When to kill an AI project: the four kill signals

A stalled AI tool rarely gets stopped. It gets quietly replaced by the next one, which is how a firm ends up paying for two and using neither. So give the current one a named owner, a few hours and a date. On that date, either their work has changed and you keep it, or it has not and you cancel it before you buy anything else.

That pattern is what we see most often, and the instruction above is the whole decision. The rest of this piece is how to make it honestly, including the three things that make it harder than it sounds: you probably have no record of how long the work took before, stopping may cost you the whole remaining term anyway, and the hours needed to rescue a tool are usually worth more than the licence.

This is a 2026 question in a way it was not two years ago. Office for National Statistics figures released in July 2026, drawn from 38,637 businesses, show self-reported AI use among UK businesses with ten or more employees rising from around 12% to around 35% since late 2023. Over the same period the average number of AI technologies used per adopting business moved only from around 1.4 to around 1.6. Firms are not, on the whole, still deciding whether to try something. They are sitting on one or two things they already tried and a decision nobody has made.

Firms mostly do not decide to abandon an AI project

A firm just stops mentioning the tool.

Deloitte, surveying 3,235 business and technology leaders across 24 countries for its 2026 State of AI in the Enterprise report, describes the mechanism plainly. Firms keep funding new pilots, it says, because those are "relatively low cost and lower risk, rather than facing the harder work of scaling up existing successes". The new thing is exciting, and nobody has to say the old one failed.

Deloitte's announcement of the same report records that only 25% of respondents had moved 40% or more of their AI experiments into production, while 54% expected to reach that level within three to six months. Read those two numbers together and you get a fair picture of most firms: a lot of things started, few finished, and a persistent belief that finishing is a couple of quarters away.

You will recognise the local version. Someone championed a tool in March. Three people had a login. One of them still uses it occasionally. The subscription renews on a card nobody looks at, and when the topic comes up the answer is that it never really got going.

That is not a failed experiment. An experiment has an end. This is a subscription with a story attached.

What "stalled" actually means

It is worth being precise here, because three different problems get the same label and they do not have the same answer.

Bought and never opened. Licences were issued, and nobody logged in past the first week. This is not a verdict on the tool. Nothing was tested.

Used, but nothing downstream moved. People log in. Work still takes the same time, goes through the same hands, and comes out the same shape. Deloitte has a figure for this specific condition: 37% of the leaders it surveyed reported using AI "at a surface level with little or no change to underlying business processes".

Used, and actively worse. In our experience this is the rarest of the three, and it tends to show up as rework. The draft arrives faster and then somebody spends longer fixing it than writing it would have taken.

Only the second and third are evidence about the tool. The first is evidence about your firm, and buying a different product will produce the same result.

The four kill signals

Score your project. Each signal that is present scores one point. Two points or more means you kill it or reshape it this quarter.

0 points: it is working. Leave it alone and go and look at something else.

1 point: fixable, usually. Signal 2 on its own is the most recoverable, and in our experience the most common, because it means the thing was never actually tried.

2 points or more: you stop it this quarter, or you reshape it once with a named owner and a new date. Reshape it once only. If you want the diagnosis behind these signals, we have written separately on why AI pilots stall before they scale.

Two notes on using the table honestly. If you cannot answer a signal because nobody knows, score it as present: not knowing whether anyone was ever set up on a tool is itself the finding. That rule covers ignorance, not impatience, so it does not apply to signal 1 if the ninety days simply have not run yet. And if the subscription costs less than about fifty pounds a month, do not run this process at all. Cancel it, diary the renewal, and spend the twenty minutes on something with a bigger number attached.

Ninety days, in signal 1, is our rule and not a researched figure. We should say so plainly: we could not find an organisation that publishes how long a dead AI pilot ran before anyone stopped it. What is published is Deloitte's observation that a use case estimated at three months can stretch to eighteen once integration work starts, alongside the three-to-six-month expectation quoted earlier. Ninety days is long enough to be fair and short enough that the answer still matters. Count it from the day someone was properly set up on the tool, not from the day you bought it, because the gap between those two dates is where, in our experience, most projects quietly die.

Before you cancel, check whether the project was ever tried

Do not skip this step. It is how a good idea gets thrown out alongside a product that nobody was ever properly given the time to use.

The Department for Science, Innovation and Technology surveyed 3,500 UK businesses for its AI Adoption Research, published in January 2026. Among businesses already using AI, the things getting in the way were, in order: limited AI skills, expertise or knowledge, cited by 54%; a lack of tools or platforms, 37%; no identified use for AI in the organisation, 30%; and complexity of integration, 26%. Vendor lock-in barely registered, at 6%.

So the most common obstacle is not that the idea was wrong. It is that nobody was equipped to carry it out. Before you cancel anything, answer three questions in writing:

1. Who owned it? A name, not a team. If there is no name, it was never tried.
2. How many hours did they actually get? Not hours they were told to find. Hours protected in a diary against a fee-earning alternative.
3. What were they trained on? An hour of supplier demo is not training. Being shown how to do one specific recurring task in their own workflow is.

If all three answers are thin, you have not tested the idea. Reshape it once: one owner, a few protected hours a week, one named task, ninety days. Set it up properly this time, the way we run a tool trial. Do not buy a different product to solve a staffing problem, because you would give the next product the same three answers.

Price those hours before you commit to them. A half day a week for a quarter from a fee earner is roughly a tenth of their chargeable time. Set that against the licence: DSIT put the median UK AI spend at £2,000 among firms able to give a figure, so for most small firms the hours are the larger number by some distance. That comparison is the real decision, and it often argues for cancelling something that could probably be rescued. Our note on working out what an AI tool actually returns sets out how to put a figure on both sides. Rescuing it has to be worth more than the hours it takes.

If all three answers are solid and the work still looks the same, that is your verdict, and it is a real one. Cancel it.

How to judge when you have no baseline

The difficulty is this. To judge whether work changed, you want to know what it looked like before, and in our experience most firms never recorded it. We rarely meet a practice that knows how long a set of accounts took last March.

You cannot recover a baseline you never took. So do not pretend to measure. Ask the owner four questions instead, and take the answers as they come.

• What did you stop doing? Not what got faster. What has actually left your week.
• What would you have to go back to on Monday if the licence was cancelled tonight? If the answer is "nothing much", you have your finding.
• Has anything left your desk earlier than it used to? A specific thing, named, with a rough sense of how much earlier.
• Would you pay for this out of your own budget? Asked of the person using it, this is the most reliable single question we know of.

None of that is measurement. It is testimony, and it is what you have. It is also considerably better than a business case built backwards to justify a renewal.

One warning about judging this way. Value does not always land on one person. Twelve people saving twenty minutes a week adds up to real money, yet no individual's week visibly changes. If your firm's benefit is spread thin, ask the four questions of three or four people rather than one, and treat a consistent shrug from all of them as the answer.

And beware the seasonal trap. Firms whose work peaks at month end, year end or a filing deadline should judge the tool against that peak, not against a quiet Tuesday in August. A tool that does nothing for eleven months and saves a week in January is worth keeping.

That cuts against the ninety-day clock, and the clock should give way. If your peak falls inside the window, judge it at the peak even if that means waiting. If the peak is months off, do not hold a dead tool open until January to be fair to it: score the other three signals and decide on those.

Stopping cleanly costs money, and the amount is knowable

Cancelling is not free, and the terms are harsher than the small firms we work with tend to expect. Two catch the small firms we work with repeatedly, and both are worth checking before you decide.

The annual plan you chose because it was cheaper. Google's own admin help sets out the trade-off directly. On the Flexible Plan, "You cancel your subscription at any time without penalty". On the Annual/Fixed-Term Plan, "If you cancel your subscription before the renewal date, you're charged for the remaining balance of your contract and no refunds are issued", and you "can't reduce the number of licenses until it's time to renew". Google also states, in the same comparison, that the annual plan "offers the lowest per user per month price". The discount that made it affordable is the same term that keeps you paying until renewal. You can switch plans, but only during a trial or at renewal, which means the decision you are living with was taken a year ago.

Microsoft's refund window, which is seven days. Microsoft's published policy for business subscriptions states: "You can only cancel and receive a prorated credit or refund if you cancel within seven days after the start or renewal of your subscription." After that, turning off recurring billing "prevents your subscription from renewing at the end of its term", and "You keep access to your products and services for the remainder of your subscription". Eligibility varies with how you bought it, so check your own purchase route rather than assuming this applies.

That gives you a number, and the number changes the decision. If the licence runs to March and you cannot get a penny back, the choice is not "cancel or continue". It is whether to spend a fee-earner's protected hours reshaping something you have already paid for. Sometimes the answer is yes, precisely because the money is gone either way. Make that trade explicitly rather than drifting into a renewal.

Get your work out before you press cancel

The clock starts when you cancel, not when you decide. Microsoft's guidance is specific: "When the cancellation becomes effective, your users lose access to their data", and "Any customer data that you leave behind might be deleted after 30 days, and is deleted no later than 180 days after cancellation." Some cancellation routes are faster still, with SharePoint and OneDrive content "deleted immediately".

Note also what your data protection rights do and do not cover. Under the UK General Data Protection Regulation, a processor contract must provide that the processor will, at the controller's choice, delete or return all the personal data it has been processing. That is a right over personal data. It is not a right to your prompts, your configuration, your custom instructions or your outputs. Those are governed by the supplier's terms, and if you want them, you export them yourself before the account closes.

So, in order:

1. Export anything you would miss. Prompts, custom instructions, saved templates, chat history if it holds working knowledge, any documents living only inside the tool.
2. Write down the configuration that took effort to get right. It is easy to lose, easy to reuse, and in our experience standard exports leave it behind.
3. Note the integrations that will break, and who will notice on the Monday.
4. Then cancel, or turn off recurring billing if the term runs on regardless, and diary the renewal date so nothing renews by accident.

Keep the lesson, and say it out loud

The point of stopping is to be able to start again well. Two things are worth writing down while they are fresh, in a paragraph, not a document.

What you learned about the task, not the tool. Often the discovery is that the job you tried to automate was not one job. Camden Council's public record on the government's Algorithmic Transparency Recording Standard shows the same separation in the open: one supplier's product was trialled and "has been discontinued", while the same underlying use case was carried forward into a fresh, deliberately time-boxed six-month pilot, with procurement held back until it was proven. The product was dropped. The problem was kept.

What you learned about your firm. Whether anyone had protected hours. Whether training happened. Whether the owner had authority to change the process or only to use the software. That knowledge is what makes the next attempt cheaper, and it is the part that gets lost when a project ends by going quiet.

There is a reason to be relaxed about restarting. Most firms buy these tools rather than build them. DSIT found that 71% of businesses using text generation had bought ready-to-use software rather than developing it. The sums are small, too. DSIT found that 27% of current AI users reported spending nothing at all in the year it measured. One of the small firms in that study put it plainly: "With any software development there will be fairly significant cost, whereas if you buy something off the shelf, you can pick it up and drop it."

That is the actual reason to stop cleanly. Not that AI does not work, but that the next attempt is easy to start and easy to drop, so nothing is served by keeping a dead one on the books.

Pause or kill?

A pause is legitimate in exactly two situations, and wanting to avoid an awkward conversation is not one of them.

Pause when the blocker is dated and external. The system you are waiting for is released in November. The person who would own it is back in six weeks. Write the date down and set a reminder. A pause without a date is a kill nobody wanted to say out loud.

Pause when the licence is paid for and reshaping it is cheap. If you have already paid to March, one protected half day a week from a named owner may still be the best thing to do with a licence you cannot get a refund on.

Kill in every other case. In particular, kill when the argument for continuing is that you have already spent the money, when the champion has left, or when the only remaining case is that stopping would look like an admission of failure. In our experience firms are rarely judged for stopping something. They are judged for the second subscription.

The one-page stop script

Five messages. Send the first four, in this order, on the day you decide, and write the fifth into your own diary.

1. To the people who actually used it, which may be two of them rather than the whole firm: "We are stopping [tool] on [date]. It did not change how [task] gets done, and that is a result, not a failure. Here is what we learned."
2. To the owner: "Thank you for running it. Please write a paragraph on what you would do differently, and export anything worth keeping by [date minus one week]."
3. To the supplier: "We are not renewing on [date]. Please confirm the cancellation terms, the data export options, and the deletion timetable in writing."
4. To whoever holds the card, or a note to yourself if that is you: "Cancel the mandate for [tool] after [date], and check the statement again a month later in case a charge still lands."
5. To yourself, in the diary: a note ninety days after cancellation, saying what you would need to see before trying this again.

The fifth line is the one that matters most. It is the difference between stopping a project and forgetting one.

FAQ

How long should we give an AI tool before deciding? Ninety days from the day someone was properly set up on it, counted from setup rather than from purchase. That is our rule, not a researched standard.

We are locked into an annual contract. Should we still stop using it? Probably yes, and decide it separately from the money. Continuing to half-use something because you paid for it costs attention as well as licence fees. If the term runs on regardless, the real question is whether one protected half day a week could make it work, and that is worth asking once, with a name and a date attached.

Is it not wasteful to cancel something we spent months on? The months are gone either way. The waste is the next twelve months of a subscription nobody uses, plus the replacement bought without understanding why the first one failed. Stopping is what makes the next attempt affordable.

How do we know whether the tool or the idea was wrong? Ask who owned it, how many protected hours they got and what they were trained on. If those answers are thin, the idea was never tested, so the tool is not what failed. If they are solid and nothing changed, the tool or the task is the problem, and you now know something useful about both.

What if nobody wants to be the one who kills it? That reluctance is itself evidence, not an obstacle. Put the decision on a date, in advance, before anyone is invested in the answer. A decision with a diary entry against it is much easier to make than one that requires somebody to raise it.

---

If you are carrying a project you suspect is dead and you would rather think it through with someone who has no stake in the answer, book a conversation.