Why we log every AI-agent draft and every human approval as separate audit events, not one combined action

Picture the audit log entry after something goes wrong: "AI resolved ticket #4021 - closed as duplicate." That single line answers what happened. It does not answer the question that actually matters afterward: did a human genuinely evaluate that decision before it went out, or did they click approve on autopilot because the queue was long and the draft looked plausible? A combined log entry cannot distinguish those two very different situations, and they call for very different responses.

That's the specific reason every AI-agent draft in Nexus and every human approval of it are logged as two separate, linked events rather than one. The draft event records what the agent proposed, when, and what it based the proposal on. The approval event records who reviewed it, when, whether they edited it before approving, and whether they approved, rejected, or escalated it. Neither event overwrites or absorbs the other.

What splitting the log actually lets you answer later

• Was this a good AI draft or a bad one? — visible independent of what a human did with it, which is the only way to actually measure and improve the drafting quality over time.
• Did the human materially review it, or rubber-stamp it? — a draft approved unedited in two seconds looks different in the log than one a tech spent time revising before sending, and that difference is exactly the signal a quality process needs.
• Where did a bad outcome actually originate? — a wrong diagnosis is a drafting problem; a correct diagnosis approved without being read is a review problem, and a combined log entry cannot tell you which one you're looking at.
• Is a specific human consistently under-reviewing AI drafts? — a pattern only visible when approval events are attributable and separate from the drafts they're attached to, not folded into a single anonymous "action taken" record.

A combined log entry can tell you what happened. Only a separated one can tell you whether a human actually decided it.

This is the same principle behind draft-then-approve as a design, applied to the audit trail instead of the workflow: accountability for a consequential action should stay traceable to the specific human who approved it, not blurred into "the system did it" once an AI was involved anywhere upstream. It's a small architectural choice compared to the AI capabilities themselves, and it's one of the ones we think matters most once any of those capabilities are actually trusted with real client work.