“Who approved this?” — the question that ends agent pilots

The pattern repeats across industries. An agent pilot works — quotes go out faster, orders reconcile, the demo lands. Then a commitment surfaces in a forwarded thread: an auto-renewal, a liability figure, a term nobody remembers agreeing. Someone senior asks who approved it.

The honest answer — "the agent did, within a prompt we iterated on" — ends the pilot. Not because the term was wrong, but because nobody could produce the authority behind it.

Pilots don’t fail on capability. They fail on accountability.

Three people, three questions

General counsel asks: is this enforceable, and what do we produce in a dispute? Procurement asks: against which terms, and are they ours? Finance asks: above whose limit, and where was the sign-off? Free-form agent transcripts answer none of these. A negotiation record with named approvals answers all three, in one export.

Designing for the question

The fix is not more oversight meetings; it is making the question answerable by construction. Every commitment above a threshold carries its approval — who, when, under what authority — because the runtime would not bind it otherwise. When the question comes, the answer is a record, not a reconstruction.

If your pilot is working, this question is coming. Better to greet it with a log.