A pilot brief should fit on one page. If it needs ten, the scope is not yet small enough to pilot. Below is the template I would use, with a note on why each field is there. Copy it, fill it in a meeting, and refuse to start until every line has an answer, even if the answer is "unknown, to be measured in week 1".
The template
AI PILOT BRIEF 1. Process One process, named in a sentence. 2. Owner One person who can say "accepted". 3. Users Who uses the output, how many. 4. Baseline Volume per month; minutes per item; error rate today (measured on a sample of N = ___ items, date ___). 5. Target Number(s) that mean success, per metric, with the measurement method. 6. Data Source, format, who owns it, personal data? (Y/N), can it go to an external service? (Y/N, who decided). 7. Golden set Size, who labels it, where it lives, frozen on ___. 8. Human path What goes to a person, who handles it, time limit. 9. Out of scope What we are explicitly not doing. 10. Timeline Start, checkpoint, decision date. Length: 2 to 6 weeks. 11. Cost Pilot budget, expected running cost per month. 12. Risks Top three and what we do about each. 13. Decision What happens at the end: scale, change, or stop, and the rule that picks between them. Notes on the fields that get skipped
1 and 9, scope. The most common failure is a vague process ("customer experience"). Write one sentence naming the input and the output. Then write what you are not doing; it prevents scope creep in week 3.
4, baseline. If nobody has timed the manual work, put the measurement into week 1. Without it you will end the pilot unable to say whether anything improved.
5, target. Write the number before any result exists, and say how it is measured. Example: "at least 70% of items handled without a person, at least 97% accuracy on those, measured on the golden set and on a week of live traffic."
6, data. Two questions people avoid: is there personal data in it, and has someone with authority decided whether it may leave your systems? A name next to the decision is better than a hope.
8, human path. The pilot is not "AI replaces the person". It is "AI handles what it is sure about, a person handles the rest". Say who and how fast, or the doubtful items will sit in a pile.
12, risks. Typical top three: data quality worse than expected, reviewers overloaded, and the target set too high or too low. Each needs a response, such as "re-measure after week 1" or "lower the share automated".
13, decision. Fix the rule in advance: for example, "if the target is met on both the golden set and live traffic, scale to the next process; if met only on the golden set, extend two weeks; otherwise stop". It protects the team from arguing about a result after the fact.
A word on length and tone
One page is a discipline, not a formatting preference. Short sentences, numbers where possible, plain words. The audience includes the owner, the person who pays and the person who will review the output every day.
Takeaways
- Use one page and refuse to start with blanks.
- Measure the baseline and write the target as numbers first.
- Name the owner, the human path and the stop rule.
- Fix the decision rule before you see results.
Portfolio and contacts: @Kotenev01_bot.
I write about practical, measurable AI adoption for business: demos on test data, no invented case studies. Telegram channel: @kotenev01_en


