This article is by Adam Riccoboni, see bio below.
The number arrives before the question
Picture the month-end close at a mid-sized insurer. An AI agent has pulled the claims data, reconciled it against the ledger, drafted the commentary on movements in the reserves and laid out the pages of the board pack. It is quick and tidy, and it is wrong in one place. Nobody notices, because the agent has been right for six months running.
That is the scenario that worries me more than any dramatic AI failure. A wildly wrong number gets caught. A plausible wrong number, in a document that has earned everyone’s trust, goes straight to the board.
Insurance Edge made the point well on 13 October: accountability for agentic AI cannot be outsourced to a vendor or an algorithm, and governance belongs with the board rather than IT. I agree. The practical question that follows is less comfortable. When an AI agent has touched a reserving report, the close or the board pack, who signs off the number, and what do they need in front of them to do it honestly?
Why insurance finance is a special case
Most of the AI governance conversation is about customer-facing use: pricing, claims triage, chatbots. Finance gets less attention because it feels internal. That is a mistake. The numbers finance produces flow into reserving, solvency reporting, regulatory returns and the board’s view of the business. They are the numbers other people rely on without checking.
Insurers also already run on judgement layered over models. Reserving has always meant an actuary taking a model’s output, applying experience and signing a view. The risk with AI agents is not that judgement disappears. It is that the agent quietly moves into the space between the systems, the gathering, matching, adjusting and drafting, and the person signing off no longer sees what happened there. The Bank of England and FCA survey of AI in UK financial services shows how far adoption has already spread across banks and insurers, and how much of it relies on third-party models. Use is running ahead of the controls built for it. Finance teams are where that gap will show first.
Five things to have in place before the agent touches a reported number At Critical Future we have spent over a decade building custom AI, much of it for the finance function. The projects that hold up in production have a few things in common, and none of them is a cleverer model.
1. A named owner for every AI-touched output. Not the vendor, not “the AI team”. A person in finance whose name sits against the reserving commentary or the close pack, who understands what the agent did and is accountable for the result. If nobody will put their name to it, the agent should not be producing it.
2. A record of what the agent actually did. Which data it read, which version of which source, what it changed, what it flagged and what it ignored. If the reviewer cannot reconstruct how a figure was produced, they are not reviewing it. They are trusting it. An audit trail is not a nice extra for the auditors. It is the thing that makes sign-off mean something.
3. Defined limits on what the agent may decide. Drafting commentary is one level of autonomy. Posting a journal is another. Changing an assumption in a reserving model is another again. Write down which actions the agent may take alone, which need a person to approve, and which it may never take. Then test that the limits hold.
4. Validation that matches the stakes. Before go-live, run the agent alongside the existing process for several cycles and compare the results line by line. After go-live, keep sampling. The PRA’s model risk management principles were written for banks, but their core ideas travel well to insurance finance: know your models, validate them independently and keep testing them in use. An AI agent in the close is a model, whatever the vendor calls it.
5. A way to stop it. Every agent that touches reported numbers needs a simple, rehearsed way to switch it off and fall back to the manual process. If the close cannot run without the agent, you have created a dependency nobody approved.
What the person signing off should ask
The CFO, the chief actuary or the audit committee chair does not need to understand how the model works. They do need to ask a handful of plain questions, every time:
Which figures in this pack did an AI agent produce or change? Who reviewed them, and what did that review consist of? What did the agent flag as uncertain, and what happened to those flags? When was the agent last tested against the manual process, and what did that test find? If we turned it off tomorrow, could we still close?
If the answers are vague, the number is not ready to sign, however good it looks.
Speed is not the prize
The pitch for AI in finance is usually speed: a faster close, instant commentary, fewer late nights. Those gains are real. But in insurance the prize is a number you can stand behind in front of the board, the auditors and the regulator. Speed without that is just getting to the wrong answer sooner.
The firms that get this right will not be the ones with the most agents. They will be the ones where every AI-touched number has a name against it, a record behind it and a person who could explain it. That is what board-level accountability for AI looks like when it reaches the finance team. It is unglamorous, and it works.
About the author
Adam Riccoboni is Founder and CEO of Critical Future, an AI strategy and data engineering partner with over a decade of custom AI development and a specialism in finance function AI. He is the author of “The AI Age” (2019), lectures in artificial intelligence at ESCP Business School and convenes the CFO Council on AI.
Be the first to comment