How did the agent decide? The question your regulator will ask

The question behind the question

Every AI conversation in banking eventually arrives at the same place: not "is it accurate?" but "can you show your working?". A model that is right 99 times and inexplicable once is, to a supervisor, inexplicable. The bar isn't intelligence — it's accountability.

Agents that can answer for themselves

The difference between an agent you can deploy and one you can defend is built before the agent runs:

• Decisions live in rules the agent uses, not weights it hides. Credit criteria, eligibility and escalation thresholds sit in decision tables your risk team can read, test and version — the agent operates them; it doesn't invent them.
• Every action writes its evidence. What the agent saw, which rule version fired, what it did, who reviewed it — captured at the moment of action, immutable after it.
• Guardrails are platform controls. What data an agent can touch and what actions it can take are permissions, enforced and logged like any user's — because the agent is a user.
• A named human answers. For every workflow an agent touches, one accountable owner — the same principle SMCR already taught the industry.

Why the foundation decides the outcome

None of this can be added after deployment. It's a property of the data foundation and the workflow design — which is why every implementation we build is structured so that when you switch the agent on, the audit trail is already there, waiting for it.

Where to start

Choose the first workflow you'd want an agent in, and write the regulator's question for it: "show me how this decision was made." If today's systems can't answer it for your humans, they won't answer it for your AI — and that, not the model, is the first project.