What the deal depends on
Turn the pilot fee, the annual contract and the expansion into the three numbers your pilot has to produce, and know which documents the buyer reads.
On this page
Three numbers sit behind an AI pilot: the pilot fee, the annual contract it is meant to become, and the expansion that justified approving it. The engineer is paid from the first, measured on the second, and the third is why the vendor agreed to send one engineer on site for eight weeks. Knowing the ratio between them changes what you build, because it says what the pilot is for.
Three numbers and their ratio
Assume the claims pilot is sold at $50k for eight weeks, the annual contract for the triage floor is $250k, and the plan sales wrote has three more lines of business (travel, marine cargo, pet) at the same $250k each from year two, so $1M a year across all four. Over three years that is $50k + $250k + $1M + $1M, call it $2.3M, and the pilot fee is about 2% of it.
Two percent means the pilot is not a revenue event. It is an evidence event: it exists to produce the numbers the annual contract is signed on. An engineer who treats the eight weeks as a fixed-price delivery to be completed optimises the wrong thing. The deliverable is not the integration; it is the renewal case, with the integration as the thing that made the renewal case true.
The ratio also says what the vendor is risking. A $50k pilot costs the vendor roughly its fee in one engineer's loaded time and the travel. That is a bet, and the vendor made it on sales's expansion number. If the expansion is fiction, the pilot was overpriced at any fee; part of your job on site is to find out, by week 4, whether the other three lines have a sponsor, a pain and a budget of their own.
How the pilot's numbers become the renewal case
The renewal case is one page that the sponsor forwards to the economic buyer unedited. It needs three numbers from the pilot and nothing the pilot cannot measure.
The first is the outcome metric from the scoping doc's goal line, which the definition of done qualifies: the median triage time measured in the assisted phase, on the teams using the tool from week 7, against the 11-minute baseline. The second is the agreement number, because the economic buyer's first question is "how do we know it's right": 95% agreement with the senior handlers on the 200-case set. The third is cost per claim, as a multiple of the contracted per-million-token input rate r, with output at 4r: at roughly 4,300 input tokens and 160 output tokens per claim, the cost is about 0.0043r + 0.00064r, inside the 0.005r line in the scoping doc; at 1,200 claims a day that is about 6r a day and 1,500r a year, and the customer multiplies by their own r.
The arithmetic on the page is back-of-envelope short. 30% off 11 minutes is 3.3 minutes a claim, 66 handler-hours a day across 1,200 claims, about $495k a year at an assumed loaded $30 an hour and 250 working days. Against that: $250k a year in licence and 1,500r in tokens. For any r a model contract plausibly states, the tokens are small next to the licence, and the case is $495k saved against about $250k paid. The economic buyer wants a payback inside a year and gets one in about six months.
What the pilot cannot measure stays off the page. "Handler satisfaction improved" is not a number; "the median on 300 assisted claims in week 8 was 8.1 minutes, measured from the customer's own timestamps" is. The page carries the measured median, the agreement, the cost multiple, and the date each was measured, with the champion's name as the person who checked them.
What the commercial team reads, and what they never open
You will write a dozen documents in eight weeks. Three of them are read by the people who decide the contract.
| Document | Who reads it | Why |
|---|---|---|
| The definition of done, five lines | sales, the sponsor, the economic buyer | it is the contract's acceptance test |
| The weekly update, one page | the sponsor, sales, your own manager | the number, the delta, the decision needed |
| The renewal case, one page | the economic buyer, procurement | the sum above |
| The scoping doc, pages 1–2 | the sponsor and the champion, once | they signed it |
| The eval harness and its 200 cases | the champion, the handlers, your successor | nobody commercial, ever |
| The prompt, the retrieval index, the ETL | your successor and the customer's owner after week 8 | never |
| The runbook and the threat model | the customer's on-call and security team | never, but security's sign-off is on the decision path |
The pattern: the commercial team reads anything with a number and a date, and nothing with a mechanism in it. That has two consequences. First, the numbers in the documents they read must match the numbers in the ones they do not: an eval that says 93% while the weekly update says "on track for 95%" is a discrepancy the economic buyer will find in week 9. Second, the documents nobody commercial reads are still on the decision path, because the security team's answer on the threat model is a gate in procurement, and a runbook the customer's on-call cannot follow is the reason year two does not renew.
The dependency nobody wrote down
The pilot ends in week 8. Where does the contract decision sit relative to the customer's budget cycle? If the annual plan for next year closes in week 6, the week-8 result lands in a quarter with no money in it, and the contract waits for the next cycle, months away. This is the decision-process zero from the scorecard, and the dependency it hides is a date.
The engineer's part is to make the midpoint number real. If the week-4 update carries an agreement figure on the first 140 labelled cases, with the shadow run started on live claims, sales has something to put into the budget conversation before it closes; minutes per claim arrive only with the assisted phase in week 7. A week-4 number that is honest and partial beats a week-8 number that is complete and late. It also means the first four weeks are planned backwards from "what can be measured by week 4", which puts the eval set and the shadow run ahead of the integration polish. The three-day prototype exists for the same reason.
The second unwritten dependency is the security review. Left alone, it sits on the critical path between week 8 and signature, it takes the customer's security team two to six weeks depending on their queue, and it can start in week 2 if someone sends them the data-boundary page. Nobody will ask you to. Send it.
In an interview
The customer-scenario round often opens with "the pilot passed; why didn't it close?" What is being tested is whether you see the pilot as an input to a commercial decision with its own calendar, rather than as a delivery that ends when the code works. Say the three numbers the renewal case needs, say where the budget cycle and the security review sit relative to week 8, and say which of your documents the economic buyer will read and what number is in it. Then reject the alternative out loud, the way communicating trade-offs has you do: "not a longer pilot, because the budget closes before it would end."
The mistake that loses points is describing the pilot's success in your own units, agreement, latency, a demo that went well, when the question was about the customer's: minutes per claim, handler-hours, payback. An answer that never converts the 3.3 minutes saved per claim into 66 handler-hours a day has not produced a renewal case.
Check yourself
1. The pilot fee is $120k and the annual contract is $150k, with no expansion in the plan. What does that tell you to do differently in week 1?
The pilot is 44% of the first year's value ($120k of $270k), and still 29% over two years ($120k of $420k), so it is a revenue event, and the vendor is treating it as a project. Ask sales in week 1 what the expansion is supposed to be; if there is none, the renewal case has to be made on this one line, and the week-8 number has to clear the economic buyer's payback threshold alone: $495k of costed pain against $150k is clear if the 30% lands, so protect the 30% over everything else and cut scope to reach it.
2. In week 4 the tool agrees with the 140 adjudicated gold cases on 91%, up from 90% a week earlier, and the shadow run has just started at 100 claims a day; no time figure exists until week 7. Does this go in the budget conversation?
Yes, labelled as partial, with the trend and the date. 91% on their own cases, with the cause of most misses found, is evidence the mechanism works; "we'll have a number at week 8" is not. Give sales the sentence: "91% agreement on 140 of your cases at week 4, against the 95% line; minutes per claim measured from week 7, against the 30% target at week 8." The budget conversation needs a number before it closes, not a perfect one.
3. The customer's security review takes six weeks and the budget closes four weeks after the pilot ends. When must the review start?
By week 6 at the latest, and the data-boundary page it needs can be written in week 2, so send it in week 2. Started at week 8, the review finishes at week 14, two weeks after the budget closed, and the contract waits a cycle. A dependency on a calendar is planned backwards from the last date, not forwards from today.