The Pilot Was Approved on the Sixty

Every scoping call for an AI deployment reaches the same question, usually from the person who signs. What does this cost?

The forward deployed engineer gives a number. It's a good number, measured and small. Eight months later the bill is many times that, and every decision that got it there made sense.

A Substack course put real figures on this last week, and I think its best line points at a page I already published.

The short version

Lesson 04 of Becoming Forward Deployed Engineer measures the model cost of a claims question-answering system for 120 adjusters at $60.19 a month, then shows three ordinary decisions, an eight-step agent, ten times the context, and retries, multiplying it to $2,106.65. Its conclusion is that nobody decides to spend 35 times more. They make three sensible decisions, and by then the pilot was approved on the sixty. I'd say the quote is a promise made before anyone owns it. The handshake page I require before a contract with an FDE attached gets signed sorts every outcome into ships, builds, or no, and it has no cost line. The rewrite adds four: cost per answered question as measured, the multipliers with their triggers, a name beside each multiplier, and three dates to re-quote the number.

What the lesson measured

The course is written under the publication's name, with no author on the page. Lesson 04 wraps the model client in a meter that records tokens and latency without touching the code being measured, then prices one run.

The rates are illustrative hosted API rates and the model is a small local one, so these are teaching numbers. The method is the point.

Two things in it I'd keep.

The first is the pair of denominators. Divide by questions asked and you get one figure. Divide by questions answered correctly and you get nearly twice that, because the failures retrieved documents and produced tokens too. The lesson calls pricing only the successes the most common mistake in the exercise, and says it always points the same way. That matches what I argued in Three Token Bills, One Unit: product's unit is cost per outcome, and it's the only unit that tells you whether to cut the tokens or leave them alone.

The second is the sensitivity table.

The lesson's reading is the sentence I'd put on a wall: "Nobody decides to spend 35 times more." Each of the three is a reasonable call. They arrive months apart. Multiplied, they're the overrun, and the approval happened back on row one.

It also says the leading reported cause of finished AI systems being cancelled is infrastructure cost landing several times over projection. It names no source for that, so I'll leave it as the lesson's claim.

The quote is a promise

Here's the part the lesson leaves open. Who holds the sixty after it's said out loud?

I've written about this problem from the outcome side. In In the Room, Not on the Quota: Who Owns the FDE Promise, the instrument is a one-page handshake attached to the contract. Every outcome in the proposal gets sorted by the FDE into three columns. Ships, with a date. Builds, with an estimate and the engineer who has to say yes. No, written down before the signature. Sales owns the number, the FDE owns the columns, both names go on the page.

That page was built to stop "mostly" from being sold as "yes." It has no cost line, and after this lesson I think that's a gap in my own template.

A cost quote behaves exactly like an outcome promise. It's made in a scoping call, it's believed, and the people who later change it weren't in the room. The engineer who adds a planning loop in month four isn't breaking a promise. They never saw it.

The rewrite: four lines

"What does this cost" stops getting one number. It gets four lines, on the handshake, before the signature.

1. Cost per answered question, as measured. In the customer's unit of work. Per claim question, per resolved ticket, per reviewed contract. State the accuracy it was measured at, since the figure moves with it.
2. The multipliers, each with its trigger. More steps per task. More context per call. Retries. Write the decision that would switch each one on, in plain words: "if we let the agent plan and re-check its own work."
3. A name beside each multiplier. Whoever would make that decision owns the re-quote. If turning on the eight-step agent is a product call after handback, it's a product name. A multiplier with no name gets switched on by whoever finds it useful.
4. Three dates. The number gets re-measured two weeks in, at go-live, and at day 60. Those are the three moments in Absorb Pain, Excrete Product: The FDE Handback, so the cost lines travel in the same packet as the customer's eval rows.

One line on latency belongs next to these. The lesson's load test shows one machine saturating at two concurrent questions, with throughput flat and the median doubling. Its point is that you buy more machines to keep the answer quick, which is a different purchase from buying for volume. Quote the token cost as a number and the latency as a range.

A multiplier of my own

The lesson's table assumes an eight-step agent costs eight times as much. I'd treat every multiplier as something to measure, because mine didn't come out linear.

Number 10 in 10 AI Agents I Built That Failed. The Honest Retrospective. is a four-agent content pipeline: a researcher, a writer, an editor, and a publisher. Four agents, not eight steps. Cost was 12 times a single agent. Latency was 8 times. Quality was lower than one well-prompted agent.

Nobody approved a 12x. I approved "let's split this into specialists," which sounded like an architecture decision and was a pricing decision.

That's why line three matters most. The multiplier and the person who triggers it have to be on the same page, or the decision gets made as engineering and discovered as finance.

One small note on reproducing the table, since the lesson's own fourth rule is to round once and print what you computed with. The retries row prints 1.2 beside $75.24, and $75.24 is 1.25 times sixty. A one-decimal format will do that. The 35 only works with the 1.25. If you rebuild the sheet for a customer, carry the extra digit.

Where this sits

This is a standalone piece next to The FDE Transition, and a companion to Run the Oracle Before You Swap the Model, which took the previous lesson in the same course. That one splits a wrong answer into whose problem it is. This one splits a cost into whose decision it is.

It belongs to the argument on AI Product Management: the build got cheap, and what's left is the part nobody trained for. Measuring a token bill is cheap now. Owning the number through three reasonable decisions is the work.

This week: find the last cost figure your team quoted for an AI deployment. Ask which accuracy it was measured at, what would multiply it, and whose call each multiplier is. If the answer to the third is "nobody's," that's the line to add.

Related answer: What should an AI pilot cost estimate include?

Sources: Becoming Forward Deployed Engineer (author unnamed), "Lesson 04: What the Working Half Costs," October 10, 2026, read in full; the figures are from the course's measured run at illustrative rates. 10 AI Agents I Built That Failed. The Honest Retrospective., falkster.com, May 4, 2026, for the four-agent pipeline.