What must an audit trail answer?
An audit trail for AI decisions must answer three questions for every action: who decided, from what evidence, and when. If any of the three is missing, the trail cannot support an audit, an incident review, or a regulator's question, and the organization is left reconstructing events from memory.
This is the standard applied to human decisions in finance and compliance, and there is no reason to apply a lower one to decisions an AI employee proposed.
What are the seven things it must record?
The trail must record the proposal itself, the evidence it carried, the decision made on it, the reason if it was sent back, the time of each step, the autonomy level that applied to that class of action at the time, and the effect: what was published, sent, posted, or spent as a result.
Together these let a reviewer reconstruct any action fully. A published page traces back to its draft, its fact-gate result, its approver, the date, and whether it was gated or auto-approved at the time.
- Proposal: what the Digital FTE proposed to do
- Evidence: sources, matches, scores, or fact-gate results attached
- Decision and reason: approved, edited, or sent back, and why
- Timing: when proposed, when decided, when effected
- Autonomy level: gated or auto-approved for that class at the time
- Effect: what was published, sent, posted, or spent
- Changes: credentials, budgets, autonomy grants, settings
Why is it built as work happens?
The trail on ftes.ai is built as a by-product of the approval gate: every proposal passes through it, every decision is recorded there, and every effect is logged when it happens. There is no separate step where someone writes up what occurred, which is where trails usually go wrong.
Token spend is recorded in an immutable ledger, so the cost of each piece of work is part of the trail rather than a separate report. Per-run trails show what a Digital FTE did within a single piece of work, step by step.
What does a reviewer see?
A reviewer sees, for any action, the full chain: the proposal, its evidence, the approver's name and decision, the timing, the autonomy level, and the effect. For a reconciled month, that is every match with its documents and every approval with its name. For a shortlist, it is every score with its evidence and the human decision that followed.
For governance changes, the reviewer sees who granted autonomy for which class on which date, who changed a budget, and who connected or disconnected an integration. The trail covers the management of the FTE as well as its work.
How should you use it?
Use the trail for three purposes: incident review, when something went wrong and the question is what happened; autonomy decisions, where the approval history for a class is the evidence for promotion; and periodic audit, where a sample of actions is traced end to end to confirm the controls are working.
Ask any vendor what their trail contains and whether it can be exported; the vendor security questions guide on ftes.ai includes it. A trail that cannot be exported is a trail you do not own.
What does a single traced action look like?
As a labelled example, tracing one published page from a Search Visibility FTE: the trail shows the topic proposal with its worth-it verdict and the search data it drew on, dated; the draft with its sources attached; the fact-gate result listing each claim checked and one claim blocked in the first draft with the revision that resolved it; the approval by a named marketing lead on a stated date with one edit to the title; the publish event through the CMS connection with the confirmation returned; the measurement samples for the page's target prompts on the following weeks' schedule; and the tokens consumed at each stage from the ledger.
Tracing one auto-approved reconciliation match from a Finance Controller FTE shows the proposal with the bank line and invoice attached, the note that it was auto-approved under the exact-match class at Level 2, the grant of that level by a named controller on a stated date on a record of a stated number of unchanged approvals, the effect in the ledger, and the tokens consumed. Either trace takes a minute to read and answers who, from what evidence, and when, which is what the trail is for.