Your AI pilot needs a kill date

Are you paying for AI tools that nobody in your business could tell you the return on?

Not tools that failed. Tools that are still running.

A pilot that started with some energy, a subscription that renews every month, a dashboard someone checks when they remember. Ask what it has earned and you get a story about how promising it looks.

It's tempting to blame the technology. The models are too new, the tools too immature, your team not ready.

That explanation feels fair. It is also the one that lets everyone off the hook.

An MIT report published in August 2025 put a number on how common this is. Researchers interviewed fifty-two organisations, surveyed a hundred and fifty-three senior leaders and reviewed more than three hundred public AI initiatives. Their conclusion was that despite $30–40 billion in enterprise spending, ninety-five percent of organisations were getting zero return.

The sample was modest and success was self-reported, so treat the exact figure with care. The direction is hard to argue with.

Most businesses I walk into have at least one of these running right now.

Treat it like a capital decision

The fix is a better approval, not a better tool.

Think about how you sign off a fitout or a refurbishment. Nobody approves one because it looks exciting.

You want to know what it costs, what it earns, how long before it pays for itself, and whose job it is to make sure it does. If the numbers don't stack up, it doesn't get built.

AI gets approved the other way round. It is cheap to start, it feels urgent, and nobody wants to be the owner who missed it.

So it goes in as an experiment. Experiments don't have payback periods. They don't have owners either — they have enthusiasts.

The MIT researchers blame tools that don't learn and don't fit the way people actually work. That is true. It is also the kind of problem an owner would have caught in month two, if there had been one.

So before any AI project starts, it gets the same four lines a capital request would:

• One named outcome. A result a customer or a bank statement would notice. Hours saved on quotes, days off debtor collection, complaints resolved on first contact. Not "efficiency".
• One owner. A person, not a team. Their name is on it and they report on it.
• A payback date. The date by which the outcome should have covered the cost, including the hours your people spent setting it up.
• A kill date. The date it gets switched off if the outcome hasn't shown up. Written down before it starts, while nobody is attached to it yet.

The honest objection is that some AI value is hard to measure, and a strict payback test will kill a few projects that would have come good. That is a real cost. It is a much smaller one than a dozen pilots that never end.

A pilot with a kill date is a decision. A pilot without one is a subscription.

Klarna measured the wrong thing

In February 2024, Klarna announced that its AI assistant was handling two-thirds of its customer service chats in its first month. The company said it was doing the equivalent work of seven hundred agents. Chief executive Sebastian Siemiatkowski became one of the loudest voices for AI replacing service staff.

By May 2025 the tone had changed. Speaking to Bloomberg, Siemiatkowski said that cost "seems to have been a too predominant evaluation factor", and that the result was lower quality. Klarna began recruiting human agents again, promising customers they could always reach a person.

That was not AI failing. The assistant did what it was pointed at. It took cost out.

The problem is that cost was the target, and nobody had written down what a good outcome for the customer looked like, or who was responsible for protecting it.

Klarna didn't have a technology problem. It had a specification problem. The payback was measured in headcount, and the customer paid for it.

Why I do it this way

I use this test, and I have for years.

It came from running my own properties. As an owner-operator of serviced apartments and a venue, I carried the full profit and loss.

Every capital spend had to earn its place against a payback — a refurbishment, a new system, a piece of equipment. If I couldn't say what it would return and by when, it waited.

That discipline doesn't care whether the thing being bought is a carpet or a piece of software. Money leaving the business is money leaving the business. The question is always the same: what does it earn, who makes sure it does, and when do we admit it hasn't.

The kill date is the part people resist. It feels pessimistic to plan the funeral on day one.

In practice it does the opposite. The owner works harder on something with a deadline, and the ones that don't earn their keep end quietly instead of dragging on for a year.

Do this before Friday

List every AI tool and pilot your business is paying for right now, including the ones on someone's personal card. Next to each one, write the four lines: outcome, owner, payback date, kill date.

Anything you can't fill in gets a kill date anyway — the last day of this month.

If nobody steps forward to own it by then, switch it off. You were never running a pilot. You were paying for a hope.