Most AI assessments find what they were paid to find. This one is built so that it can return nothing — which is the only reason to believe it when it returns something.
Most of the value of this method sits in the engagements it declines.
Is there one process that runs at least weekly, with more than one person touching it? If not, there is nothing to automate yet.
Can you name one number that would move if it got faster or cheaper? If not, neither of us can tell afterwards whether the work succeeded.
Who signs, and are they in the room? If not, this is free education for someone who cannot buy.
Declining here costs an hour. Not declining costs the engagement, the reference and three weeks.
1
Day 1Observe
One recorded call · 45–60 minutes
The only job is to find out what actually happens, which is never what the org chart says happens.
We talk to the person who does the work, not only the person who owns it — and not both in the same room, because nobody describes their workarounds in front of their manager.
It opens with "walk me through yesterday". Asking what your problems are gets answered with policy; yesterday gets answered with behaviour.
The most valuable question in the set is what you have already tried to automate, and what stopped it. A failed attempt names the real constraint.
No tools are named. The moment one is, everyone starts performing for it and the rest of the interview is worthless.
2
Day 2Analyse
Transcript first, model second
Every candidate task is listed with no limit and no target. The number that survives is the finding, not a quota to fill.
Repeatable — does it run on a schedule or a trigger, at least weekly? If not, it is noise.
Legible — could the rules fit on one page a competent new hire could follow? If not, knowledge capture comes first.
Bounded — if it produces a wrong output, is that caught before it reaches a customer or a ledger? If not, it is still buildable, but the review step gets designed and priced rather than wished away.
The candidates we killed are kept, named, with the test each failed. That list is usually longer than the survivors, and it is the most credible page in the report.
3
Day 3Price the friction
Your figures, our arithmetic, shown
Frequency × duration × loaded rate = the annual cost of doing nothing.
Duration is elapsed, not touch time. The waiting is the cost, and it is the part nobody counts.
The arithmetic is shown in full. A client who can reproduce the number will defend it internally when we are not in the room, and that is where the decision is actually made.
It is stated as a range. A range survives a finance director; a point invites an argument about the third decimal that we would lose, because it was never that precise.
4
Day 4The report
Five pages, one page each
What we saw, in your words. What survived and what did not. The first build — one, not five. What comes after. What to commit.
The first page is quoted from the transcript, because it is the page you believe: you said it.
Wave two is named in advance as the place where the curve flattens and people lose their nerve. Naming it is what carries a firm through it.
Tools are named last, and only for steps that passed all three tests.
If we cannot choose one first build, Day 2 is not finished.
5
Day 5Decide
Thirty minutes, live, screen shared
The pack is never sent ahead. A report read alone converts at a fraction of one walked through.
Which of these is most urgent for you?
Do you want to build it yourselves, or have it built?
What is your timeline?
Then we stop talking. The silence after the third question is doing work.
The sizing rule
Commit a quarter of your best estimate, and never more than a small stated share of the year’s discretionary budget on a first build.
The reason is the shape of the loss, not caution for its own sake. Committing less than optimal costs time, and the penalty is gentle. Committing more costs the programme, and the penalty is not linear at all — at around twice the right size, more than half of otherwise-sound bets end below where they started.
The Day 3 arithmetic is exact. The probability quietly assumed underneath it — that the build actually delivers — is a guess. When the inputs are guesses, the honest engineering response is to commit less than the formula says. And where a prerequisite is missing — no owner, no data, nobody who can decide — the recommended commitment is zero, whatever the arithmetic returns.
The ladder
Each rung is recommended by a report you have already read, not pitched cold. Most engagements stop at the rung that solved the problem, which is the correct outcome.
The gateAlways. Fifteen minutes, free — it qualifies or it declines.
AssessmentPassed the gate. Five days, one recommendation.
Knowledge captureCandidates failed legible — the rules live in someone’s head.
Process redesignThe process itself is the problem, not the tooling.
First buildOne workflow end to end, with the review step priced in.
RetainerYou want capability, not a project.
Prices are quoted in the conversation, against your own friction figure, in your own currency. The anchor we hold ourselves to: the assessment should cost less than one month of the friction it finds. If it does not, you are right to refuse.
Start where it costs an hour
Three questions, about a minute, and any one of them can rule the work out on its own. If it does, you get the specific first step instead of a proposal — and that is the whole point.
