AI Agents in Financial Services: What Ships and What Regulators Expect

AI agents in financial services ship today where a person still makes the binding call: AML and transaction-screening triage, document review for onboarding and lending, small-claims intake, advisor and servicing support, internal knowledge search, and code. Credit decisions and customer-facing advice are still risky. No US, EU, or UK regulator has written agent-specific rules, so existing rules apply.

One thing changed in April 2026: US bank regulators replaced SR 11-7 with model risk guidance that excludes generative and agentic AI.

Where do AI agents in financial services run in production?

The disclosures share one pattern: AI prepares, checks, or drafts, and a person or a validated model makes the binding decision.

JPMorganChase. In the bank’s 2025 line-of-business letters, published in April 2026, the documented agent example is in wealth management: the Connect Coach tool, used by 12,000 people, includes 25 specialized AI agents that suggest client outreach to advisors. Elsewhere the bank reports AI results without naming an agentic system. The Commercial and Investment Bank says AI has let it “review more than double the volume while halving the number of manual operator checks” in transaction screening.

The consumer bank letter describes mature predictive AI and machine learning use cases “in credit, operations and fraud,” and an “increased focus on agentic capabilities in coding, marketing, sales optimization, pricing and personalization.” My read: credit stays with predictive models, and agents go to other work.

Citi. A September 2026 Citi article says its Stylus Workspaces platform has reached more than 180,000 people. In U.S. Services onboarding, AI cut document review from about an hour to 15 minutes. Citi used the Devin coding agent to re-map more than 30 years of legacy code in two days. Citi says “a human is always in the loop.”

Allianz. Project Nemo, launched in Australia in July 2025, runs seven task-specific agents on food spoilage claims up to AUD 500. Allianz reports an 80% cut in processing and settlement time. “By design, payout decisions are never automated.”

All figures are company-reported, not audited.

How widely are banks and insurers using AI agents?

Broad AI use is common; agents at scale are not. The latest joint survey from the Bank of England and the FCA, published November 21, 2024, covered 118 UK firms. 75% were already using AI, including 95% of insurers and 94% of international banks. 55% of AI use cases had some degree of automated decision-making, but only 2% were fully autonomous. 46% of firms said they had only partial understanding of the AI they use. I found no newer edition.

For agents, KPMG’s banking edition of its Q2 2026 AI Quarterly Pulse, fielded April 28 to May 25, 2026, reports 51% of banks piloting AI agents, 24% scaling them across multiple functions, 15% orchestrating multiple agents across workflows, and 10% still exploring. KPMG describes the survey as 204 U.S. leaders at organizations with $1 billion or more in revenue. The banking PDF does not say how many are banks, so treat the split as directional.

A warning sign: 24% said overrides of AI output happen “on a case-by-case basis without formal criteria.”

Failure mode to catch early: if your override rule is “the analyst uses judgment,” you cannot show an examiner when a person should have stepped in. Write the triggers down. The most common one in the KPMG data (52%) is an output that conflicts with regulatory or legal requirements. Start there.

Use-case table: what ships, what is risky, and what each needs

Maturity is my read of the disclosures above. Two rows have their own guides: KYC onboarding and insurance claims intake.

Why credit decisions and customer advice are still risky

Credit: the reasons problem. Under Regulation B, the statement of reasons for adverse action “must be specific and indicate the principal reason(s) for the adverse action” (12 CFR 1002.9(b)(2)). A validated scorecard can rank the factors behind a decline. An LLM agent that reads a file and decides cannot reliably say which inputs drove its answer. The CFPB withdrew its 2022 and 2023 circulars on adverse action and complex algorithms on May 12, 2025. The regulation text did not change, so the duty stays.

In the EU, AI used “to evaluate the creditworthiness of natural persons or establish their credit score” is high-risk under Annex III, point 5(b) of the AI Act. Risk assessment and pricing for life and health insurance is point 5(c). Deployers of both must also run a fundamental rights impact assessment under Article 27.

Advice: the supervision problem. FINRA’s 2026 oversight report, published December 2025, lists agent risks, including autonomy without human validation, acting beyond the user’s intended scope and authority, and multi-step reasoning that is hard to audit. Rule 3110 requires a supervisory system reasonably designed for the business. FINRA does not say how to sample. My risk judgment: sampling alone may be enough for a chatbot answering advisor questions, but it may not be enough for an agent recommending products to thousands of customers.

Decision rule: if the agent’s output reaches a customer as a decision or a recommendation with no person in between, treat it as a separate, higher tier. My recommendation: start that tier only after your internal use cases have two quarters of clean monitoring.

What do US regulators expect from AI agents in banking?

No agent rule exists, and the model risk guidance now excludes agents.

On April 17, 2026, the Fed, FDIC, and OCC issued revised model risk guidance: SR 26-2 at the Fed and OCC Bulletin 2026-13. It replaces SR 11-7 and the 2021 statement on BSA/AML models. It says: “Generative AI and agentic AI models are novel and rapidly evolving. As such, they are not within the scope of this guidance.” The guidance says a bank’s own governance should set controls for tools outside its scope. It is most relevant to banks above $30 billion in assets and “does not set forth enforceable standards or prescriptive requirements.” The agencies plan a request for information on AI, including agentic AI, “in the near future.” I could not find it published as of October 15, 2026.

The gap: an agent that is “not a model” can land in nobody’s inventory. Yet most agents wrap real models, and the guidance still warns that unsafe or unsound practices can draw supervisory action. My recommendation: run the agent through your model risk process anyway.

NYDFS’s October 16, 2024 industry letter adds no new requirements. It explains how Part 500 applies to AI. Risk assessments must cover deepfakes and should cover your own and your vendors’ AI. MFA has been required for all authorized users since November 2025, though § 500.12(b) lets a CISO approve in writing reasonably equivalent or more secure compensating controls. An inventory of all information systems has been required since November 1, 2025. The letter also recommends, without requiring, that firms “consider monitoring for unusual query behaviors that might indicate an attempt to extract NPI.”

EU and UK rules: AI Act, DORA, FCA, and PRA

EU AI Act. The Digital Omnibus, Regulation (EU) 2026/1744, moved the start date for Annex III high-risk systems to December 2, 2027. Classification depends on each system’s intended purpose. Alert prep and onboarding review are not 5(b) or 5(c) uses as such, but check each against every Annex III category. An onboarding flow that uses one-to-many biometric identification, or feeds a credit assessment, can land in scope. Article 50 transparency duties for AI that interacts directly with people have applied since August 2, 2026. The EU AI Act guide covers both.

DORA. The Digital Operational Resilience Act has applied since January 17, 2025. Article 28(3) requires a register of information on “all contractual arrangements on the use of ICT services provided by ICT third-party service providers.” A hosted model API, an agent platform, and a managed vector database each need an entry. Article 19 requires reporting major ICT-related incidents, which an agent failure can become.

UK FCA. The FCA published the Mills Review on July 6, 2026, with seven recommendations, including monitoring the move to autonomous models. The press release announces no new AI-specific rules. The FCA chair points to the Consumer Duty and the Senior Managers Regime. My read: name the senior manager who owns the agent’s outcomes before launch.

UK PRA. The PRA’s SS1/23 model risk principles apply to UK banks, building societies, and PRA-designated investment firms with internal model approval for regulatory capital. The current version took effect April 23, 2026, and covers the risks of AI used in modeling techniques.

Decision rule: if you operate in the EU and an agent touches credit scoring or life and health pricing, plan for December 2027 now. Either way, DORA applies today alongside Article 50 transparency, GDPR, and your sector rules.

A model-risk checklist for an AI agent

Run this before production, then quarterly. It applies SR 26-2 ideas (inventory, materiality, effective challenge, monitoring) to an agent the guidance excludes. The thresholds are mine.

Checks 3, 8, 9, and 13 are the four I would never waive. Details: what to log for an agent audit trail, change control for prompts, models, and tools, and acceptance tests for agents.

Worked example: sizing an AML alert-prep agent

Hypothetical. Say a financial crime team works 12,000 alerts a month at 40 minutes each, and 25 of those minutes go to gathering evidence across systems. An agent assembles the evidence packet and drafts the narrative. The analyst reviews and decides.

If gathering drops from 25 to 8 minutes, the team saves 17 minutes per alert: 204,000 minutes, or 3,400 hours a month. Now pay for control. A second reviewer rechecks 5% of packets: 600 alerts at 15 minutes, or 150 hours. Net saving: 3,250 hours a month. At 140 productive hours per analyst, that is about 23 analysts’ worth of capacity.

Run the low case. If gathering only drops to 15 minutes, the saving is 10 minutes per alert, or 2,000 hours. Net of the same 150 review hours, that is 1,850 hours, about 13 analysts. The case still holds at a little over half the benefit.

The math assumes no headcount cuts. My recommendation: point the capacity at the backlog and harder cases first.

Failure mode: the approval rate climbs toward 100% while review time per alert falls under two minutes. That pattern suggests reviewers are rubber-stamping. Check 9 catches it before an examiner does.

Bottom Line

AI agents in financial services ship when they prepare the work and a person or a validated model decides.

SR 26-2 left agents out of scope. That is a reason to govern them with your own model risk process, not to skip it.

Pick one internal prep workflow. Run the 14-point checklist. Measure capacity and escapes before you widen scope.

FAQ

Does SR 11-7 still apply to AI agents at US banks?

No. SR 26-2 superseded it on April 17, 2026. The OCC also rescinded its Model Risk Management handbook booklet and Bulletin 1997-24 on credit scoring models.

Can an LLM draft adverse action notices?

It can draft them. But Regulation B says reasons such as the creditor’s internal standards or failing a qualifying score are insufficient. A person should confirm the notice matches what drove the decision.

Is AI fraud detection high-risk under the EU AI Act?

Not under point 5(b). That point excludes AI systems used to detect financial fraud from the creditworthiness category. The exclusion applies only to 5(b), so check a fraud tool’s intended purpose against the other Annex III categories. Fraud tools still fall under DORA and data protection law.

Do consumers want AI agents managing their money?

Some do. In an April 2026 survey of more than 5,000 UK retail financial services consumers for the FCA’s Mills Review, a fifth said they are likely to use AI that can act autonomously within pre-set goals. The FCA puts that at 11 million UK adults.

How much financial-services AI is bought rather than built?

In the 2024 Bank of England and FCA survey, a third of AI use cases were third-party implementations, and the top three model providers made up 44% of all model providers reported. That concentration is why I treat the DORA register as a live control.

Three Ways to Work With TMA

Need an agent built? We deploy production AI agents in your infrastructure. Working pilot. Real data. Measurable ROI. → Schedule Demo

Want to co-build a product? We’re not a dev agency. We’re co-builders. Shared cost. Shared upside. → Partner with Us

Want to join the Guild? Ship pilots, earn bounties, share profit. Community + equity + path to exit. → Become an AI Architect