Before the AI pilot, write down when you’ll stop

Some AI pilots never quite end: they don’t prove themselves, nobody switches them off, and they keep taking attention. Many others are stopped only after proof of concept. In S&P Global Market Intelligence’s 2025 survey of 1,006 IT and business professionals in North America and Europe, 42% of companies had abandoned most of their AI initiatives before production, up from 17% a year earlier, and on average 46% of projects were scrapped between proof of concept and broad adoption.¹

Stopping a project that doesn’t work is healthy. The waste is in how long it takes to find out, and that is decided before the pilot starts.

Why pilots drift

• There is no baseline. Nobody measured the task before, so nobody can say whether the pilot is better.
• Success is defined afterwards. When the target is set after the results arrive, almost any result can be called promising.
• The scope grows. A pilot for one task becomes a platform for ten, and none of them is finished.
• Costs are measured in the demo. Model usage, checking time and monitoring look small at ten cases a day and very different at a thousand.
• Nobody owns the decision. The pilot team wants to continue, the business is unsure, and the date to decide never arrives.

Write four things down before you start.

1. The baseline. Today’s numbers for the task (volume, time per case, error rate, cost), measured the way you will measure the pilot.
2. The success line. The result that justifies the next step, in the same units as the baseline.
3. The stop line. The result that ends the pilot, and the date by which you will decide either way.
4. The cost at volume. Model usage, the time people spend checking, monitoring, and what it would cost to switch provider or switch it off.

None of this takes long. It usually fits on one page, and it makes the end of the pilot a reading of the numbers.

An example: orders by email

An order desk receives about 300 orders a week by email and types each one into the ERP. Measured on a sample, entry takes 6 minutes per order, and 3% of orders contain an error. In the pilot, AI reads each email, rules check the order, and a person confirms it before it reaches the ERP.

Agreed before the start, and read on 500 real orders: continue if entry takes 2 minutes or less per order, including the check, with errors at 1% or below. Stop if errors rise above today’s 3%, or if the average order takes longer than today’s 6 minutes. Decide at the end of week four, whichever way the numbers point. The sample pilot sheet shows this case on one page.

The figures are illustrative. What matters is that both lines and the date are fixed before the start.

Cash and capacity

Time a pilot saves is capacity: hours freed for other work. Hours become cash only when spending falls, through less overtime, fewer agency staff or a contract that ends. Keep the two in separate columns of the business case, and write down what the freed hours will be used for. Both columns can justify the next step, as long as nobody presents the capacity column as cash.

What the surveys suggest

In McKinsey’s survey of 1,719 respondents in 97 countries, published on 25 August 2026, the AI high performers (6% of the sample: they attribute at least 5% of EBIT to AI and report significant value from it) were twice as likely as others to report defined processes for measuring the impact of AI, and nearly three-quarters of them had fundamentally redesigned workflows, against about a quarter of other respondents.²

Surveys like this show what successful companies have in common. They can’t show what caused the success. They do suggest that results and measurement go together: the companies with results know what they are measuring.

A forecast, and what followed

On 29 July 2024, Gartner predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, citing poor data quality, inadequate risk controls, escalating costs or unclear business value.³ On 26 January 2026, Gartner reported that at least 50% had been abandoned after proof of concept by the end of 2025, and set out five failure points: lack of business value, data that isn’t ready, escalating total cost of ownership, responsible AI treated as an afterthought, and poor change management.⁴

A baseline, a stop line and an honest cost at volume are designed to expose three of those problems early: data that isn’t ready, rising costs and unclear value.

When the answer is no

A pilot that stops at its rule has produced a result. You know that the task, the data or the process isn’t ready, and roughly why. Write that down as carefully as a success: it is usually the starting point for the next, better-chosen pilot.