Only if you captured the right record when it acted. You cannot rebuild the reason afterwards by rerunning the model, because the same prompt can produce a different answer, and the agent's own explanation of its reasoning is not reliable evidence. What you can reconstruct is what the agent saw, what it proposed, which bounds checked it and who approved.
Why doesn't rerunning the agent work?
Models are not deterministic in practice. Run the same request twice and you may get two different drafts. The sources the agent read may also have changed since: a customer record updated, a policy revised, a price moved. A rerun tells you what the agent would do now, not what it did then.
The model's own narrative of its reasoning is no better. It is generated text, written after or alongside the action, and it can sound convincing while omitting the real cause. Treat it as a useful note, not as the explanation.
What must be recorded at the moment of action?
The deciding factor is whether the decision context was captured when the action was checked, not later.
| Element | What to capture | Why it matters |
|---|---|---|
| Request | Who or what asked, and the exact wording | Starts the chain |
| Inputs | Every fact the agent used, with provenance | Shows what it believed |
| Source versions | Which policy, price list or record version | Sources change after the event |
| Brief and model | Version of the agent's brief and the model | Behaviour depends on both |
| Tool calls | Each call, its parameters and result | Shows what it did step by step |
| Proposed action | Exactly what it tried to do | Separates intent from outcome |
| Bounds checked | Each bound and its result | Shows why it was allowed |
| Verdict | Allowed, held, refused or withheld, and the bound that decided | The decision itself |
| Approver | Who approved a held action and what they saw | Human accountability |
Our guide to reconstructing an AI agent's decision from the record walks through using this record in practice.
Why does the verdict matter more than the reasoning?
Because the verdict is the part you control. You cannot fully explain why a model drafted what it drafted. You can explain exactly why the action was allowed: the agent's role scope covered it, the amount was under the customer's limit, the sources were sealed, the certificate was current. That is the logic of governing the action, not the model, and it depends on checks happening when the agent acts. Our answer on how to enforce an AI agent's limits at the moment of action covers the mechanics.
A worked example
Take a bank's operations agent that unblocked a customer's card three weeks ago. The card was then used fraudulently, and the customer complains that the bank should have kept it blocked.
The investigator opens the record. The request came from the customer through the app, authenticated. The agent read the block reason, "suspected card loss", and a later customer message saying the card was found. That message had declared provenance. The bank's limits allowed the agent to unblock cards blocked for loss when the customer confirmed recovery through an authenticated channel. The verdict was allowed, decided by the customer limit, and no approver was needed.
The reconstruction shows the agent followed the rule. The weakness was the rule: a declared statement was enough. The bank changes the limit so unblocking after suspected loss is held for a person, and the record of the change sits beside the original verdict.
What may the agent do in an investigation, and what stays with a person?
An agent can assemble the record, line up events in order and draft a summary. The investigator decides what caused the outcome and whether a rule, a source or the agent was at fault. If something must be undone, a person decides on the remedy; what a compensating action is explains how harm is corrected when an action cannot simply be reversed.
Which controls make reconstruction possible?
- Record at the action, not after. The platform files the context as part of the check.
- Append-only storage. Entries cannot be edited or removed.
- Versions, not just values. Keep the version of every policy and source used.
- Retention long enough for complaints, disputes and audits in your sector.
- Practice. Reconstruct a few decisions each quarter. Gaps show up quickly, and a response plan for actions outside the agent's bounds depends on the same record.
The topic guide, agent governance is a bound, not a model, explains why the record is part of the bound.
Questions people ask
Should we store the model's reasoning text?
Yes, as a supporting note. It can help a reviewer understand a draft. Do not treat it as the explanation of the decision.
How long should the record be kept?
As long as the decision could be challenged: complaint windows, limitation periods and audit cycles in your sector. Personal data in it still needs to meet data protection rules.
Can customers see the record?
They are entitled to an explanation and, for personal data, access to their data. They do not usually see the full internal record, which may contain other people's information.