Why an AI agent's actions need the same audit discipline as a human technician's — no exceptions

Audit discipline exists to answer one question after the fact, cleanly: who decided this, and on what basis. That question doesn't get easier to answer when an AI agent is involved — it gets harder, because there's now an extra step in the chain, and the temptation in a lot of AI tooling is to compress that step into something vague. 'The AI resolved it' sounds like an audit-friendly sentence. It's actually the opposite: it collapses two distinct events, a machine's proposal and a human's decision, into one entry that can't answer which one actually happened.

Why a combined log entry fails under real scrutiny

• It can't distinguish a technician who carefully reviewed a draft from one who clicked approve without reading it — both look identical in a single blended record
• It can't show a regulator or an insurer that a human was actually in the decision path for a consequential action, only that something happened and an AI was involved somewhere
• It makes root-cause analysis after an incident nearly impossible, because you can't tell whether the failure was a bad draft, a bad approval, or a bad interaction between the two
• It quietly shifts accountability onto 'the AI' in a way that isn't actually true and isn't actually useful — software doesn't attend the postmortem

What separate events actually buy you

Two linked but distinct audit entries — one for the AI's draft, one for the human's approval — reconstruct the real sequence of decisions instead of a summary of the outcome. You can see what the model proposed, verbatim, before any human touched it. You can see who approved it, when, and whether they modified it first. That's the difference between an audit trail you can actually defend and one that just asserts a good outcome happened.

An audit trail's entire value is in resolving ambiguity after the fact. A record that says 'AI and human, combined' preserves the ambiguity instead of resolving it — which means it isn't an audit trail, it's a summary.

This is the same standard we'd apply to a human technician's actions, and there's no principled reason to relax it just because the actor drafting the action is a model instead of a person — if anything the newer, less-understood actor deserves more scrutiny, not less. Every AI draft in Nexus and every human approval on it are logged as separate, linked events for exactly this reason, and it's a mechanical property of how the audit log is structured rather than a policy statement layered on top — which means it's also something you can go verify directly on a real ticket or job record rather than take on our word for it.