Write the AI pilot's stop rule before you write the prompt

Mention last spring's AI pilot in most offices and you get a shrug. It did not fail, exactly. People used it for a few weeks, then less, then the subscription renewed for two months before someone cancelled it. Nobody wrote down what was learned. Six months later a different team starts a pilot on the same idea.

The big surveys from last year suggest this is the normal path.

• S&P Global's 2025 survey of 1,006 respondents found 42% of companies had abandoned most of their AI initiatives, up from 17% a year earlier. On average, 46% of proofs of concept were scrapped before production.
• Gartner forecast in June 2025 that over 40% of agentic AI projects would be cancelled by the end of 2027, citing rising costs, unclear business value and weak risk controls. It also warned that many products sold as agents were not, a practice it called agent washing.
• BCG's autumn 2025 study of about 1,250 firms found only 5% getting value from AI at scale, and 60% getting little or none.

Abandoning a pilot is fine. Many should be abandoned. The waste is in abandoning it slowly and learning nothing.

The six lines

Before anyone writes a prompt or signs up for a trial, write these six lines on one page. If you cannot fill one in, that is your first finding.

1. The owner. One named person who answers for the pilot. A committee does not count.

2. The task. One task, described the way the person doing it would describe it. "Draft the weekly client update for our ten largest accounts," not "improve client communication with AI."

3. The baseline. How long the task takes today, and how often it goes wrong. A rough number from a week of sticky notes beats no number.

4. The success test. What would have to be true at the end for you to keep it. Time saved, after counting review time. Errors down, or at least not up. People still using it in week six without being reminded.

5. The stop date and the stop rule. A date, and the condition under which you end it on that date. For example: if review time eats more than half the time saved by week six, we stop.

6. The write-up. Who writes down what was learned, and where it goes.

Here is what the six lines might look like for a small accounting practice:

1. Owner: the practice manager.
2. Task: draft the monthly covering email that goes with each client's management accounts.
3. Baseline: minutes per client today, measured over one week with a timer.
4. Success test: drafting plus checking takes half the baseline or less by week six, no factual errors reach a client, and the two staff who send the emails still choose to use the tool.
5. Stop date and rule: end of week six. If checking alone takes more than the old drafting time, or any client receives a wrong figure, stop.
6. Write-up: the practice manager, one page, saved in the shared memory, within a week of the stop date either way.

Notice that line four counts checking time, and that line five names a failure that ends the pilot on the spot.

Why the sixth line matters most

Line six is where a company memory comes in. A pilot produces knowledge even when it fails. Which tasks the tool handled well. Which facts it needed and did not have. Which rules people had to repeat to it every day. Which risks showed up that nobody predicted.

If that knowledge stays in the heads of the three people who ran the pilot, it leaves when they move on. If it goes into a written, shared memory, the next pilot starts from it. And if the pilot succeeds, the same notes become the rules the tool reads every day: the facts it needs, the rules it must follow, the point where a person approves its work.

We keep ours the same way. Corrections and lessons are written down with the reason they exist and how to apply them, in files our AI sessions read before starting. A lesson that only lives in a chat history is, for practical purposes, lost.

What six lines cannot fix

They will not save a pilot that exists for political reasons. If a senior person wants the pilot to succeed regardless, the stop rule gets quietly moved. Writing it down first makes that move visible, which is the most you can ask of a document.

They also will not choose the right task for you. A well-run pilot on a task that does not matter still produces a well-documented result that does not matter.

If a pilot is already running

Find any AI tool or trial currently running in your business. Write the six lines for it today, even if it started months ago. If line four is blank, set it now and pick a date three weeks out. When the date comes, decide, and write down why.

The AI readiness check is a quick way to see which of your candidate tasks are ready for a pilot at all. Book 20 minutes at pgco.world/book.