AI SecurityZero TrustSpeaking

When the Agent Acts, Who Answers? My Take From Reuters Momentum AI Finance

November 19, 2026 · Brian Deitch

On Tuesday I sat on an executive panel at Reuters Events Momentum AI Finance in New York. The title was the question every bank is quietly asking right now: "When the Agent Acts, Who Answers?"

I shared the stage with Allyn Shaw, Chief Business Execution and Risk Officer at Truist, and Moataz Mahmoud, SVP of Enterprise Data Management at First-Citizens Bank, with Elena Guz moderating. Great room, sharp people, and a lot of risk officers who were clearly not there for the free coffee.

Quick disclosure: I work at Zscaler. Everything below is my own view, not a recap of what anyone else on the panel said.

The Problem in One Sentence

Financial services spent decades building risk frameworks around a clean split: models make predictions, people make decisions, vendors provide technology. Agentic AI blurs all three. An agent can read data, decide, and act, using a vendor's model, inside your systems, on behalf of an employee who may not even know it ran.

So when it wires the money, approves the exception, or emails the wrong customer, whose name goes on the incident report?

My Position: Accountability Doesn't Move. Controls Do.

I'll take the firm stance here. The agent is never accountable. A named human owner always is. "The AI did it" is not an answer any regulator, board, or customer is going to accept, and it shouldn't be.

What does have to change is how that owner can realistically stay in control. You can't ask a human to approve 40,000 agent actions a day. That's not oversight, that's a rubber stamp with extra steps. So the controls have to move into the architecture.

Treat Agents Like Identities, Not Features

This is where my Zero Trust brain kicks in. Every agent should be treated like a new kind of user:

  • Its own identity. Not a shared service account, not borrowed human credentials. If you can't tell which agent did something, you can't tell who owns it.
  • Least privilege, scoped per task. An agent that summarizes loan files does not need write access to the core banking system. Grant the narrowest access for the job, and let it expire.
  • Policy at every action, not just at login. Checking an agent once when it starts is like checking a contractor's badge in the lobby and then letting them roam the vault. Evaluate each call it makes against what it's allowed to do.
  • Hard limits on irreversible actions. Money movement, account changes, customer communications: put thresholds in place where a human approves, and keep everything below that line logged and reviewable.
  • A kill switch you've actually tested. Know how to pull an agent's access in minutes, not after a change board meets next Thursday.

Logs Are the Accountability Layer

If accountability sits with a human, that human needs evidence. Every agent action should leave a trail: which agent, acting for whom, using what data, calling which tool, with what result. Without that, your "human in the loop" is a human in the dark.

This also makes third-party risk tractable. When a vendor's model sits inside an agent's workflow, you can't audit the model's reasoning, but you can absolutely audit what it was allowed to touch and what it did.

Guardrails Are How You Say Yes

The part I liked most about this conversation: nobody was arguing for blocking AI. Banks that try to ban agents just get shadow agents. The point of guardrails isn't to slow AI down. It's what lets you give it more responsibility safely, one well-scoped, well-logged task at a time.

Your Action Item

Tomorrow morning, ask your team three questions:

  1. How many agents do we have?
  2. What access to systems and data do they have?
  3. Last but not least, if we needed to pull the plug and stop them, can we?

If you can't answer those questions confidently and definitively, we should have a one hour chat really soon.

Bottom Line

Agents can act. They can't answer. Put a named owner on every one, give that owner real controls at the architecture level, and the question in the panel title gets a boring, clear answer. Boring is exactly what a risk officer wants.

Thanks to Reuters Events for having me, and to Allyn, Moataz and Elena for a great conversation.

Event: Reuters Events Momentum AI Finance agenda

agentic-aiai-governancefinancial-servicesaccountabilityreuters-eventsmomentum-ai-financezero-trustnon-human-identity
All postsDiscuss on LinkedIn