Champion, blocker, executive sponsor
Give the champion, the sponsor and the blocker what each needs weekly, handle the blocker who is right, and re-plan an 8-week pilot when the champion leaves.
On this page
Three people decide whether your eight weeks become a contract. You see the champion most days, the executive sponsor perhaps four times, and the blocker whenever the blocker chooses. Each needs a different thing from you each week, and the pilot fails quietly when one of them stops getting it.
Who they are on the claims floor
The champion is the Head of Claims Operations. She got the first week of triage logs released, she put her name on the scoping doc, and she runs the Friday meeting where the handlers hear what changed. She has influence without budget: she cannot sign, but the sponsor will not sign against her.
The executive sponsor is the VP of Claims. He owns the 30% number, he opened the kickoff asking for a chatbot, and he carries the case to the COO who signs. He will give the pilot about 30 minutes a fortnight, and he reads one page.
The blocker is the information-security lead. She was not at the kickoff. She found out in week 2 that claim text, which contains names, addresses and medical detail, was going to a model endpoint, and she has the authority to stop the pilot's data access with one ticket. The head of underwriting, whom a week in the role calls the blocker in week 3, is a second one, and like her he turns out to be right.
Then there are the 28 handlers, who are not stakeholders in the deal sense and who decide, from week 7, whether the thing gets used.
What each needs from you every week
| Stakeholder | Needs, every week | In what form | If they stop getting it |
|---|---|---|---|
| Champion | the number, the misses, and something to show the floor | Friday, 20 minutes: the eval results and five real claims the model got wrong | she stops defending the pilot in rooms you are not in |
| Sponsor | the number, the delta since last time, the one decision he has to make | one page, sent the day before, in his units: minutes per claim and handler-hours | he forms his view from whoever else talks to him |
| Blocker | written answers to specific questions, inside the week | two lines per question, with the mechanism: which fields leave the tenant, what is logged, who reads the logs | the questions become a formal review, and the review becomes a gate |
| Handlers | a change in the screen they already use, and the sense that their corrections are read | the retrieved clauses inside the triage screen; a visible count of corrections applied | they stop correcting, and the eval goes blind |
Three rules run across the rows. The number is the same number for everyone: the agreement figure, the measured median once the assisted phase starts in week 7, the cost multiple. The form differs: the champion gets the claims, the sponsor gets the page, the blocker gets the mechanism. And the cadence is yours to keep: the weekly update goes out whether the week was good or not, because a missed update is read as a bad week.
The blocker who is right
In week 2 the security lead asks a question nobody on the vendor side wants: "Are you sending our claimants' medical details to an external model endpoint?" The answer is yes. She says the pilot stops until it does not.
She is right. The scoping doc said "claims data" and nobody listed the fields. The engineer's first instinct is to go to the sponsor and ask him to overrule her, and that instinct costs the deal twice: the sponsor spends authority he needs for the contract, and the blocker becomes a permanent no that outlasts the pilot. The second instinct, to argue that the model provider's terms cover it, is a legal answer to an engineering question.
The move that works is to agree, in writing, inside the day. Then produce what she needs: the list of fields the triage prompt uses, with the structured personal columns excluded, names, phone numbers, account numbers and addresses in the loss description replaced by placeholders inside the tenant, and the redaction's measured miss rate, about 3% of mentions on a 200-description sample, stated; the retention setting on the endpoint; who can read the request logs; a one-page data boundary she keeps. Claim text with the claimant's name replaced by PERSON_1 triages just as well, and the eval shows it, which is a second reason to have the eval early. Her ticket closes in three days, and in week 7 she is the person who tells procurement the security review is already done.
A blocker who is wrong is rarer and is handled the same way: in writing, with the mechanism, and never through the sponsor. The difference is that the answer is "no personal columns leave the tenant, and here is the measured miss rate on the free text", and the boundary diagram behind it is the one a threat model starts from.
The champion leaves in week 4
On Monday of week 4 the Head of Claims Operations gives notice; Friday is her last day. Count what she was holding.
The gold set: 140 of 200 cases labelled by her two senior handlers, 60 to go, with her adjudicating the 18 disagreements so far. The Friday meeting, which is the only time the handlers hear the number from someone they trust. The sponsor's attention: he reads the page because she tells him to. And the scoping doc, which has her name beside the definition of done.
Four weeks remain. A successor who did not label cases needs roughly two weeks to be able to defend the number to the sponsor, so without intervention the plan is a week-10 result: two weeks past the date the sponsor gave the COO, with one week to spare before the budget closes in week 11. The arithmetic gives you the options. Keep week 8 and shrink scope: the gold set stops at 160 cases, the 110 motor and 50 property, and the volume gate drops from 90% of daily claims to those two segments, 80% of volume. Or keep the scope and move the date to week 10, said to the sponsor in week 4, not discovered in week 7.
What you do in the five days: finish the labelling, since 60 cases by two reviewers at 11 minutes is 22 reviewer-hours, a little over two hours a day each for the two senior handlers if nothing else is asked of them; sit with the champion and her deputy for the week-4 Friday meeting so the handlers see the handover; get her to put the successor's name on the scoping doc before she leaves; and write the week-4 update yourself, with the number, the departure, and the decision the sponsor has to make, on one page. The case exercise works the same week with the sponsor away.
In an interview
The customer-scenario round puts one of the three in front of you: the blocker's question, the sponsor's "where's my 30%", the champion's resignation. What is being tested is whether you know what each person needs, in their form, and whether you handle the blocker without the sponsor. Name the three, say what each gets from you weekly and in what units, and when the blocker appears, agree with her first.
The mistake that loses points is escalating a blocker to the sponsor. The interviewer hears an engineer who spends the sponsor's authority on a problem the engineer could have solved with a field list and a day, and who will do it again in week 7 when the stakes are higher.
Check yourself
1. The sponsor replies to your week-3 page with "great, keep going" and no questions. Good sign or bad?
Bad, slightly. A page with the right number in it produces a question or a decision; "keep going" means the page carried nothing he had to act on, or he did not read past the first line. Check which: next week, put the one decision at the top in bold, with a date. If the reply is the same, he is forming his view elsewhere, and the champion needs to be in the room when he does.
2. The security lead wants the model run inside the customer's cloud tenant. Assume the listing there costs twice the per-token rate, not the usual same r. The pilot is in week 3 of 8. Agree?
Not yet. Her concern is personal data leaving the tenant; redaction inside the tenant answers most of it at no rate change, and the eval can show in a week that triage quality holds on the redacted text. Offer that first, in writing, with the eval result and the measured 3% miss rate attached, since that residual is the part only an in-tenant model removes. If she still requires in-tenant after seeing it, that is a scope change for the sponsor and sales, costed at twice 6r a day, 12r, and the where the model sits trade-offs apply.
3. The champion's successor is the deputy who said at kickoff that "the handlers won't trust a machine". Do you change the week-8 plan?
No; you change who presents the number. The deputy's scepticism is an asset if he labels 20 cases himself in week 5 and finds the agreement figure is real. Ask him to adjudicate the remaining disagreements, put his name beside the definition of done, and have him present the week-6 Friday meeting. A sceptic who has done the labelling is a better champion to the handlers than an enthusiast who has not.