Seventy-three percent of the world’s systemically important banks now permit AI agents in production for select use cases. Only 8% have a comprehensive, standardized framework to monitor and control them. That gap sits at the center of AI Safety and Governance in Banking 2026, a benchmark Boston Consulting Group published on November 13, 2026, covering 26 global and domestic systemically important banks.
The same benchmark found that nearly half of those banks (46%) 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. Agentic AI in banking has moved past the pilot stage. The ability to see what those agents do, what they cost and what they return has not kept pace.
Banks are not alone in this. A day earlier, HFS Research and Rackspace reported that only 33% of AI-active enterprises have met or exceeded the business case for their most significant AI program, and 96% report significant governance gaps. Two studies, two different samples, one conclusion: the policies exist, and the operating controls behind them mostly do not.
What BCG found
The headline is not that banks are reckless with agents. They are cautious. Ninety-two percent are piloting or permitting agentic AI, but none permit broad production use, and at 88% of banks, agents make up 10% or less of the AI portfolio. Fifty-four percent have written restrictions or prohibitions on agentic AI into their risk-appetite frameworks.
Read that carefully. More than half of these banks are managing agent risk by limiting what agents are allowed to do, because they cannot yet watch what agents actually do. Restriction is a reasonable stopgap. It is also a ceiling. A bank that can only govern agents by prohibiting them will scale them slowly, if at all.
The governance paperwork, by contrast, is largely in place. Almost all banks in the benchmark (96%) manage AI risk through their existing risk frameworks. Seventy-three percent have defined their appetite for AI risk, up from 57% in 2025, and 77% have a dedicated AI policy. But fewer than half have a centralized, tool-based AI inventory, and none have a comprehensive human-oversight framework that is consistently implemented across AI systems.
Anne Kleppe, BCG’s global lead for AI safety and governance and a coauthor of the report, framed the next step as “turning governance from a framework into an operating capability.” With agents still a small share of AI portfolios at most banks, there is a window to get that right before their use expands, according to the same statement.
Why agentic AI in banking breaks the old oversight model
Traditional model risk management was built for models that produce a score or a forecast and hand it to a person. Agents are different. BCG describes them as systems that can plan and execute multistep tasks, take actions and adjust their approach without human input at every step. Each of those properties weakens a control that banks rely on today.
- A periodic model review assumes the model behaves the same way between reviews. An agent that chooses its own tools and steps can behave differently on every run.
- A human sign-off assumes a person sees each consequential output. An agent that executes a multistep task may take several actions before anyone looks.
- An inventory assumes you know what is deployed. Fewer than half of the banks in the benchmark have a centralized, tool-based AI inventory, and agents can be built inside platforms that business teams already own.
This is why an inventory is the first control, not an administrative afterthought. You cannot monitor an agent you have not registered, and you cannot attribute its cost or its incidents to an owner if nobody knows who owns it. We made a similar argument about the need for an AI system of record: every other function in the enterprise has one, and agents are about to make the absence expensive.
Governance and speed are not a tradeoff
The most useful finding in the benchmark is the one that cuts against conventional wisdom. Banks with more mature risk management are 2.5 times more likely to be advanced AI adopters. Banks that use standardized, previously approved patterns for lower-risk AI applications are nearly twice as likely to move those applications from idea to production in under six months.
The mechanism is not mysterious. When a bank can classify a new use case against an approved pattern, it skips a bespoke review and reserves deep scrutiny for novel, higher-risk uses. Standardizing what repeats is how governance becomes a throughput tool. Matteo Coppola, global leader of BCG’s Risk and Compliance practice, put it plainly: the banks applying governance well in day-to-day operations are the ones scaling AI most extensively.
The HFS data points the same way from a different angle. Among enterprises with AI embedded in standard operations, 56% have met or exceeded their business case, compared with 28% of those deploying AI across several functions. HFS is careful to say these are different groups and the data does not prove causation. Still, the pattern matches what we saw in JP Morgan’s agentic AI rollout: scale follows operating discipline, not the other way around.
The measurement angle: incidents and lost value are the same problem
BCG makes a point that deserves more attention. AI risk, it says, can include lost value in addition to major risk events. The 46% of banks with an AI incident and the more than one in four whose AI is falling behind its business case are looking at two views of the same blind spot. If you cannot see what an agent did, you cannot prove it was safe, and you cannot prove it was worth the money.
That is the gap Olakai was built to close. Olakai gives enterprises one record of AI activity across models, agents and AI tools, so the same data answers the risk officer’s question (what did this agent do, under which policy, and who owns it) and the CFO’s question (what did it cost and what did it return). For banks already spending hundreds of millions a year on AI, and facing questions like the ones raised by Barclays’ job cuts, having those two answers come from one dataset is the difference between a governance framework and a governance capability.
Visibility also has to include the AI nobody registered. Agents and assistants adopted outside formal channels are where incidents tend to start, which is why shadow AI detection belongs in the same program as agent monitoring rather than in a separate security project.
What bank leaders should do now
BCG’s own framing suggests the window is open but closing: agents are a small share of portfolios today, and that will not last. Four moves turn the benchmark’s findings into a plan.
- Build a tool-based inventory of every AI system and agent, including those running inside vendor platforms and business-owned tools, with a named owner for each.
- Define standardized, preapproved patterns for low-risk agent use cases, so they move to production quickly and review capacity goes to novel, high-risk uses.
- Monitor agent behavior continuously in production, including actions taken, data accessed, policy exceptions and escalations, rather than relying on periodic review alone.
- Track value alongside risk for every agent: cost per completed task, outcome against the business case, and the human oversight it still requires.
Each of these depends on the same foundation: a continuous, attributable record of what AI is doing across the institution. AI governance built on that record can show a regulator the controls in action, and analytics and custom KPIs built on the same record can show the board which agents earn their place. For institutions in financial services, where supervisors are already asking how agents are overseen, that record is quickly becoming an expectation rather than an option.
Where to start
Take the agents your bank already permits in production and answer three questions for each: who owns it, what it did last week, and what it returned against its business case. If those answers require a manual data call to several teams, you are in the 92% without comprehensive agent monitoring, and the restrictions in your risk-appetite framework are doing work your controls cannot yet do. The banks pulling ahead in BCG’s data are the ones that fixed that first, and Olakai can help you do the same.
Talk to an expert about monitoring every AI agent, model and tool in one record, and see how the Olakai platform connects AI governance to measurable value.