Why AI returns stall, and how to set AI up properly

The return problem

Can you name the one number your last AI project moved?

A pilot, sometimes called a proof of concept, is a small trial to test whether an idea works before committing to it. Generative AI is software that produces text, images or code in response to a request. Most businesses have run at least one trial of this kind. The independent evidence on what came next is sobering.

MIT's NANDA initiative published a report in July 2025 finding that 95% of the organisations it studied saw no measurable profit and loss return from their generative AI pilots. It was early, preliminary research based on interviews, a survey and a review of public deployments, and its method has been debated. I treat it as a signal, not a precise rate.

Gartner predicted in July 2024 that at least 30% of generative AI projects would be abandoned after the proof of concept stage by the end of 2025, citing poor data quality, inadequate risk controls, escalating costs or unclear business value. In October 2024, a Boston Consulting Group survey of 1,000 senior executives in 59 countries found that 74% had yet to show tangible value from AI. In June 2025 Gartner added a forecast that over 40% of agentic AI projects would be cancelled by the end of 2027. Agentic means software that can act, not only write.

Where the value is

I have spent my career in regulated, operations-heavy businesses, where a process either works every time or someone writes a report about why it did not. That shapes my view. Once AI is set up properly, the potential is large. It can take routine work off skilled people, move work that queues, and let a business respond to change without hiring for every new demand.

The businesses that struggle tend to have the same gaps. Nobody owns the result. The job is vague. There was no measure before the trial, so there is nothing to compare with afterwards. Controls arrive late. This is a view from experience and not a research finding.

That is why I describe the gap as one of management. Technology has moved quickly, and the models are capable. The harder part is deciding what you want from them and who answers for it. In my experience that conversation is easier to have before a trial starts than after it has run for six months and nobody can say whether it worked.

Five steps to set AI up properly

This is the frame I would use for any single use of AI. Start with one, not ten. A single well-run use gives you a pattern you can repeat.

1. Name an owner. One accountable person, with authority over the process and the outcome, not only the tool.
2. Define the job. Describe one task with a clear start and finish, such as preparing a first draft for a person to review. Avoid goals like 'use AI in operations'.
3. Measure a baseline. Record how the work performs today: cost, time, errors and volumes. Without it you cannot show change.
4. Set controls and human oversight. Decide what the AI may and may not do, which data it may see, and who approves consequential actions. Where personal data is involved, I also ask where it goes. As GenWA's data protection officer, that question is never far from my mind.
5. Decide in advance. Agree the review date and the conditions for scaling, adjusting or stopping. Stopping a weak trial is a good result.

Why the order matters

The sequence is deliberate. An owner without a defined job has nothing to be accountable for. A job without a baseline cannot be shown to have improved. Controls added after the fact cost more, and tend to arrive after the first problem.

In regulated sectors, this is simply how change is introduced. A new process has an owner, a procedure, a test and an approval before it goes live. AI deserves the same treatment. Done this way, a result can be believed by a finance director, a regulator or a customer, and that is worth the extra care.

What leaders should measure

Measure the outcome the business cares about, not the activity of the tool. The number of prompts typed is not a return. I would track these:

• Cost per completed task, including AI spend and the time people spend reviewing.
• Time from start to finish, against your baseline.
• Errors and rework, including what a person had to correct.
• The business outcome the process exists for, such as backlog cleared, complaints reduced or cash collected.
• Use by the people the tool was meant for, and total AI spend against budget.

How GenWA helps

Analyse, delivered through our advisory service, starts with where you are. The AI health check is a short questionnaire, and our advisers, human and agentic, review how AI is used, what it costs and where it is earning its place. From there our advisory work helps you set the owner, the job, the measure and the controls. Automate puts AI to work on defined tasks within those controls.

I will not give you a return figure in advance. It depends on your processes, and any figure offered before looking at them would be a guess.

What to do on Monday

Five things to ask in your next leadership meeting.

• Pick one AI use in the business and name its accountable owner.
• Write the job in one sentence, with a start and a finish.
• Ask for the baseline: what did this work cost, and how long did it take, before AI?
• Check who approves consequential actions and where personal data goes.
• Set a review date, with the conditions for scaling, adjusting or stopping written down.