Show me the record of the last thing it did and why.
Say it near the end, when the laptop is still open and the run everybody watched is a few clicks behind. You are asking for an artefact rather than an explanation, in this room, in the next two minutes, and most of what you learn arrives before anybody finds the file. Watch how many people have to be telephoned first, and whether one of them works for you.
An entry, and what it is made of
Every agent writes something down. The differences that matter are what goes in an entry, whether an entry can later be changed, and who is permitted to open the file.
An entry that answers your question carries the action, the time, the reason, and the rule or the person that permitted it. A line recording that a message was sent tells you a message was sent. A line recording that it was sent because a named person approved a named draft under a named rule tells you who decided, which is what anybody asks six months later.
The reference design adds one property and one refusal, both on the Register, the standing page of every setting in ArkOne’s reference design for an executive agent. Entries are added and never edited, so a correction is a new entry and the old one stands. And the ledger never holds a meeting up: a write that fails is itself written down elsewhere, and the room carries on. The cost of that choice is stated rather than hidden, because a minute of lost ledger is a minute whose evidence is a note saying the evidence is missing.
One correction, read six months later
Suppose the renewal invented for this work, in which a discount was offered, found to be wrong within the hour, and corrected. Half a year on, somebody asks what happened. Two designs, both of which recorded the correction.
| What you can establish afterwards | Corrections are added | Corrections overwrite |
|---|---|---|
| What the client was offered on the day | Both figures, in order | The corrected figure only |
| That anything was ever wrong | Visible | Invisible |
| Who noticed, and when | Named, with the hour | Absent |
| How long the wrong figure stood | The gap between two entries | Unknowable |
| What the file proves about the fix | That it was made, and by whom | That the figure is correct today |
Exhibit 1. Illustrative. One corrected action, asked about six months later, under an added entry and an edited one.
The second column is not a cover-up, and it is worth being precise about that, because it is what an ordinary database does by default when a value changes. It holds the truth as of now, which is exactly what most software is for. It simply cannot hold a sequence, and a sequence is the whole of what an audit is.
Money sits in the question of how long the wrong figure stood, because that decides whether one client saw it or forty did. One column answers with subtraction. The other cannot answer at all, and both designs behaved impeccably on the day; what differs is only what survives to be read.
Watching who has to be telephoned
A good answer is somebody opening a page and reading it out, with the reason beside the action, in words you can follow without an engineer. It usually comes with an offer you should accept: pick a different action and they will show you that one too.
The reply that sends your request somewhere else is the common one. We have full logging, or the system captures everything. Both may be entirely true and neither is the artefact you asked for, and the tell is who they have to ask. A second ordinary answer opens a dashboard that only the vendor’s staff may sign in to, which is a real record in somebody else’s custody. Ask whether your own people can read it without a support ticket, whether you may export it, and what happens to it when the contract ends.
Then two follow-ups worth more than the page itself. Can an entry be edited or removed, and by whom? And can a failure to write one stop the agent working? The second sounds technical and is not: a system that halts whenever its ledger hiccups is a system your staff will quietly ask somebody to switch the ledger off.
You asked for one entry about one action. What you were really measuring is custody, and the meeting answered it before anybody opened a file. The question about who it was talking to is the other request on your list that asks to be shown rather than told.
component: answer-card
Asked plainly
What should an AI agent's record of an action contain?
The action itself, the time, the reason it was taken, the rule or the person that permitted it, and enough of the material it relied on to judge whether the reason was sound. A line saying a message was sent is a log. A line saying a message was sent because a named person approved a named draft under a named rule is evidence, and only the second answers a question asked six months afterwards.
Can an AI agent's audit trail be edited after the fact?
In some designs, yes, and it is worth asking directly. A record where a correction overwrites the original cannot show that anything was ever wrong, so it can neither exonerate anyone nor prove a change was made. A record where a correction is added as a new entry, leaving the old one standing, shows the mistake and the fix, which is what makes the sequence usable as evidence.
Who is able to read an AI agent's records, and where do they live?
That is the half of the question a demonstration rarely reaches. If the history sits in a screen only the vendor may open, then producing it depends on a support request and on the vendor still being your vendor. Ask whether your own staff can read it without asking anyone, whether you can export it, and how long it is kept after the contract ends.