We ran an AI pilot and it went nowhere. What did we get wrong?

You ran the pilot. Everyone clapped at the demo. The slide looked clean, the use case was sharp, the room nodded, and someone from the board said the word "transformative" out loud. That was six months ago. Now there is nothing on the P&L (the profit and loss account, the line that decides whether the work paid) to show for the effort, the tool sits in a tab nobody opens, and no one will say so in a meeting. So you are sitting with the quiet version of the question: what did we get wrong?

Here is the short answer, and it should land as relief rather than indictment. You almost certainly got the technology right and the organisation wrong. The pilot worked in the demo because demos are easy; the value never arrived because the work around the tool never changed. That is the single most common reason AI pilots fail, and it is the most fixable. The model was never the problem.

Why do AI pilots fail so consistently?

Start with the number that takes the shame out of it. In 2025, the MIT NANDA initiative (a research programme at the Massachusetts Institute of Technology studying how AI is actually used inside companies) studied 300 public AI deployments, interviewed 150 leaders and surveyed 350 employees, and found that 95% of enterprise generative AI pilots were delivering little to no measurable impact on profit. Only about 5% reached rapid revenue acceleration. So whatever happened in your building happened in nineteen out of twenty buildings. You are not an outlier. You are the rule.

The instructive part is what MIT named as the cause. Not weak models. Not the wrong vendor. They called it a learning gap: the failure to integrate the tool into workflows, processes and culture. The pilot was treated as a technology experiment when the value was always going to come from an operating change. Gartner saw the same shape from a different angle, forecasting that at least 30% of generative AI projects would be abandoned after proof of concept (the trial stage that proves a tool can work before any real rollout) by the end of 2025, citing unclear business value alongside poor data quality and weak controls. As Gartner's Rita Sallam put it, executives are impatient to see returns, yet organisations are struggling to prove and realise value. Impatience plus a vague goal is a combination that ends in silence.

Was it the goal, or the change management?

It was almost certainly both, and they are the same wound. Walk back to the week the pilot launched and ask three questions. Did anyone write down, in commercial terms, what success looked like before a single prompt was typed? Did one named person on the business side own the outcome, rather than the IT side owning the deployment? Was there a plan for how a real team would change the way they actually work? When a pilot dies, the post-mortem nearly always finds the same three blanks: no commercial case, no owner, no change plan. The tool performed. The decision to change the work was never made.

This is why the failure feels so strange. The thing worked, and still nothing happened. That gap between a working tool and a moved number is not a software gap; it is an adoption gap, and adoption is a human and process discipline. The bottleneck is no longer the technology. What matters now is whether the organisation reshapes itself around what the technology makes possible, and that reshaping is leadership work, not engineering work.

What do the 5% who win actually do differently?

They redesign the work, and they refuse to guess at the result. Wharton's Human-AI Research group surveyed enterprise leaders for its 2025 adoption report and found that 82% now use generative AI at least weekly, that around three in four see positive returns, and that 72% formally measure those returns against a business outcome rather than a feeling. That last figure is the quiet differentiator. The leaders capturing value do not run a pilot to see what happens; they name the number they intend to move, rebuild the workflow that moves it, then hold the work to that standard. They aimed at a defined target, then changed the operation to reach the target.

This is the part the dead pilot teaches by its absence. Deloitte's 2024 study of enterprise generative AI found that more than two-thirds of organisations expected 30% or fewer of their experiments to reach full deployment within three to six months, and named the reason plainly: organisational change runs behind technological advancement. Most pilots spend their energy on the model, which demos well, and starve the redesign that actually moves the number. You can buy the model in an afternoon. You earn the new way of working over a deliberate season of effort.

And the encouraging part, the part the dead pilot obscures, is where the readiness already sits. US Federal Reserve researchers, analysing the Census Bureau's Business Trends and Outlook Survey (the government's running pulse-check on what firms are actually doing), found that roughly 41% of workers report some AI use at work while only about 12% use it daily. Read that again. Your people are further ahead than your dashboards reveal, yet the use is shallow, occasional, unembedded. The capability is already in the building. The work now is to give it a goal, an owner and a redesigned process the people can use.

How do we run the next one so it lands?

Treat the next pilot as a change programme with a tool inside it, in this order:

1. Write the commercial case first. Name the number you intend to move, in dollars or hours or conversion, before you choose a tool. If you cannot state it in one sentence, the pilot is a demo waiting to happen.
2. Appoint a business-side owner. One named person who owns the outcome on the P&L, sitting in the function the work belongs to, away from IT. IT enables; the business owns.
3. Redesign the workflow on paper. Map how the task is done today, then how it is done once the tool is in place. That redrawn process is where the value lives. Build it before you scale.
4. Pick a target with growth in it, not only efficiency. The high performers aimed at new value, not just a cheaper version of yesterday. Set the larger goal.
5. Measure on the P&L, and review like an adult. Decide the metric and the date you will judge it against, then keep that appointment in full, whatever it reveals.

There is a quieter point underneath all of this, for the leader who wants it. The reason the redesign gets starved is rarely budget. It is that reshaping how people work is harder, slower and more personally demanding than buying software, and under pressure we default to the easier option. There is a striking study of judges, more than 1,000 parole rulings, where the rate of favourable decisions fell from around 65% at the start of a session toward zero by the end, then returned to 65% after a food break. As decision-makers tire, they reach for the safer, easier choice. Buying the tool is the easy choice. Changing the work is the demanding one. The leaders who win the next decade are the ones who upgrade themselves first, who find the capacity to make the demanding choice when it counts.

About the author

Thomas Green

British technology futurist, AI keynote speaker and advisor. Thirty years across enterprise technology and AI strategy, helping leaders navigate the future of work. The futurist who died.