Skip to content
Spark
Writing

Governance and compliance

Building an AI audit trail: how regulated firms keep AI decisions defensible

An AI audit trail is five fields and a habit. Capture the input, the model and prompt version, the output, the approver, and the timestamp, from day one. Retrofit it later and you are reconstructing history you no longer have.

By Seb Lawson · 12 May 2026 · 4 min read

An AI audit trail is five fields and a habit. For every material decision a system helps make, you record the input it received, the model and prompt version that produced the output, the output itself, who approved it, and when. Capture those from the first day a system goes live and you can defend any decision it touched. Skip them, and you are reconstructing history you no longer have.

This is the single control regulated firms most often wave their hands at. Everyone agrees they should be able to explain an AI output. Far fewer have wired up the plumbing that makes it possible after the fact. The gap only shows when a client complains, a supervisor asks, or a decision goes wrong, which is exactly the moment you cannot afford to shrug.

The five things to capture

Keep the record boring and complete. If you can reconstruct what happened without asking a person to remember, you have enough.

  1. The input. What actually went into the model: the document, the client data, the query, the context retrieved. Not a paraphrase. The real thing, or a stable reference to it.
  2. The model and prompt version. Which model, which version, and which prompt. A prompt is code. When it changes, the behaviour changes, and you need to know which version produced which output. We wrote about why in prompt versioning.
  3. The output. What the system returned, verbatim, before any human edited it. The edited final version matters too, but the raw output is what proves what the model did.
  4. The approver. Who reviewed it, who signed it off, and whether they changed anything. A name, not a role. ‘Reviewed by compliance’ is not an approver.
  5. The timestamp. When each step happened. Input received, output generated, human approved. Order and timing are half the story when something is contested.

That is the whole schema. Five fields, logged automatically, on every decision that could reach a client, a payment, or a regulatory return. Low-stakes uses, like a first draft of an internal note, do not need the full treatment. Match the record to the risk, the same way the JFSC’s guidance expects effort to track harm.

Why retrofitting is painful

Built in from the start, an audit trail is nearly free. The system writes each field as it works, and nobody has to remember anything. Bolted on after a supervisor asks, it is expensive and sometimes impossible. You cannot log a prompt version you never recorded. You cannot recover the input if it was overwritten. You cannot prove who approved something when the system never asked for a name.

In our experience most firms that skip this end up rebuilding the workflow anyway, usually under time pressure, and often after the exact incident the trail would have explained. Two weeks of clean logging design at the start saves a scramble later. That is the trade, and it is not close.

You cannot log a prompt version you never recorded. Build the trail before you need it, because you will need it on the day you least expect.

What a supervisor actually wants to see

A regulator is not trying to catch you using AI. They are trying to establish whether you are in control of it. When they ask about a specific decision, they want a short, honest answer to three questions. What did the system do? Who checked it? And what would have happened if it got it wrong? A clean audit trail answers all three in minutes.

What unsettles a supervisor is not a mistake. Firms make mistakes. It is the inability to explain one. A firm that can pull the input, show the prompt version, point to the human who approved it, and demonstrate the checkpoint that would have caught an error, looks like a firm in control. A firm that says ‘the AI did it and we are not sure how’ does not.

Make it real on one workflow first

Do not try to instrument every system at once. Pick your highest-stakes AI workflow, the one that could actually hurt if it failed, and build the five-field trail properly there. Prove the pattern, then reuse it. In our work this is a fortnight of effort on a well-scoped case, and it is the part of the work that most reliably survives contact with a regulator.

The audit trail is only half the control. The other half is the human doing the approving, and whether that review is genuine or theatre. We covered that in human-in-the-loop AI. Get both right on one workflow and you have a template for every AI system you build after it.

If you want a candid read on whether your current AI decisions are defensible, book a review. We will look at one live workflow, tell you what is missing, and put a number on what it takes to fix. No deck, no hype. Just what to capture and why.

Let’s map how your business works

A free 30-minute call with Seb or Mark. Bring the spreadsheet you hate most and we’ll tell you honestly whether a system would fix it.