What two rooms of bank leaders taught me about trusting AI in KYC

In November I hosted two roundtables on AI in know your customer (KYC), one in London and one in New York. Around each table were senior KYC, operations, and transformation leaders from a range of banks, talking openly under the Chatham House Rule.

I opened both sessions with the same question: where are you really on your AI journey, and is capability outpacing trust? It’s the question our Sibos panel, Trust at agentic speed, raised this year. Our answer then was yes.

After these roundtables, I’m more convinced than ever. The technology works. Building trust is the next frontier, and that depends as much on policy and people as on the AI itself. Here is what I took away.

Everyone is on a different part of the curve

The first thing that struck me was how far apart banks are. Some are still largely manual. Some use single agents or Model Context Protocol (MCP) servers to support analysts. A few already let agents prepare cases end to end.

At the leading edge, one bank told us agent-prepared cases are being completed in minutes at over 95% accuracy, on a small share of volume. Another has agents acting as the maker for certain low-risk populations, such as listed entities in specific jurisdictions, with human checkers still in place. For context, an analyst typically spends three to five hours on a comparable case.

Most firms are earlier on, though, and some deliberately prefer rules-based workflows with full logging. I think that is healthy. Deterministic automation, generative AI, and agentic AI are different things, with different risks and different controls. Being deliberate about which one you use, and where, matters more than racing to “agentic.”

Trust is earned through data and evidence

In both rooms, trust came back to two things: the data underneath the agent, and the agent’s ability to show its working.

One bank found its people were challenging agents more, not less. Another saw validators marking correct output as wrong, simply because the agent got there by a different source or route than an analyst would have. That is not a technology problem. It is a confidence problem.

The fixes people described were practical. Capture every step an agent takes to reach a recommendation. Use reason codes wherever the agent and the human disagree. Study every override, find the root cause, and feed it back into both the agent and the policy. The test I kept hearing was simple: can a non-technical reviewer follow the evidence?

And data comes first. Give an agent work instructions that differ between language versions of the same policy, and it gets confused in exactly the way a person would.

Policy must become rules, and someone must own them

This was the debate I found most interesting. KYC policies are written for people to interpret. Agents need rules they can apply. In between sit the “unwritten rules” that experienced analysts carry around in their heads.

Some argued we rely too heavily on rules that aren’t always logical, and that this drives false positives. Others seek to codify analyst reasoning or rewrite the rulebook for agentic decision making altogether. Exceptions are still hard. Client refusals that compliance has accepted, for example, do not fit neatly into any rule, and codifying every permutation would take far too long.

Then there’s ownership. The New York group agreed that compliance, not KYC procedures teams, should own the rules and the rules engine. Yet ownership alone doesn’t settle the question of who holds compliance accountable. My own view is that AI can help here too, by helping to write the rules so a smaller team can run them. Governance is maturing in parallel, with agent autonomy reviewed through central risk governance and each MCP server going through model risk management.

Who is right when the human and the agent disagree?

We tend to measure AI against human performance. But is the human baseline the right target? Participants reported human accuracy of 70 to 80% even after four-eyes checks, while AI is held to a 100% standard. So, I put the question to the room: what are you willing to accept?

There is a harder question underneath. Years of human interpretation can quietly become the baseline, and that baseline can be wrong. When AI exposes it, the bank gets a chance to fix the problem proactively, but it also must work through the remediation and regulatory questions that come with it.

The risk profile is different, too. Retrain an analyst who makes a mistake and you have fixed one person. One agent error can mean rechecking every case it touched. That is why bounded autonomy, clear accuracy thresholds and strong governance must come before you reduce human involvement.

The encouraging news is that regulators are on board. London participants felt supervisors are supportive, and that approval is possible where a model is bounded, explainable and risk-based, with humans owning the key decisions. The industry should not let itself be outpaced by its regulator.

People are the real key

If the technology works, adoption, and behavior really determine whether it delivers in a meaningful way. This came up repeatedly.

Analysts who do not understand a new capability double-check everything, which cancels out the time saved. Some still work offline and key cases in afterwards, which flatters the metrics. And testing with operations teams now takes longer than building the solution in the first place.

Roles are shifting. When agents do the maker work, checkers become the front line, and validators end up reviewing both human and agent output. Traditional quality scoring of individuals stops making sense. One leader aiming for an 80/20 split between agent and human effort said it had not transformed the organization yet, because nobody had decided what people should do with the time freed up.

Talent is what worried me most. Banks are losing good people who fear AI, and a pure efficiency play risks losing exactly the expertise the new model relies on. Reskilling is expected to take three to five years.

So, what’s working? Networks of internal automation experts. Spotting and empowering AI power-users. And involving analysts in building the solution because trust follows understanding. As one participant put it, a lack of trust usually comes from a lack of knowledge.

It is not just about cost

Cost take-out is only part of the story. Both groups are already seeing better risk escalations and exception handling, and the New York group wants to go further: using AI to catch fake entities, doctored documents, and fraud.

That brings it back to corporate digital identity (CDI). The most basic question in corporate KYC is whether an entity really exists and is who it says it is. Agents can only answer that reliably if they work from verified, authoritative data with a clear audit trail. That is exactly where we focus at Encompass.

What it adds up to

Three things came through clearly from both rooms:

1. The technology works. Adoption is now about risk appetite and confidence, not capability.
2. Behavior change decides success. The roles that remain need retraining, new measures, and a clear purpose. Embracing the potential of AI at the coalface is key.
3. The prize is bigger than cost. The strongest cases combine efficiency with better risk outcomes.

If I were planning the next step, I would fix the data foundation, turn policy into rules with clear owners, build explainability in from day one and bring analysts into the design. And I would choose a pace of change the organization can absorb.

With EC360, we gather evidence from official public sources and from a bank’s own data providers, apply the bank’s own policy, and deliver the result into core banking applications. Agents and analysts then work from the same auditable view of every corporate client, and both can follow the evidence.

Thank you to everyone who joined us in London and New York. If you would like to compare notes on how your peers are approaching AI in KYC or test your own approach at an Encompass Agent Build Day, I would be glad to talk.