Why AI pilots don't pay back

Ask a leadership team how their AI pilots went and most will say they went well. The model did what it was asked to do. The people who tried it liked it.

Ask what the business got for it and the answer takes longer.

That gap is common. In BCG’s 2026 CEO survey, most companies were running AI pilots, but only about a quarter had changed the wider business around them. The pilots succeed at something that was never connected to a result.

In my experience there are three reasons, and they tend to arrive together.

Nobody made an investment decision

A pilot usually starts because a team had an idea and a tool was available. That is a fine way to learn. It is a poor way to allocate money.

An investment decision looks different. It says what should improve, by how much, what has to be true for that to happen, and how anyone will know. It also says what the business is choosing not to do, which is the part most often missing.

Without that, a pilot has no test to pass. It can only be judged on whether it ran. So it is declared a success, and then nothing follows, because success was never defined as something the business would act on.

The fix is to agree, before a pilot starts, the measure it will be judged by and the decision that follows. There are only three: proceed, adapt or stop. If none of those would change what you do next, the pilot is not worth running.

The work around it didn’t change

This is the largest of the three. A pilot adds a capability to a process that was designed without it. The process, the roles and the decisions stay as they were.

Take a team that handles customer queries. An assistant drafts replies, and drafting gets faster. But the agent still reads every case, still checks the same systems, and still waits for the same approval. The slow steps were never the drafting.

Automating one step of a poor process preserves the rest of it. The return comes from redesigning the work: what should stop, what stays with a person, what AI assists, and what AI can take on within agreed limits. That is an organisational change, and a pilot cannot make it.

The fix is to treat the workflow as the unit of change. Redesign it first, then decide what to automate.

No one owns the result

Pilots tend to belong to the people who ran them: an innovation team, an enthusiastic manager, a vendor. The benefit, if it comes, belongs to a different budget.

So when the pilot ends there is nobody whose job it is to carry it into everyday use, find the money, retrain the team and answer for the outcome. It stays where it started.

The fix is to name an owner for the business result before the pilot begins. That person should not be the one running the pilot.

The saving that wasn’t

There is a fourth problem, and it hides inside business cases that look healthy.

Many AI cases rest on time saved. Ten minutes a case, thousands of cases a month, multiplied by a salary. The number is large and it is not a saving.

Time saved creates capacity. It becomes a cash saving only when spending is reduced or avoided. Until then the hours have moved, and the cost is the same.

Capacity is worth having. It can mean shorter queues, better service or growth without hiring. But it is a different kind of value from cost reduction, and a business case that blurs the two will be found out at the first budget review.

I keep four kinds of value in separate columns: revenue, cost savings, released capacity and customer outcomes. I also record how firm each estimate is. A board should be able to tell a hypothesis from a validated case, because they deserve different amounts of money.

What a pilot that pays back looks like

It is smaller than most, and stricter.

• It tests the riskiest assumption in the investment case, not the easiest thing to demonstrate.
• It is time-boxed, with success measures agreed in advance.
• It sits inside a workflow that has been redesigned, at least on paper.
• It has an owner for the result, who is not the person running the pilot.
• It ends in a decision: proceed, adapt or stop.

Stopping counts, because a pilot that shows an idea is not worth scaling has saved the cost of scaling it.

Where to start if this is you

If you have several pilots and no clear return, do not start another. Look at what is running and give each one a verdict against the outcome it was meant to produce. Some will proceed. Some need the work around them redesigned. Some should stop.

That review takes about a week when someone independent does it. It is what the AI Advantage Diagnostic is for, and it is the first step in recovering a programme that has stalled.