The Three Words That End an AI-Agent Audit Badly: "We're Not Sure"
September 6, 2026
The question you will eventually be asked
If your organization uses AI agents for anything that touches real business processes, this question is coming, whether from an internal audit, a regulator, or a customer's due-diligence team: "Show me who approved this specific action, and on what basis."
There's a good answer to that question. There's also a genuinely bad-but-survivable answer ("this was a mistake, here's exactly what we changed as a result"). And then there's the answer that actually damages trust, more than either of the other two: "we're not entirely sure — we have logs, but not a clear approval trail." That third answer isn't a technicality or a rounding error. It's the difference between a controlled process and an ungoverned one, and auditors are trained specifically to tell which is which — the uncertainty itself is the finding, independent of whether anything actually went wrong.
Why the pressure to move fast made this question harder, not easier, to answer
Organizations aren't slow-walking AI adoption out of caution for its own sake. They're moving because the case for speed is real: agents that draft, triage, and act faster than a team of people could is a genuine advantage, and every quarter of delay has a real cost. That's not the part that needs fixing. What needs fixing is the assumption, common in fast-moving teams, that documenting the approval trail is something you can bolt on later once the pilot proves out. It rarely gets bolted on later. It gets discovered missing, at exactly the moment someone asks for it.
Article 12 isn't abstract anymore
The EU AI Act's Article 12 requires automatic logging of events over the lifetime of high-risk AI systems — not as a nice-to-have, but as an enforceable obligation, with enforcement dates that are no longer comfortably far away. We're not citing this as a claim about what any particular product does or doesn't certify you for; no vendor can hand you compliance in a box. We're citing it as the real external context that's changing what "we'll figure out the audit story later" actually costs. For any EU organization deploying agents into real workflows, that sentence is quickly becoming "we'll figure out the audit story after the deadline has already passed" — and the fix, at that point, is retroactive and much more expensive than it needed to be.
This isn't really new territory for compliance teams — it's the same discipline already applied to financial controls or access management, applied to a new kind of actor that happens to be software instead of a person. The bar for evidence doesn't change just because the thing making the request is an agent instead of an employee.
Why "we have logs" isn't the same as "we have governance"
Logs answer "what happened." They don't answer "was this authorized" or "could it have been stopped." A genuinely defensible answer needs three things stacked together: a policy that was actually evaluated before the action ran, a human decision tied to the specific proposed change (not a general sign-off on "the AI program"), and a record that can't be quietly edited after the fact. Most organizations have the first piece — logging — and are missing the second and third: the actual preventive control, and the tamper-evident proof that a person, specifically, exercised it.
Note that these are two separate things, and conflating them is exactly how organizations end up with the false confidence that produces "we're not sure." An audit trail tells a story about the past. A preventive control changes what could happen in the future. You need both, and having one is not evidence of having the other.
What "defensible" actually looks like
When the question comes — and for a regulated or EU-based organization deploying agents into real workflows, it will — the answer you want to give is specific and fast, because it's already true, not because you scrambled to reconstruct it: this exact action was evaluated against this policy, this named person reviewed and approved the exact change being proposed, and here is the record, unaltered, tying all three together. That's a fundamentally different, far stronger position than "here's a log file, we believe nothing unusual happened" — and it's also, not incidentally, the position that lets you answer the auditor in minutes instead of weeks, because there's nothing to reconstruct.
If your current answer to "who approved that agent action" leans more on hope than on evidence, that's precisely the gap David Gate is built to close, before the auditor asks, not after. Walk us through one workflow you'd need to defend under real scrutiny, and let's see what a genuinely provable answer looks like.
