Review the audit trail
Settings → Audit trail answers the question an access review actually asks: who opened this? It records reads as well as changes, it cannot be edited once written, and it's the surface you bring to anyone auditing who reached what.
Two different things carry similar names. Keep them apart:
- Settings → Audit trail (this page) — one row per access or change to a record: who opened a conversation, who deleted a user. It's about people reaching data.
- Governance → Audit log (Monitor usage and cost) — one row per inference: which model, how many tokens, what it cost, how fast. It's about the AI doing work.
If the question is "who looked at this?", you want this page. If it's "what did the model do and what did it cost?", you want Governance.
Prerequisites
Before you start, make sure you have:
- Administrative access: the trail is admin-only, and deliberately so — it records who reached what, which is itself sensitive.
Read the trail
Go to Settings → Audit trail. Each entry carries five things:
- When the action happened.
- Who did it. An entry with no actor reads System — background work, not missing data.
- Action — what was done. Read is a first-class action here, which is the whole point: it's the one nothing else in the platform records.
- Resource — what was acted on.
- Detail — a one-line summary of the entry.
The table is paged and sortable, newest first by default, because the question that brings anyone here is almost always what just happened. Filter by action to narrow it; changing the filter resets you to the first page rather than stranding you on page four of two.
Reads are the bulk of the trail, so colouring them would turn the column into decoration. The colour is reserved for the destructive action, which keeps it meaning something when you scan the page.
Why entries can be trusted
Entries are append-only, enforced by the database itself: an attempt to update a row is rejected outright, not merely disallowed by the application. That distinction is the point — a trail the application can rewrite is a log, not evidence, and "we only ever insert" is a claim a reviewer would otherwise have to take on faith.
Deletion is still permitted, because entries have to age out eventually and a retention sweep is the legitimate way to do that.
So the guarantee is precise, and worth stating precisely when someone asks: falsifying an entry is impossible; destroying one is not. If your obligations require entries to survive for a fixed term, that's a matter of the retention policy you set, not of this protection.
Two further details matter when you're reading old entries:
- The actor is recorded verbatim, not looked up when you view the page. An entry still names the person correctly after that user has been deleted — which is exactly the moment someone goes looking.
- A refused access is not recorded as a read. Entries are written after the access check passes, so the trail holds reads that actually happened rather than reads that were attempted.
What is recorded today
Instrumentation is deliberately partial rather than a blanket sweep of every endpoint. Recorded now:
- Reading a conversation — the access review case, and the one no other part of the platform captured.
- Deleting a user — written in the same transaction as the deletion, so there's no window in which one happened without the other.
- Erasing a person's data — written before anything is destroyed, since afterwards there's nothing left to tie an identifier back to a person. See Erase a person's data.
Expect this list to grow. An action you don't see here isn't being hidden — it isn't instrumented yet.
What the list deliberately doesn't show
Entries can carry a before and after payload, and the list view doesn't display them. Those payloads can hold the very content the entry exists to protect, and a table anyone can scroll is the wrong place to spread it.
Related
- Set data retention — the policy that governs how long entries live.
- Erase a person's data — an erasure records itself here before it runs.
- Monitor usage and cost — the other, differently-named log: one row per model call.
Result
You can answer who opened this record, and when from a page built for the question, and you can say precisely what the trail guarantees: entries cannot be altered once written, though they can age out under your retention policy. You know that reads are captured as first-class events, that a refused access isn't recorded as one, and that entries keep naming the right person even after that account is gone.