceptance tests. Public benchmarks and vendor case studies help choose probes, but they rarely show a firm's own permissions, customer files, trade records, policy forms, claim images or investigator decisions.

Turn each public-evidence gap into a procurement requirement. If the vendor cannot support the test, the route is not ready for the work.

The split question is also an evidence gap. Create narrower tracks only when the public source pack and private buyer questions are specific enough to support a separate control map.

Gaps to turn into tests

Gap	Risk	Private test
Credit adverse-action outputs	Unsupported or unfair customer explanation.	Reason-code, source, fairness and appeal-path replay.
Insurance line-specific fairness	A route passes in one line and fails in another.	Line, state, factor and protected-class proxy tests.
Live market-data entitlements	Model uses stale or unauthorised market data.	Entitlement-bound retrieval and timestamp checks.
AML and fraud false negatives	Bad cases pass as clean.	Labelled replay with investigator disposition and severity weighting.
Model-provider concentration	Common outage or defect hits many routes.	Dependency map, fallback drill and route criticality review.
Audit export quality	Firm cannot reconstruct output after exam, dispute or incident.	Replay prompt, source, tool call, model version, reviewer and decision.
Demand signal for narrower tracks	Premature split creates thin evidence and duplicated controls.	Track owner, workflow volume, source specificity and buyer questions by route.

▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒Recommended next steps

Build the approval artefacts before scaling usage. The first deliverable is a three-lane route register. Create one register for banking, one for capital markets and one for insurance. Each row should name the work, source, data class, model lane, host, action permission, reviewer, log fields, fallback and reapproval trigger.

Then run private evals on the rows with the highest consequence. Do not expand from research assistance into customer, market or claims decisions until the route can show accepted-output evidence.

Use the split rule as governance, not catalogue housekeeping: a new subvertical track earns its place when the route register shows enough distinct demand and evidence.

Action plan

Step	Owner	Output
Create route registers	Model-risk owner with business route owners	Banking, capital-markets and insurance registers with source, data, model, reviewer and trigger fields.
Choose first private eval pack	Route owner and validation lead	Worst-document, worst-customer, worst-market and worst-claim cases with rejection rules.
Map provider controls	Third-party risk and security	Data retention, region, subprocessor, resilience, exit and incident evidence by route.
Run accepted-output economics	Finance and operations	Cost per approved answer, resolved alert, reviewed memo, reconciled cell or claim decision.
Set reapproval triggers	Governance forum	Model, parser, retriever, source, tool, regulator, provider and benchmark triggers.
Decide split timing	Portfolio owner	Demand-signal threshold for narrower banking, market or insurance tracks.

▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒Source ledger

Primary sources, benchmarks, regulatory anchors and source-layer references.

Primary source list

Ref	Source	Verification	Evidence	Open source
1	FSB - financial stability implications of artificial intelligence	Publisher: FSB; Kind: Regulator; Checked: 2026-06-16; Status: verified	E1: The FSB says AI adoption in finance may amplify vulnerabilities including third-party dependencies, market correlations, cyber risk and model risk, while financial authorities face monitoring challenges because data on AI usage is limited. \| E2: The FSB describes current financial AI use as concentrated in internal operations and regulatory compliance, with generative AI opening document summarization, information retrieval and code-generation use cases. \| E51: The FSB calls for authorities to enhance monitoring of AI developments and supervisory capability, so regulator signals should reopen approval.	Open source
2	Bank of England and FCA - AI in UK financial services 2024	Publisher: Bank of England and FCA; Kind: Regulator survey; Checked: 2026-06-16; Status: verified	E3: The Bank of England and FCA survey groups respondents into UK banks, international banks, insurance, investment and capital markets, non-bank lending, FMIs, payments and other firms, which maps directly onto separate model routes. \| E4: The 2024 UK survey received 118 responses and was designed to understand AI capabilities, development, deployment and use across regulated financial firms, including generative AI. \| E48: The Bank of England and FCA sector taxonomy supports a three-lane control map for banking, investment and capital markets, and insurance before splitting into narrower subvertical tracks.	Open source
3	FCA - AI and the FCA approach	Publisher: FCA; Kind: Regulator; Checked: 2026-06-16; Status: verified	E5: The FCA frames AI adoption through existing financial-services rules and safe, responsible adoption rather than a single new AI licence route.	Open source
4	SEC - examination priorities	Publisher: SEC; Kind: Regulator; Checked: 2026-06-16; Status: verified	E6: The SEC examination-priorities page lists 2026 priorities and prior years; the 2025 priorities explicitly named artificial intelligence among emerging risk areas for registered entities. \| E52: SEC examination priorities are a recurring signal for investment advisers, broker-dealers, clearing agencies and other capital-market entities using automated investment tools, trading algorithms or AI-related representations.	Open source
5	FINRA Regulatory Notice 24-09	Publisher: FINRA Regulatory Notice 24-09; Kind: Regulator; Checked: 2026-06-16; Status: verified	E7: FINRA Regulatory Notice 24-09 reminds member firms that existing obligations still apply when using generative AI and large language models.	Open source
6	FINRA - GenAI adversarial-use threat intelligence series	Publisher: FINRA; Kind: Regulator; Checked: 2026-06-16; Status: verified	E8: FINRA's GenAI threat-intelligence series covers adversarial uses including cybersecurity risks, investment club scams, new-account fraud, account takeover, business email compromise, market deception and ransomware.	Open source
7	NAIC model bulletin on use of AI systems by insurers	Publisher: NAIC model bulletin on use of AI systems by insurers; Kind: Regulator; Checked: 2026-06-16; Status: verified	E9: The NAIC model bulletin expects insurers to govern AI systems across the insurance life cycle through governance, risk management, internal controls, audit and third-party oversight. \| E50: The NAIC bulletin applies across insurer AI systems rather than one product line, which supports keeping insurance in the same regulated-finance control map until line-specific evidence warrants a split.	Open source
8	NAIC - implementation map for AI model bulletin	Publisher: NAIC; Kind: Regulator; Checked: 2026-06-16; Status: verified	E10: NAIC's implementation map shows state-by-state adoption of the model bulletin, making insurance AI approval more fragmented than a global model policy.	Open source
9	Federal Reserve SR 11-7 model risk management	Publisher: Federal Reserve SR 11-7 model risk management; Kind: Regulator; Checked: 2026-06-16; Status: verified	E11: SR 11-7 treats model risk management as a lifecycle discipline involving development, implementation, validation, governance, controls and ongoing monitoring.	Open source
10	Federal Reserve SR 23-4 third-party risk management	Publisher: Federal Reserve SR 23-4 third-party risk management; Kind: Regulator; Checked: 2026-06-16; Status: verified	E12: SR 23-4 says third-party risk management should be tailored to the size, complexity, risk profile and nature of a banking organisation's third-party relationships.	Open source
11	EU AI Act Regulation 2024/1689	Publisher: EU AI Act Regulation 2024/1689; Kind: Regulation; Checked: 2026-06-16; Status: verified	E13: The EU AI Act places certain creditworthiness, credit scoring and life or health insurance risk-assessment or pricing systems in high-risk territory, which changes the approval route for customer-impacting models. \| E53: EU AI Act dates and high-risk classifications can change the diligence path for cross-border credit and insurance systems faster than annual model reviews.	Open source
12	EU DORA Regulation 2022/2554	Publisher: EU DORA Regulation 2022/2554; Kind: Regulation; Checked: 2026-06-16; Status: verified	E14: DORA sets ICT risk-management, incident-handling, digital-operational-resilience testing and ICT third-party risk-management requirements for EU financial entities. \| E54: DORA creates resilience and third-party-control evidence that should be attached to critical model routes, cloud routes and shared source providers.	Open source
13	OpenAI - solutions for financial services	Publisher: OpenAI; Kind: Vendor documentation; Checked: 2026-06-16; Status: verified	E15: OpenAI positions financial-services use cases around research, risk, client service, secure deployment, data connections and model use for complex financial environments. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
14	OpenAI - Morgan Stanley financial services case study	Publisher: OpenAI; Kind: Case study; Checked: 2026-06-16; Status: verified	E16: Morgan Stanley's OpenAI case study says more than 98% of advisor teams actively use AI @ Morgan Stanley Assistant and that the deployment used an eval framework, retrieval tuning and quality assurance. \| E17: The Morgan Stanley case study describes growth from answering 7,000 questions to effectively answering from a corpus of 100,000 documents, making source scale and retrieval evaluation central to the workflow.	Open source
15	Anthropic - agents for financial services	Publisher: Anthropic; Kind: Vendor documentation; Checked: 2026-06-16; Status: verified	E18: Anthropic's financial-services agent release lists ready-to-run templates for work such as pitchbooks, KYC file screening and month-end close, showing frontier model providers moving into financial workflow wrappers. \| E55: Anthropic's 2026 financial-services agent templates include banking, investment and finance-operations work, reinforcing the route distinction between models, agents and financial workflow wrappers. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
16	Anthropic - Claude for Financial Services	Publisher: Anthropic; Kind: Vendor documentation; Checked: 2026-06-16; Status: verified	E19: Anthropic's financial-services page describes Excel, market-data and portfolio-analytics connectors, Financial Analysis Solution routes and financial-work benchmarks; vendor-reported results are useful pattern evidence, not independent approval. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
17	OpenAI enterprise privacy	Publisher: OpenAI enterprise privacy; Kind: Privacy and data terms; Checked: 2026-06-16; Status: verified	E20: OpenAI enterprise privacy material is a primary source for data-use and control diligence on direct OpenAI routes.	Open source
18	Microsoft - Azure OpenAI data, privacy and security	Publisher: Microsoft; Kind: Cloud documentation; Checked: 2026-06-16; Status: verified	E21: Microsoft's Azure OpenAI data, privacy and security page is the primary route source when the same model family is procured through an Azure control plane. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
19	Google Cloud - Vertex AI data governance	Publisher: Google; Kind: Privacy and data terms; Checked: 2026-06-16; Status: verified	E22: Google Vertex AI data-governance docs are primary diligence sources for Gemini and managed-model routes deployed through Google Cloud.	Open source
20	AWS - generative AI for financial services	Publisher: AWS; Kind: Cloud documentation; Checked: 2026-06-16; Status: verified	E23: AWS financial-services generative AI material positions cloud-model routes around regulated financial workloads and managed cloud controls. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
21	BloombergGPT paper	Publisher: BloombergGPT paper; Kind: Financial benchmark; Checked: 2026-06-16; Status: verified	E24: BloombergGPT is a 50B-parameter language model trained on a finance-heavy corpus, with a dataset built from Bloomberg financial sources plus general-purpose text.	Open source
22	FinGPT paper	Publisher: FinGPT paper; Kind: Financial benchmark; Checked: 2026-06-16; Status: verified	E25: FinGPT presents an open-source, data-centric financial LLM framework, arguing that financial data freshness and curation are central to financial model performance.	Open source
23	FinanceBench paper	Publisher: FinanceBench paper; Kind: Financial benchmark; Checked: 2026-06-16; Status: verified	E26: FinanceBench contains 10,231 open-book financial QA questions with answers and evidence strings; on a public sample, GPT-4-Turbo with retrieval incorrectly answered or refused 81% of questions. \| E49: FinanceBench's evidence-string design shows why a route eval should keep source strings and accepted-output records, not only model answers.	Open source
24	FinBen paper	Publisher: FinBen paper; Kind: Financial benchmark; Checked: 2026-06-16; Status: verified	E27: FinBen evaluates financial LLMs across 36 datasets, 24 tasks and seven critical aspects, giving a broad task map rather than a production approval test.	Open source
25	TAT-QA benchmark	Publisher: TAT-QA benchmark; Kind: Financial benchmark; Checked: 2026-06-16; Status: verified	E28: TAT-QA tests question answering over financial tables and text, preserving the table-plus-narrative failure mode that appears in filings, capital calls, statements and policy documents.	Open source
26	ConvFinQA paper	Publisher: ConvFinQA paper; Kind: Financial benchmark; Checked: 2026-06-16; Status: verified	E29: ConvFinQA tests multi-turn numerical reasoning over financial reports, which helps expose calculation-chain and context-carry failures.	Open source
27	FinQA paper	Publisher: FinQA paper; Kind: Financial benchmark; Checked: 2026-06-16; Status: verified	E30: FinQA treats financial reasoning as program-style question answering over reports, making source cells and intermediate calculations part of the eval target.	Open source
28	Vals AI - Finance Agent benchmark	Publisher: Vals AI; Kind: Financial benchmark; Checked: 2026-06-16; Status: verified	E31: Vals AI's Finance Agent benchmark measures multi-step finance work involving search, calculations and final answer submission, making it useful for tool-use and workflow tests.	Open source
29	SEC - EDGAR APIs	Publisher: SEC; Kind: Public data API; Checked: 2026-06-16; Status: verified	E32: SEC EDGAR APIs are an official public route for company filings and metadata; they are more source-authoritative than unofficial filing copies or generic web search.	Open source
30	NIST AI Risk Management Framework	Publisher: NIST; Kind: Standard; Checked: 2026-06-16; Status: verified	E33: NIST AI RMF provides a general risk-management vocabulary for AI governance, mapping and measurement, but financial use still needs sector-specific model, conduct and resilience controls.	Open source
31	LSEG Workspace	Publisher: LSEG Workspace; Kind: Financial data; Checked: 2026-06-16; Status: verified	E34: LSEG Workspace positions the licensed-data layer around market data, news, analytics and workflow access; for capital markets, source entitlement is part of model capability.	Open source
32	FactSet Mercury	Publisher: FactSet Mercury; Kind: Financial data; Checked: 2026-06-16; Status: verified	E35: FactSet Mercury is positioned as an AI workflow over financial data and analytics, illustrating the source-plus-workflow path rather than a standalone foundation-model purchase.	Open source
33	S&P Global Kensho	Publisher: S&P Global Kensho; Kind: Financial data; Checked: 2026-06-16; Status: verified	E36: S&P Global Kensho offers financial NLP, extraction and AI products over market and reference-data contexts, making financial-source ownership a route advantage.	Open source
34	AlphaSense generative search	Publisher: AlphaSense generative search; Kind: Financial data; Checked: 2026-06-16; Status: verified	E37: AlphaSense generative search uses filings, research, transcripts and expert-call content as a financial-research route; source rights and provenance determine what the model may know.	Open source
35	Moody's research assistant	Publisher: Moody's research assistant; Kind: Financial data; Checked: 2026-06-16; Status: verified	E38: Moody's research assistant sits on credit research and analytics, which is a different approval problem from a generic chat route.	Open source
36	NICE Actimize Xceed	Publisher: NICE Actimize Xceed; Kind: Financial crime; Checked: 2026-06-16; Status: verified	E39: NICE Actimize Xceed is a financial-crime and case-management route, where false negatives, investigator workflow and audit trails matter more than fluent summaries.	Open source
37	FICO Falcon Fraud Manager	Publisher: FICO Falcon Fraud Manager; Kind: Financial crime; Checked: 2026-06-16; Status: verified	E40: FICO Falcon is a fraud decisioning model route, making latency, false positives, false negatives and explainability central acceptance criteria.	Open source
38	Feedzai risk operations platform	Publisher: Feedzai risk operations platform; Kind: Financial crime; Checked: 2026-06-16; Status: verified	E41: Feedzai positions its platform around financial-crime risk operations; these routes need investigation outcomes, not only better alert text.	Open source
39	Featurespace ARIC Risk Hub	Publisher: Featurespace ARIC Risk Hub; Kind: Financial crime; Checked: 2026-06-16; Status: verified	E42: Featurespace ARIC Risk Hub is a behavioural-analytics and fraud route, illustrating that banking model choices often sit inside specialised fraud systems.	Open source
40	Mastercard Decision Intelligence	Publisher: Mastercard Decision Intelligence; Kind: Payments AI; Checked: 2026-06-16; Status: verified	E43: Mastercard Decision Intelligence is a network-scale payment authorization and fraud route, where response time, decline rates and fraud loss are production metrics.	Open source
41	Visa Advanced Authorization	Publisher: Visa Advanced Authorization; Kind: Payments AI; Checked: 2026-06-16; Status: verified	E44: Visa Advanced Authorization is a payment authorization risk route, making payment-network AI materially different from document or research models.	Open source
42	Guidewire predictive analytics	Publisher: Guidewire predictive analytics; Kind: Insurance AI; Checked: 2026-06-16; Status: verified	E45: Guidewire predictive analytics shows insurance AI embedded in core insurance systems, not merely in an external model API.	Open source
43	Shift Technology claims fraud detection	Publisher: Shift Technology claims fraud detection; Kind: Insurance AI; Checked: 2026-06-16; Status: verified	E46: Shift Technology's claims-fraud route shows insurance models operating against claims evidence, fraud signals and examiner workflows.	Open source
44	Tractable insurance AI	Publisher: Tractable insurance AI; Kind: Insurance AI; Checked: 2026-06-16; Status: verified	E47: Tractable's insurance AI route shows computer-vision and claims workflows as separate model classes from language-model research assistants.	Open source
45	OCC Bulletin 2026-13 model risk management revised guidance	Publisher: OCC Bulletin 2026-13 model risk management revised guidance; Kind: Regulator; Checked: 2026-06-16; Status: verified	E56: OCC Bulletin 2026-13 issues revised model-risk guidance while stating that generative AI and agentic AI models are novel, rapidly evolving and outside the scope of that guidance. \| E67: OCC Bulletin 2026-13 says agencies expect to consider banks' use of AI, including generative and agentic AI, in future model-risk work.	Open source
46	Federal Reserve SR 26-2 revised model risk management attachment	Publisher: Federal Reserve SR 26-2 revised model risk management attachment; Kind: Regulator; Checked: 2026-06-16; Status: verified	E57: The revised interagency model-risk guidance covers model development and use, validation and monitoring, governance and controls, and vendor or third-party products.	Open source
47	U.S. Treasury - artificial intelligence in financial services	Publisher: U.S. Treasury; Kind: Regulator; Checked: 2026-06-16; Status: verified	E58: The U.S. Treasury report maps AI use across financial services, including underwriting, fraud detection, AML/CFT, customer service, risk management, cybersecurity and operations. \| E69: Treasury's cross-sector use-case map supports early value in controlled research, customer operations, cybersecurity, compliance support and risk operations rather than autonomous regulated decisions.	Open source
48	FSB - monitoring adoption of AI and related vulnerabilities	Publisher: FSB; Kind: Regulator; Checked: 2026-06-16; Status: verified	E59: The FSB's 2025 monitoring report treats AI adoption, related vulnerabilities, provider concentration and supply-chain dependency as financial-stability monitoring questions. \| E70: FSB monitoring advice makes operational signals such as adoption concentration, outsourcing dependency and common-provider exposure relevant route-review triggers.	Open source
49	IOSCO - artificial intelligence in capital markets	Publisher: IOSCO; Kind: Regulator; Checked: 2026-06-16; Status: verified	E60: IOSCO maps capital-markets AI use cases across advice, trading, research, surveillance and AML/CFT, with data, model, outsourcing and human-interaction risks. \| E68: IOSCO's capital-markets paper supports using AI where source-backed research, surveillance support and supervisory workflows gain speed without handing over execution authority.	Open source
50	FINRA - 2026 annual regulatory oversight report on GenAI	Publisher: FINRA; Kind: Regulator; Checked: 2026-06-16; Status: verified	E61: FINRA's 2026 GenAI material identifies summarization and information extraction as common use patterns while flagging supervision, data and guardrail issues for member firms.	Open source
51	CFPB Circular 2022-03 on complex algorithms and adverse-action notices	Publisher: CFPB Circular 2022-03 on complex algorithms and adverse-action notices; Kind: Regulator; Checked: 2026-06-16; Status: verified	E62: CFPB Circular 2022-03 says creditors using complex algorithms must still provide specific and accurate reasons for adverse action.	Open source
52	EBA - machine learning in internal ratings-based models	Publisher: EBA; Kind: Regulator; Checked: 2026-06-16; Status: verified	E63: EBA's machine-learning work on internal ratings-based models shows that bank credit-model validation remains use-case-specific and evidence-heavy.	Open source
53	NYDFS Circular Letter No. 7 on AI and external consumer data in insurance	Publisher: NYDFS Circular Letter No. 7 on AI and external consumer data in insurance; Kind: Regulator; Checked: 2026-06-16; Status: verified	E64: NYDFS Circular Letter No. 7 sets expectations for insurers using AI systems and external consumer data in underwriting and pricing, including governance, proxy, impact, testing, documentation and complaint-handling controls.	Open source
54	EIOPA - AI governance principles for insurance	Publisher: EIOPA; Kind: Regulator; Checked: 2026-06-16; Status: verified	E65: EIOPA's insurance AI governance principles anchor ethical, trustworthy, fair and explainable AI expectations for European insurance routes.	Open source
55	SEC - AI washing enforcement release	Publisher: SEC; Kind: Regulator; Checked: 2026-06-16; Status: verified	E66: The SEC's 2024 AI-washing enforcement release shows that public claims about AI use in investment processes must match substantiated operations.	Open source
56	Artificial Analysis - model comparison and Intelligence Index cost	Publisher: Artificial Analysis; Kind: Model benchmark; Checked: 2026-06-16; Status: verified	E71: Artificial Analysis publishes per-model intelligence, token pricing, speed, latency, context and the cost to run the Artificial Analysis Intelligence Index. \| E105: Artificial Analysis checked on 31 May 2026 supplies the extracted historical rows used for the full price-decline chart series, including date, model, Intelligence Index, blended token price and benchmark bill. Current blended-price tier winners are 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; current Intelligence Index cost winners are 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.	Open source
57	a16z - LLMflation: from a cost miracle to the margin wars	Publisher: a16z; Kind: Market analysis; Checked: 2026-06-16; Status: verified	E72: 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.	Open source
58	OpenAI status page	Publisher: OpenAI status page; Kind: Status page; Checked: 2026-06-16; Status: verified	E73: 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.	Open source
59	Claude status page	Publisher: Claude status page; Kind: Status page; Checked: 2026-06-16; Status: verified	E74: 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.	Open source
60	Microsoft Azure OpenAI Service	Publisher: Microsoft; Kind: Cloud documentation; Checked: 2026-06-16; Status: verified	E75: Microsoft's Azure OpenAI Service page states that Azure OpenAI is guaranteed to be available at least 99.9% of the time. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
61	Google Cloud Service Health JSON history	Publisher: Google; Kind: Status page; Checked: 2026-06-16; Status: verified	E76: 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.	Open source
62	Azure status history - Azure OpenAI Service	Publisher: Azure; Kind: Status page; Checked: 2026-06-16; Status: verified	E77: 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.	Open source
63	Vertex AI Platform Service Level Agreement	Publisher: Vertex AI Platform Service Level Agreement; Kind: Cloud documentation; Checked: 2026-06-16; Status: verified	E78: The Vertex AI Platform SLA lists monthly uptime percentage commitments including at least 99.9% for covered training, deployment and batch-prediction services. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
64	OpenAI GPT-5.5 model rate limits	Publisher: OpenAI GPT-5.5 model rate limits; Kind: Official model documentation; Checked: 2026-06-16; Status: verified	E79: 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.	Open source
65	OpenAI Batch API	Publisher: OpenAI Batch API; Kind: Official model documentation; Checked: 2026-06-16; Status: verified	E80: 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.	Open source
66	Claude API rate limits	Publisher: Claude API rate limits; Kind: Official model documentation; Checked: 2026-06-16; Status: verified	E81: 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.	Open source
67	Groq API rate limits	Publisher: Groq; Kind: Official model documentation; Checked: 2026-06-16; Status: verified	E82: 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.	Open source
68	Azure OpenAI quotas and limits	Publisher: Azure; Kind: Cloud documentation; Checked: 2026-06-16; Status: verified	E83: 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: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
69	Azure OpenAI provisioned throughput	Publisher: Azure; Kind: Cloud documentation; Checked: 2026-06-16; Status: verified	E84: 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: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
70	Vertex AI Standard PayGo throughput	Publisher: Vertex AI Standard PayGo throughput; Kind: Cloud documentation; Checked: 2026-06-16; Status: verified	E85: 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: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
71	Vertex AI Provisioned Throughput	Publisher: Vertex AI Provisioned Throughput; Kind: Cloud documentation; Checked: 2026-06-16; Status: verified	E86: Vertex AI Provisioned Throughput is a dedicated-capacity route with request-type controls and monitoring, distinct from shared PayGo throughput. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
72	Amazon Bedrock quota increase mechanism	Publisher: Amazon; Kind: Cloud documentation; Checked: 2026-06-16; Status: verified	E87: 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: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
73	Amazon Bedrock cross-Region inference	Publisher: Amazon; Kind: Cloud documentation; Checked: 2026-06-16; Status: verified	E88: 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: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
74	Amazon Bedrock Provisioned Throughput	Publisher: Amazon; Kind: Cloud documentation; Checked: 2026-06-16; Status: verified	E89: 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: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
75	OpenAI Flex processing	Publisher: OpenAI Flex processing; Kind: Official model documentation; Checked: 2026-06-16; Status: verified	E90: 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.	Open source
76	OpenAI prompt caching	Publisher: OpenAI prompt caching; Kind: Official model documentation; Checked: 2026-06-16; Status: verified	E91: 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.	Open source
77	Anthropic Message Batches	Publisher: Anthropic; Kind: Official model documentation; Checked: 2026-06-16; Status: verified	E92: Anthropic Message Batches support high-volume asynchronous Claude requests with discounted pricing and can be combined with prompt caching for repeated context.	Open source
78	Anthropic prompt caching	Publisher: Anthropic; Kind: Official model documentation; Checked: 2026-06-16; Status: verified	E93: 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.	Open source
79	Vertex AI generative AI pricing	Publisher: Vertex AI generative AI pricing; Kind: Pricing page; Checked: 2026-06-16; Status: verified	E94: 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: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
80	Vertex AI context caching	Publisher: Vertex AI context caching; Kind: Cloud documentation; Checked: 2026-06-16; Status: verified	E95: 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: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
81	Vertex AI Gemini batch prediction	Publisher: Vertex AI Gemini batch prediction; Kind: Cloud documentation; Checked: 2026-06-16; Status: verified	E96: 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: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
82	Amazon Bedrock pricing	Publisher: Amazon; Kind: Pricing page; Checked: 2026-06-16; Status: verified	E97: 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: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
83	Amazon Bedrock prompt caching	Publisher: Amazon; Kind: Cloud documentation; Checked: 2026-06-16; Status: verified	E98: Amazon Bedrock prompt caching uses cache points with TTL controls and model support limits, reducing repeated-context processing where the route is eligible. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
84	Amazon Bedrock batch inference	Publisher: Amazon; Kind: Cloud documentation; Checked: 2026-06-16; Status: verified	E99: 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: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
85	Groq prompt caching	Publisher: Groq; Kind: Official model documentation; Checked: 2026-06-16; Status: verified	E100: 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.	Open source
86	Groq Batch API	Publisher: Groq; Kind: Official model documentation; Checked: 2026-06-16; Status: verified	E101: 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.	Open source
87	Groq Flex processing	Publisher: Groq; Kind: Official model documentation; Checked: 2026-06-16; Status: verified	E102: 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.	Open source
88	Harvey - Legal Agent Benchmark initial results	Publisher: Harvey; Kind: Agent benchmark; Checked: 2026-06-16; Status: verified	E103: 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.	Open source
89	Rogo - BigFinanceBench announcement	Publisher: Rogo; Kind: Financial benchmark; Checked: 2026-06-16; Status: verified	E104: Rogo's May 2026 BigFinanceBench announcement describes 928 workflow-grounded finance questions, 15,656 rubric criteria and 36,241 weighted points; it reports no single top model dominates all workflows and a workflow/source router improves rubric score by 4.5 percentage points over the best single-model baseline.	Open source
90	Silicon Valley Bank - State of Enterprise Software Report 2026	Publisher: Silicon Valley Bank; Kind: Market survey; Checked: 2026-06-16; Status: verified	E106: Silicon Valley Bank's 2026 State of Enterprise Software survey of 120+ companies reports that about 48% saw 25% or greater employee-productivity improvement from AI, while half said revenue-growth attribution was hard to measure and one-quarter put revenue lift below 10%. \| E107: Silicon Valley Bank reports that nearly 40% of surveyed enterprise-software companies are primarily seat-based today, but only 26% expect to remain primarily seat-based as AI labour augmentation and compute economics change pricing. \| E108: Silicon Valley Bank reports that five hyperscalers' combined operating cash flow rose from $116 billion in 2016 to $577 billion in 2025 while CapEx increased from 28% to 66% of operating cash flow, underscoring capital intensity behind AI infrastructure capacity.	Open source

▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒Glossary

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.

AI models for banking, capital markets, and insurance

The colophon