The one page you keep.
Every AI pilot ends in a decision. Most of those decisions are never written down — so eleven months later, when the champion has left and the CFO is deciding whether to write it off, nobody can say whether the thing worked. The Go/No-Go Decision Record is the one page that fixes that: the decision, its owner, the signed baseline, the acceptance-test result, the kill rule, and the date you re-check. Fill it at the meeting, keep it for the trail.
a decision with no record is a decision nobody owns
Here is the pattern, drawn from a synthetic teardown but true to the documented evidence. A mid-market manufacturer spent $342,000 and eleven months on an AI quote-desk pilot. The CFO killed it at a budget review — and nobody in the room could say whether it had worked. Not because the model was bad. Because nothing had ever been written down that "worked" could be checked against. The whole $342K reduces to a missing page.
An unrecorded decision fails in three specific ways, and they are the three things a record is for:
- owned With no record, ownership diffuses across a committee, a vendor, and a slide deck. When it goes wrong, everyone points sideways. A record names one accountable person and puts their signature on the outcome. The champion who leaves in month two can be replaced; the signature can't be un-signed.
- audited A decision you can't reconstruct is a decision you can't defend — to a board, a lender, an insurer, or yourself in a year. "We thought quotes took three days" is folklore. "We measured 26 hours, signed on this date, and set this pass/fail line" is a record that survives questioning. The former is how a six-figure write-off arrives with no explanation attached.
- reused Discipline that lives in one person's head dies when they change roles. A record is a form: next pilot, you fill the same ten fields, and the standard travels with the paper instead of the person. That is how a buyer builds a habit of not getting burned instead of relearning it each cycle.
This is spoke 05 of the Verification Field Guide. The other spokes give you the pieces that fill this page — the problem statement, the baseline, the acceptance test, the kill rule. This one is the container that makes them a decision instead of a pile of documents.
what the record contains
Six load-bearing entries. Miss any one and the page stops being a decision and becomes a wish. None of them is optional, and three of them — the problem statement, the signed baseline, and the acceptance-test result — carry the whole weight.
Go, No-Go, or Revisit — dated, one owner.
The verdict in three words, the date it was made, and the one name and title accountable for it. Not a committee, not the vendor. The person whose signature is on the baseline is the person who owns the outcome.
The problem statement, in one sentence.
What operating metric, in what segment, moving from what to what. "Cut engineered-to-order quote cycle from 4.5 days to 2 days" — not "improve quoting with AI." If the decision doesn't answer a named problem, there is no decision to record.
The current number, measured and signed.
Today's value of that metric, the query or method it came from, the date, and the sponsor's signature that it is fair. Folklore does not qualify. If you can't fill this box from your own system data, the page stops here.
The pass/fail line, and where it landed.
The written test the pilot is judged against — thresholds, dataset, who runs it — and, at each review, whether it passed, failed, or hasn't run. "Criteria to be defined during the pilot" is an empty box and a default No-Go.
The conditions that stop it early.
Set at the start, in writing: the tripwires that end the pilot without another meeting — adoption below a line by a date, an integration slip past a limit, the sponsor leaving with no replacement inside 30 days. A kill rule set later is a kill rule that never fires.
The fixed date you check again.
A decision without a next review is a decision that quietly becomes permanent. The record names the date the baseline gets re-pulled and the page re-signed — so a pilot ends by a decision, not by fading out of the budget.
the method: fill it in at the meeting, not after
The single rule that makes the record work is when you fill it. In the room, before the money moves — never afterward. A record written after the decision is a justification for what you already did; a record written during it is the decision itself.
- before
The baseline is pulled before the meeting, not promised in it
The number in box c comes from your own system, queried before anyone sits down. If the honest answer at the meeting is "we'll measure it during the pilot," the pilot is already unmeasurable — and that finding is worth more than the meeting. Measure first, then decide.
- during
The three load-bearing boxes get filled live, out loud
Problem statement, signed baseline, acceptance test — read them into the room and fill them as you go. If any of the three can't be filled honestly while everyone watches, the default answer is No-Go. Not because the tool is bad: because you cannot yet tell whether it worked, and you don't sign what you can't check.
- after
The page is signed, kept, and re-opened on the review date
One accountable signature closes the meeting. The record goes in the file — not the vendor's file, yours. On the re-measure date it comes back out, the baseline gets re-pulled, and the page is re-signed with the new reading. A record with an empty box and a signature is worse than no record at all.
a worked example
Synthetic and clean-room — a fictional company, invented numbers, built to show the shape. This is the record the $342K pilot never had. Read it as the counterfactual: one page, filled at week zero, that would have turned an eleven-month write-off into a three-minute No-Go.
The setup. Fallowfield Fluid Components (fictional, ~$185M valve-and-fittings manufacturer-distributor) is deciding whether to run an AI quote-desk copilot, sold on "quote turnaround from days to hours." The champion has a demo, a vendor number of "$1.4M annual margin impact," and a belief that "quotes take about three days." Before the decision meeting, an analyst spends one afternoon pulling actual quote records. Here is the record the meeting produces:
GO/NO-GO DECISION RECORD — AI PILOT (worked example — synthetic)
1. DECISION DATE / REVIEW DATE
Decided 2025-06-05. Re-review 2025-09-05 (90 days).
2. ACCOUNTABLE SPONSOR
J. Okafor, VP Sales Operations. One name. Not the vendor, not the committee.
3. PROBLEM STATEMENT (one sentence, names metric + segment)
"Cut median quote-cycle time on the slow segment."
PROBLEM ON EXAMINATION: the slow segment is engineered-to-order (22% of volume,
median 4.5 days). The tool cannot read engineering drawings, so it cannot
operate on that segment at all. The segment it CAN touch — catalog quotes,
78% of volume — already runs at a median of 26 hours.
4. BASELINE — MEASURED, SIGNED
Catalog quotes: median 26 hours [counted, TradeCore Q- records, 2025-05]
ETO quotes: median 4.5 days [counted, TradeCore Q- records, 2025-05]
Folklore figure was "about 3 days." It did not survive the query.
Signed fair: ______________________ (sponsor)
5. TARGET + DATE + WHO MEASURES
Vendor promise: "under 4 hours." Applies only to the catalog segment, which is
already at 26 hrs. On the ETO segment where the target would matter, the tool
cannot run. No coherent target can be written. Measured by: our sales-ops team.
6. ACCEPTANCE TEST — WRITTEN
SOW clause reads: "success criteria to be mutually defined during the pilot."
That is an empty box. No pass/fail line exists → default No-Go.
7. KILL CRITERIA
Not applicable — the pilot fails box 3 before a kill rule is needed.
8. SPEND CAP
Proposed: $104,500 license + ~$88,000 integration + ~$118,000 internal hours,
loaded ≈ $310K first year, before the $31K item-master cleanup surfaced in wk 9.
9. BUILD/BUY + VENDOR EXPOSURE
External tool; 14% duplicate SKUs in the item master would block the matcher —
discovered on examination, would have been discovered after signing.
10. DECISION
[ ] GO [X] NO-GO [ ] REVISIT
Reason: the pilot targets a segment already fast (26 hrs) and cannot operate on
the segment that is slow (4.5 days). Redirect the AI budget to the MEASURED
$610K/yr order-entry-error problem finance already costed.
Sponsor signature: ______________________ Date: 2025-06-05The record does its whole job in box 3 and box 6. The moment the problem statement is examined against the signed baseline, the pilot is dead — the tool was bought to fix a segment that was already fast and can't run on the one that's slow. No model was ever tested. No $342K was ever spent. The decision took the length of one meeting because the number was pulled before it started. That is the difference between a record and a hope.
the template
Copy it, paste it into your own doc, and fill it at your next AI decision meeting. Ten fields, one page. Boxes 3, 4, and 6 are the record — if any of the three can't be filled honestly, the answer is No-Go by default. Free to use and share; not legal, financial, or professional advice.
GO/NO-GO DECISION RECORD — AI PILOT
Fill this in AT the decision meeting, BEFORE any dollar or signature moves.
Free to use and share. Not legal, financial, or professional advice.
1. DECISION DATE / REVIEW DATE
Decided: [date this record is signed]
Re-review: [the FIXED date it will be re-measured and re-signed]
2. ACCOUNTABLE SPONSOR
[ONE name and title]. Not a committee. Not the vendor.
The person whose signature is on box 4 and box 10.
3. PROBLEM STATEMENT (one sentence — must name the metric AND the segment)
[e.g. "Cut ___ quote cycle in the ___ segment from ___ to ___."]
Not "improve ___ with AI." If it doesn't name a metric and a segment,
there is no decision to record.
4. BASELINE — MEASURED, SIGNED
Metric: [name] Current value: [number]
Source/query: [the system + query it came from]
Date measured: [date]
Folklore ("about ___") does NOT qualify. If you cannot fill this box
from your own system data, STOP HERE — the answer is No-Go.
Signed fair by sponsor: ______________________
5. TARGET + DATE + WHO MEASURES
Target value: [number] By when: [date]
Measured by: [must be YOUR team, on YOUR data — not the vendor]
6. ACCEPTANCE TEST — WRITTEN
Pass/fail threshold: [the line]
Runs on: [which dataset / period] Run by: [name] Run date: [date]
If the contract says "criteria to be defined during the pilot," this box
is empty and the default answer is No-Go.
7. KILL CRITERIA (set now, in writing — the tripwires that stop it early)
- [e.g. adoption < __% by month 3]
- [e.g. integration slip > __ weeks]
- [e.g. sponsor departs with no replacement inside 30 days]
8. SPEND CAP
Total authorized (license + integration + internal hours, loaded): [$]
Burn to date at this review: [$]
9. BUILD/BUY + VENDOR EXPOSURE
Why external vs. internal: [reason]
If the vendor disappears in 12 months, the workflow and the data: [what happens]
10. DECISION
[ ] GO [ ] NO-GO [ ] REVISIT (date: ______)
Sponsor signature: ______________________ Date: ______
HOW TO USE IT: boxes 3, 4, and 6 are the record. If any of the three cannot be
filled honestly, the answer is No-Go by default — not because the pilot is bad,
but because it is unmeasurable, and unmeasurable pilots are how a six-figure
write-off arrives with nobody able to say whether the thing worked.
Re-sign at every review date. A record with an empty box and a signature is
worse than no record.how you reuse it to build an auditable trail
One record is a decision. A folder of records, one per pilot, re-signed at every review, is an auditable trail — the thing that lets you answer "how do we decide on AI spend?" with a file instead of a story.
- per pilot
One record, opened before the money
Every AI proposal gets a record before a dollar or a signature moves. Same ten fields, every time. The form is the standard; the standard travels on the paper, not in the head of whoever happened to be in the room. New people inherit the discipline instead of relearning it.
- per review
Re-pulled, re-signed, on the date you set
On each re-measure date the baseline gets re-pulled and the page re-signed with the new reading and the current spend-to-date. A pilot that isn't moving toward its target shows it in the numbers, on schedule — so it ends by a decision at month three, not by fading out of the budget at month eleven.
- across pilots
A file that survives turnover and questioning
The folder is what a board, a lender, or an insurer sees when they ask how you govern AI spend: dated, signed, measured decisions — including the No-Gos, which are the ones that prove the discipline is real. Every No-Go on file is money not written off, and a decision somebody can still explain a year later.
The record fills itself from the rest of this guide: the problem statement and the signed baseline come from the same discipline as the Problem-Definition Audit, and the acceptance-test result is the pass/fail line the audit writes for your own team to run. This page is where those pieces become a decision you own.
You can start one of these in a room full of operators, on a real decision, this month.
The Pilot Autopsy is a free, monthly, 60-minute live teardown of a synthetic failed AI pilot — and it walks this record field by field against a case where every empty box cost real money. Bring a live decision of your own and you leave with a started record, not a slide. If you want the paid version — a signed baseline, a written acceptance test, and a completed record on one of your own AI decisions — that is the Problem-Definition Audit, $12,500 fixed, with $2,500 invoiced only when your scoped pilot passes its own test. It all runs under a charter that sells no implementation and takes no vendor money — which is why the buyer's-side playbook is the thing we can give away.
the rest of the field guide
This record is the container. These four spokes are what fill it — each one a usable tool for the buyer's side of an industrial AI decision. Start anywhere; they all feed this page.
questions a skeptical operator asks
- Q What is an AI go/no-go decision record? It is the one page you keep after deciding whether to run an AI pilot: the decision itself, the date, the one accountable owner, the problem statement it answers, the sponsor-signed baseline, the acceptance-test result, the kill rule that was set, and the date it will be re-measured. It is not the pitch deck and not the vendor's proposal — it is the buyer's own record, signed, so the decision can be owned, audited, and reused.
- Q When do you fill it in — before or after the decision? At the decision meeting, before the money moves — never afterward. A record written after the fact is a justification, not a decision; it records what you wish you had thought, not what you actually knew. If the baseline box or the acceptance-test box cannot be filled honestly in the room, that is the finding, and the default answer is No-Go — not because the pilot is bad, but because it is unmeasurable.
- Q Why does a decision need a record at all? Because a decision with no record is a decision nobody owns. Eleven months later, when a champion has left and a CFO is deciding whether to write it off, there is no way to say whether the thing worked — nothing was written down that "worked" could be checked against. The record fixes ownership to one name, makes the decision auditable after the fact, and lets you reuse the same discipline on the next pilot instead of relearning it.
- Q Does filling in the record mean the pilot will succeed? No. A signed baseline and a written acceptance test do not make a pilot succeed — they make it measurable, which means they make it killable early and cheap. That is the whole claim. The record turns an eleven-month, six-figure write-off nobody can explain into a three-month decision somebody signed, or stops the spend at week zero. It is a discipline for not getting burned, not a guarantee of success.
- Q Where can I get one filled in for a real decision? Two free routes and one paid one. Bring a live decision to a Pilot Autopsy and walk out with a record you started in the room. Or take the free Industrial AI Scoreboard to see whether your organization is even ready to fill one honestly. The paid next step is the Problem-Definition Audit — $12,500 fixed — which delivers a sponsor-signed baseline, a written acceptance test, and a completed decision record on one of your own AI decisions, with $2,500 of the fee invoiced only when the scoped pilot passes.
Can your team fill box 4 honestly?
If the baseline box gives you pause, that is the finding. The free Industrial AI Scoreboard — twenty scored questions, no sales call — tells you which of the four disciplines will fail you first, before you sign anything.
Take the ScoreboardThat doesn't look like an email — try name@company.com