Answers
How do we show a client that our AI followed their rules instead of inventing one?
By showing the record the assistant read, rather than a summary written afterwards. Every rule in force in a project carries what was decided, when, and which person approved it — and a rule an assistant proposed that nobody approved is visibly still a proposal. So the answer to a client asking where a rule came from is a dated line with a name on it. What the record does not do is prove that one particular deliverable came out of it: it shows what governed the work, and your own review is what ties a specific piece to it. Claiming more than that would be claiming more than the record holds.
Last updated September 4, 2026
The question is provenance, not logs
A client asking for proof is not asking for a transcript. They want to know that the constraint they gave you is in force, that nobody changed it quietly, and that what your assistant presented as firm policy is not something it inferred on the spot. Those are all questions about where a rule came from. A conversation log answers none of them — it is long, it is unreadable, and it contains every wrong turn taken on the way to the answer.
Approved by a person, and it says which one
An assistant working on an account can propose a rule, and that proposal governs nothing until someone on the project approves it. When they do, the approval is part of the record: the person, the date, and the source they pointed at. When they refuse, the refusal is recorded too, which is what stops the same suggestion returning as though it had never been seen. A client can therefore be shown the difference between what your team decided and what a model suggested, in one place.
What it proves, and what it does not
The record proves what was in force at a point in time, who put it there, and that it had not been superseded. It does not prove that a given draft obeyed it, and no memory can: obedience is a property of the output, and checking outputs is what review is for. Being explicit about that line is worth more with a client than a broader claim would be — a promise that compliance happens on its own collapses the first time something slips through, and it takes the credible part down with it.
What it looks like in practice
A regulated client asks an agency to confirm in writing that a claim used across a campaign was cleared with their legal team rather than produced by the agency's assistant. The agency opens the account's rules in force. The claim's exact wording is there as an approved rule, dated four months earlier, approved by the account lead, with its source line pointing at the client's own written sign-off. Beside it, still marked as a proposal, is a broader version of the same claim that the assistant suggested in July and the account lead refused, with the refusal recorded. That second line is what settles the question. It shows the assistant did suggest going further, that a person said no, and that the narrower wording is the one in force. A log of the drafting session would have shown the suggestion and left the client to guess what happened next.
Questions people ask about this
- Can we give a client direct access to see this?
- You can add someone to a project, and that is how a client reads the account's rules directly. Consider the scope first: they see the account they are on and nothing from any other client, but they also see the agency's own decisions recorded there. Many agencies prefer to export the relevant rules instead, which gives the client the same dated lines without a standing seat on the project.
- What if a rule changed halfway through the engagement?
- Then the record says so. A rule that is replaced is replaced by the one that supersedes it, and the earlier text stays in the history rather than disappearing — so it is possible to say what was in force in March and what has been in force since. For a client question about older work, that is usually the more useful answer.
- Does this work for an internal audit, not just a client?
- It is the same record read for a different reason. What it supports is a statement about what the assistants working on an account were operating under, and who approved it. It does not stand in for whatever else the audit requires — reviews, sign-offs, delivery evidence — and presenting it as more than one input is where this kind of record loses its value.
Where this is verifiable
Product documentation on this site (How it works, Install), the sanction, refusal and provenance rules in the sanctioned product spec, and the answer on client confidentiality for how project scope keeps one client's material out of another's. Everything described here is behaviour the tools apply today, not roadmap.
https://www.arroway.app/en/answers/showing-clients-the-ai-followed-their-rules