How to Audit What Your AI Assistant Actually Did

Most AI features in business software cannot tell you what they did.

They can tell you what they can do, and they can show you the output. But ask for a record of every action taken on your behalf last month — what was proposed, what a human approved, what data it read, what it sent — and you will usually get a description of the logging rather than the log.

That gap matters more as these tools move from summarising to acting.

Propose versus act

The first question about any AI feature is whether it proposes or acts.

A proposing system drafts, ranks or suggests, and a human decides. The failure mode is a bad suggestion, and the cost is the reviewer's attention.

An acting system sends the email, updates the record, assigns the task. The failure mode is a bad action that already happened, and the cost is whatever the action cost.

These need different scrutiny, and vendors describe both as "AI-powered." Ask which one you are buying, feature by feature — the answer often differs across features in the same product.

What a real audit trail contains

For every action, six fields. Fewer than six and you cannot reconstruct what happened:

The last one is the one nearly everyone omits, and it is the one that makes an investigation possible six months later. "The AI started behaving differently in March" is unanswerable without it.

The trail must be exportable. A log you can only view in the vendor's interface is not an audit trail — it is a screen. If you cannot produce a file, you cannot hand it to anybody who asks.

Twelve questions to ask a vendor

Ask for the artefact, not the description. "Can you show me an export of everything your AI did in this demo account last month?" separates the categories immediately.

On capability

1. Which features propose, and which act autonomously?
2. Can autonomous action be turned off per feature, or is it all-or-nothing?
3. What is the largest-blast-radius action the AI can take without a human?

On the record 4. Is every AI action logged, or only some categories? 5. Can I export the log — as a file, with a stable schema? 6. Does the log include the original output and any human edit, or only the final state? 7. Does it record which model and prompt version produced each output? 8. Is the log append-only, or can it be edited or deleted?

On grounding 9. When the AI makes a claim, can I see what it read to reach it? 10. What happens when it doesn't know — does it say so, or does it produce something plausible?

On retention and access 11. How long are logs retained, and can I set that? 12. Who at the vendor can read our AI logs, and is that access itself logged?

Question 10 is the one that predicts most day-to-day pain. A system that fabricates confidently when under-informed will cost you more in verification than it saves in drafting.

Why this is becoming a procurement question

Two things are converging.

AI is moving into decisions with consequences for people — how work is allocated, how performance is assessed, who gets shortlisted. Regulators in several jurisdictions have begun treating that category differently from AI that summarises documents, and obligations in this area have been arriving steadily.

Obligations often land on the deployer, not just the builder. In several emerging frameworks the organisation using the system carries duties of its own — typically to retain the system's automatically generated logs and to assign named human oversight.

Verify your own jurisdiction. Requirements differ by country, state and sector, and they are moving. This page is operational guidance on what to ask vendors, not legal advice, and nothing here should be treated as a statement of what any specific law requires of you. Confirm your obligations with counsel.

The practical consequence is straightforward regardless of which rules apply to you: if a tool cannot produce a record, it cannot help you discharge a duty to keep one — and you will have inherited an obligation with no means of meeting it.

What good looks like

A vendor with a real answer will, without much prompting:

• Distinguish proposing features from acting ones, unasked
• Show you an actual export rather than a screenshot of a log view
• Include the pre-edit output, not just the final state
• Version prompts and models, and be able to tell you when they last changed
• Have a clear answer on retention that you can configure
• Tell you plainly what the AI does when it lacks grounding

A vendor without one will describe the logging in general terms, offer a screenshot, and move the conversation to capability. That is itself an answer.

Doing this internally

Even without a vendor, the same discipline applies to anything you build:

1. Log the proposal separately from the action. The edit between them is the most informative signal you have about whether the AI is helping.
2. Never let a model act without a named approver on anything outward-facing or destructive.
3. Version your prompts and record which version produced which output.
4. Test the "I don't know" path deliberately. A system that fabricates under uncertainty fails silently, which is the worst way to fail.

Related

• Evaluating strategy execution software — where questions 4–8 belong in a wider evaluation
• Measuring strategic alignment — if scoring is automated, this is the accountability question that follows it