If your bank has AI agents in production, you are in the majority, and you are probably watching them with controls built for models that do not act. BCG's new benchmark of 26 global and domestic systemically important banks found that 73% permit agents in production for select use cases, while only 8% have a comprehensive, standardized monitoring and control framework that explicitly covers agents. The same study found that banks using preapproved patterns for lower-risk AI were about twice as likely to get those applications from idea to production in under six months. For a bank CRO or CIO, the work before the next wave of agents is a fast lane for repeatable uses and monitoring built for agents.
The report, AI Safety and Governance in Banking 2026, was published on November 13. What follows covers what BCG measured, why the gap sits where it does, how US, UK and EU rules bear on it, and what to start this month.
What Did BCG Measure in 26 Banks?
BCG scored each bank on two axes, AI adoption and AI risk maturity, and the summary on bcg.com sorts the 26 into four broad groups, from banks still building foundations to those pairing higher adoption with more mature risk management. The adoption numbers show broad permission and narrow use:
| Finding | Share of the 26 banks |
|---|---|
| Piloting or permitting agentic AI | 92% |
| Agents permitted in production for select use cases | 73% |
| Agents permitted for broad production use | 0% |
| Agents are 10% or less of the AI portfolio | 88% |
| Restrictions or prohibitions on agents written into the risk appetite | 54% |
| Comprehensive, standardized monitoring framework covering agents | 8% |
Source: BCG via PR Newswire, November 13, 2026.
The paper foundations look better. According to the release, 96% manage AI risk through existing risk frameworks, 73% have defined an AI risk appetite (up from 57% in 2025), and 77% have a dedicated AI policy. The operating layer underneath is thinner: fewer than half have a centralized, tool-based AI inventory, and none reports a comprehensive human-oversight framework applied consistently across its AI systems.
Things are already going wrong at that scale. 46% of the banks reported an AI-related issue or incident in the past 12 months, and more than a quarter said AI systems were falling behind their promised business case.
Read the sample for what it is. Twenty-six systemically important banks are the institutions with the largest risk functions in the industry, so a regional or community bank should assume its own numbers are worse, not better. The public summary does not name the banks or give a regional split, and the maturity matrix appears only as an image.
Why Does Faster Approval Come From More Governance?
The banks with the strongest risk management ship more AI. BCG found that banks in the top third for risk-management maturity were 2.5 times as likely to be advanced AI adopters as those in the bottom third. "The banks that are applying governance best in day-to-day operations are the ones scaling AI most extensively," Matteo Coppola, global leader of BCG's Risk and Compliance practice and a coauthor, said in the release.
The mechanism BCG points to is the preapproved pattern. A preapproved pattern is a reusable design for a class of low-risk use cases (the data it may touch, the model route, the controls and the monitoring) that the risk function signs off once, so each new instance goes through a lighter review instead of a full one. Deep review is then saved for novel and higher-risk applications. Banks that work this way were about twice as likely to move lower-risk applications into production in under six months. The bcg.com summary says "twice"; the release says "nearly twice."
The strongest objection is that this is a correlation from one survey, and the causation could run the other way: a bank with a large, well-funded risk function may simply be the kind of bank that can afford to ship more AI, and its maturity score could reflect that rather than produce it. BCG does not claim to have isolated the effect. Even so, the direction of the advice holds up on cost alone. Every hour a model risk team spends re-reviewing the fortieth document summarizer is an hour it does not spend on the agent that can move money.
Compare the approach at BNY. Microsoft's case study says the bank has nearly 140 "digital employees", each with login credentials, an employee number in the corporate directory and a human supervisor who reviews its output, and that a governance flowchart lays out sixteen steps of review boards, compliance gates and model risk evaluations between an idea and a working agent. Sixteen steps is defensible for an agent with its own identity. Applied unchanged to every low-risk use case, it is the queue BCG's fast-lane banks avoid.
Where Is the Agent Governance Gap?
The gap is in monitoring what agents do after approval. BCG says even banks with relatively high overall AI risk maturity tend to be less prepared to govern agentic AI, and it names two forces that make the problem harder: technology providers can change AI capabilities after a bank has approved them, and business users can increasingly build or configure their own agents. Both make it harder to know which AI systems are running, how their risks are shifting, and whether the controls still apply.
Model governance was built around a point-in-time decision. A model is validated, approved and then checked on a schedule. An agent calls tools, reads changing data and can take actions, so its risk can move between reviews without anyone changing a line of the bank's own code. A vendor switching on a new agent feature inside a licensed product is the clearest case. Our coverage of shadow agents in the Microsoft stack and the 44% visibility gap on business-built agents found the same blind spot outside banking.
The inventory number matters here. With fewer than half of these banks running a centralized, tool-based AI inventory, most cannot answer the first question an examiner asks: which agents are live, who owns each one, and what each can touch. Products such as Microsoft Agent 365, Credo AI and Arize AX sell pieces of that inventory and monitoring layer. None of them is a substitute for deciding, as a bank, what "monitored" means for an agent.
What Do Regulators Expect From Banks on AI Agents?
In the US, the formal model risk guidance leaves agents out, and the examiner material is filling the space. The Fed, OCC and FDIC replaced SR 11-7 on April 17, 2026 with SR 26-2, and OCC Bulletin 2026-13 states: "Generative AI and agentic AI models are novel and rapidly evolving. As such, they are not within the scope of this guidance." The agencies said they plan a request for information covering banks' use of AI, including agentic AI. A BankThink piece in American Banker argued that the exclusion leaves agentic AI risk "for bankers to sort out themselves."
Examiners have not waited for the RFI. OCC Bulletin 2026-54, the November 4 examiner guide, tiers bank AI by the consequence of a failure and points examiners at the inventory, the tiers, the owners and the ability to stop a running agent. That followed the agencies' October questions on agent safeguards.
In the UK, the PRA's SS1/23 model risk principles cover all models used to inform business decisions, built in house or bought from vendors, regardless of technology, with a current version effective April 23, 2026. The Bank of England and FCA survey found UK agents rarely act alone yet.
In the EU, third-party rules already reach the vendor-change problem BCG describes. DORA's Article 28 requires a register of information on every ICT third-party arrangement and exit strategies for ICT services supporting critical or important functions, documented, "sufficiently tested and reviewed periodically." Under the AI Act, Annex III lists AI systems used to evaluate the creditworthiness of natural persons as high-risk; the provisional Omnibus deal of May 7, 2026 pushes the standalone high-risk requirements to December 2, 2027.
Concentration is the systemic version of the same issue. The Financial Stability Board named third-party dependencies and service provider concentration as one of four AI vulnerabilities that could raise systemic risk, in November 2024. BCG's recommendations ask banks to extend oversight to provider concentration, fallback arrangements and exit plans, but the public summary carries no numbers on how many banks have them.
What Should a Bank CRO or CIO Do Now?
Build the fast lane and the agent monitoring together, before the agent share of your portfolio grows past the 10% most banks sit under today.
This Week:
- Pull your AI inventory and count the agents in it, then ask each business line for the agents it built or configured itself. The gap between those two numbers is your first finding.
- List every licensed product where a vendor can switch on agent features without a change ticket at your bank. Start with your productivity suite and your CRM.
This Month:
- Draft three preapproved patterns for your highest-volume low-risk uses (internal document summarization, internal search, drafting with human sign-off). For each one, write down the permitted data classes, the approved model routes, the required logging and the reviewer, and get the CRO's signature once.
- Define what monitoring means for an agent at your bank: which actions are logged, which actions need a human before execution, who receives an alert, and how a running agent is stopped. Test the stop path on one agent in production.
- Add an "agent features changed" line to vendor reviews, so a capability change at the provider triggers a re-tier instead of passing silently.
Before the Next Exam or Vendor Renewal:
- For each AI provider behind a critical or important function, write the exit plan: the fallback model, the data you need to take with you, and how long the transition takes. In the EU, DORA already requires this; elsewhere, it covers the provider concentration, fallback and exit planning BCG recommends.
- Report the fast-lane numbers to the board quarterly: time from idea to production for pattern-based uses against fully reviewed ones, and incidents per pattern.
The Bottom Line
Banks have seen a version of this with end-user computing, when spreadsheets and macros spread through business lines faster than anyone inventoried them. Agents are spreading the same way, and an agent can take actions a spreadsheet never could. BCG's data says the largest banks have the policies and lack the operating controls: 77% have an AI policy, 8% have agent monitoring. Write your first three patterns and test one agent's stop path this month.
Continue Reading
- OCC Guide Says Bank AI Risk Depends on Failure, Not Sophistication
- US Regulators Ask Banks to Explain Their AI Agent Safeguards
- Bank of England and FCA Find UK AI Agents Rarely Act Alone
- Northern Trust Cuts Onboarding Data Prep 83% With AI Agent
- Bank AI-Related Listings Rise 49% as Agent Orchestration Grows
- Best MLOps for Regulated AI: Domino, Then a Sign-Off Layer
- AT&T Keeps Human Review on Nearly 500 AI Workflows