THE SHORT ANSWER
Four lines, not one number. Cost per answered question as measured, in the customer's unit of work. The multipliers that will grow it, each with the decision that triggers it. A name beside each multiplier. And three dates to re-measure: two weeks in, go-live, and day 60. Lesson 04 of the Becoming Forward Deployed Engineer course shows why one number fails. It measures a claims question-answering system for 120 adjusters at $60.19 a month in model cost, then shows an eight-step agent (8 times), ten times the context (3.5 times), and retries (about 1.25 times) multiplying it to $2,106.65. Each decision is reasonable and they arrive months apart, after the pilot was approved on the sixty. I put the four lines on the one-page handshake I require before any contract with a forward deployed engineer attached is signed. My own measured multiplier: a four-agent content pipeline I built cost 12 times a single agent and ran at 8 times the latency, with lower quality.
The cost question gets asked once, in a scoping call, and answered with one figure. That figure is usually honest and usually small.
What goes wrong comes later and looks like good engineering each time. Someone lets the agent plan its own work. Someone retrieves more context. Someone adds retries. Nobody in those three meetings is looking at the number that was approved.
The table, the four lines, and the handshake page they go on are in The Pilot Was Approved on the Sixty. The case for cost per outcome as the unit is in Three Token Bills, One Unit.