How to use it
- 01
State the original target
Write down the success measure agreed before the pilot began. If none was set, say so. A pilot judged against a target invented afterwards will always appear to have passed.
- 02
Answer from evidence
For each scored question, choose the option you could support with a document, a number or a log. Where you would have to rely on an impression, the lower option is the honest one.
- 03
Read the low answers
The band summarizes. The questions that scored one or zero are the content of the decision, because each names a gap that can be turned into a condition or a reason to stop.
- 04
Bring it to the meeting
Export the record and attach the evidence behind it. A decision meeting that starts from one page of scored answers is shorter and less political than one that starts from a demonstration.
What acceptance of an AI pilot should rest on
A pilot that people liked is not the same as a pilot that passed. Acceptance needs evidence in four areas, and the checklist asks two or three pointed questions about each.
| Part | What it settles | Why it matters |
|---|---|---|
| Results | Target reached, quality measured, serious failures absent | Without measurement there is nothing to accept |
| People | Real use, and review that changed things | A tool nobody uses, or a review nobody performs, will not survive rollout |
| Operation | Cost per task, data terms and ownership | These are what turn a successful pilot into a failed service |
| Decision | Open issues, conditions and a recommendation | Makes the outcome explicit: roll out, extend or stop |
A worked example
A property manager's email pilot, assessed
The example is an invented property management company at the end of a pilot. Load it to see the scores. The pilot went well by any everyday standard, and the record still says extend.
- StrongThe target was met, managers used the drafts daily and they edited them when needed, which shows review was real.
- GapsQuality was checked on an informal sample, one serious fault had to be fixed, cost is known only at pilot volume and two data questions are open.
- OutcomeFour conditions with owners. The recommendation is to extend the pilot until they are met, then decide again.
Nothing here argues for stopping. The record simply shows that the evidence for a full rollout is not complete, and it lists what would complete it.
What this tool does not do
The checklist structures a decision. It cannot make the measurements it asks about.
- It does not test your system. The answers are only as good as the evidence behind them.
- It is not a legal or compliance review. Sector rules and privacy law need their own check.
- The bands are a guide. One zero on a serious failure can outweigh a high total.
- Your answers are stored only in this browser and are sent nowhere.
What to do with the finished document
Take the record, with its evidence, to whoever funds the next stage and agree one of the three outcomes. If you would like an outside view of the pilot's results, send the record and we will tell you what we would test. Send it through the contact page or book a 30-minute call and bring it along.
Common questions
What score means an AI pilot has passed?
The top band suggests the evidence supports a staged rollout, but read the individual answers before relying on it. A zero on the question about serious failures, or on review, should hold a rollout back whatever the total. The score is there to prompt a discussion about specific gaps, and it does not replace judgement.
What if no success measure was set before the pilot?
Then the pilot cannot be accepted on its results, because there is nothing to compare them with. Say so plainly, choose the lowest option for that question, and set a measure now. A short extension run against a stated target is worth more than a long pilot that was judged by how it felt.
Why does the checklist ask whether reviewers changed anything?
Because approval without changes can mean two opposite things: the system is very good, or nobody is really checking. If reviewers never edited or rejected an output, confirm quality with an independent sample before trusting it. A review step that has become a formality gives no protection once volumes rise.
Is stopping after a pilot a failure?
No. A pilot exists to find out cheaply whether something is worth building properly. Learning that the task was unsuitable, the data was not ready or the benefit was too small is a valid result. The failure to avoid is drifting into a rollout because nobody wanted to make the decision.