Audit trails: what your security team will ask for
When a security or compliance team reviews an agent, the questions are remarkably consistent across industries. There are about eight of them, and they are all versions of the same worry: can you show me what this thing did, and can you undo it.
Here is the full list, and what Lyrion logs so you can answer each one without taking a screenshot.
Who, what, when
Every action records the agent that took it, the exact action, the systems it touched, and the timestamp. “The AI did something” is not an answer a reviewer accepts. “The support agent scoped to billing issued a refund of this amount on this account at this time” is.
This sounds obvious and it is the thing most systems cannot actually produce. If the answer to “what did it do” is a model log full of tokens, you do not have an audit trail. You have a transcript.
What changed, and can it be reversed
For any action that modified a record, the log holds the before and the after, and whether the change can be undone. A reviewer's real fear is not that the agent acts. It is that it acts wrongly and the change is silent and permanent.
Reversibility is a design choice, not a feature you add later. Actions are built so that the ones that change state can be traced back and, where it matters, rolled back. That is what turns a nervous review into a short one.
Who approved it, and under what rule
Every action either ran automatically under a named rule you can point to, or a named person approved it. There is no third category. A reviewer can ask about any single action and get a clean answer: this rule, or this person.
This is why we made approvals central. The compliance story is not a report generated after the fact. It is the by-product of how the system already works, which means it cannot drift out of sync with reality.
Scope, retention, and residency
The last questions are about boundaries. What was each agent allowed to see. How long is the record kept. Where does the data sit and where does the processing happen. These are answered by configuration you set, not by assurances.
Your data stays in your own systems. Lyrion keeps no copy it does not need to. When the review asks where things live, the answer is: with you, and here is exactly what we touch and for how long.
What to keep an eye on
Watch the ratio of exceptions to handled work over the first fortnight. If it is falling, the agent is learning the shape of your queue. If it stays high, that usually means the scope is too broad or the material it reads from is thin, and both are fixable without touching the model.