Why do AI pilots fail to scale?

AI pilots fail to scale because the conditions that made the pilot work are the conditions you cannot reproduce anywhere else in the business. A hand-picked team, a clean sample of data someone prepared by hand, an enthusiastic sponsor watching closely, and no requirement to prove anything in pounds. Take those four supports away and the thing falls over. The model was rarely the problem.

Four structural gaps account for most of it, and all four are visible before you spend a penny. There is no stated baseline, so nobody can say what the pilot beat. There is no named owner with authority outside the pilot team. The systems involved cannot pass data to each other without somebody exporting a spreadsheet. And there is no budget line beyond the pilot, so success arrives at the end of the money. MIT’s 2025 study found 95% of AI pilots deliver zero measurable P&L impact. That number is not a verdict on the technology. It is a verdict on what surrounds it.

The abandonment rate tripled, which tells you this is structural

Before the four gaps, one piece of context that stops this being read as teething trouble. S&P Global Market Intelligence found the share of firms scrapping most of their AI initiatives rose to 42% in 2025 from 17% a year earlier, with the average organisation scrapping 46% of its proofs of concept before production.

If this were a learning curve, the number would be falling. It tripled. More organisations, with more experience and better tools, are killing more projects. That happens when the failure sits in the operating model rather than the technology, because the technology genuinely did improve over that period and the outcomes got worse anyway.

PwC’s 2026 CEO Survey puts the same thing from the top of the house: 56% of CEOs report neither increased revenue nor decreased costs from AI in the last twelve months, and only 12% report both. More than half of chief executives have nothing on either side of the ledger. Not a small return. Nothing.

Gap one: nobody wrote down the baseline

Ask the person who ran your last successful pilot one question. What number did it beat, and where did that number come from?

If the answer takes longer than a sentence, you did not run a pilot. You ran a demo. As one practitioner guide to AI ROI puts it, a pilot that “went well” but cannot state the baseline it beat is a demo, not a pilot, and it will not survive the budget cycle.

This is the most common failure I see, and it is also the cheapest to prevent. The baseline has to be written down before the pilot starts, because afterwards everyone is motivated to pick a flattering comparison. Average handling time before, average handling time after. Invoices processed per person per day. Days from quote to signature. Rework rate. Pick the number, record where it came from, record who agreed it. Five minutes of work in week one.

Without it the pilot enters the budget cycle with a story instead of a figure, and stories lose to figures in every finance meeting I have ever sat in. The team says it went really well. The CFO asks by how much. The room goes quiet. The project does not get killed dramatically, it just does not get renewed.

There is a second effect. If nobody set a baseline, nobody had to define what the pilot was actually for. Teams end up optimising whatever was easy to measure in the tool rather than whatever costs the business money. That is how you get a pilot that improves draft-generation speed in a process where drafting was never the bottleneck.

Gap two: nobody owns it once it leaves the pilot team

A pilot has a sponsor. Production needs an owner. They are different jobs and people conflate them constantly.

The sponsor cares while the pilot runs. The owner has to care for years, has to sit in front of the board when it underperforms, and needs the authority to make another function change how it works. Most pilots have the first and not the second, which is why they succeed in one department and die at the boundary.

Deloitte’s 2026 Global CSO Survey found 65% of chief strategy officers do not own the top strategic decisions in their own function. If the person whose job title is strategy does not own the biggest strategic calls, ask honestly who owns AI outcomes across your business. Not who is interested. Not who talks about it most. Who is accountable for the number when it is missed.

The test is simple. Name the person. If you get a committee, a steering group, or two names joined by “and”, you have a coordination arrangement, not an owner. I have written separately on who owns AI outcomes in a company and why this question is harder to answer honestly than it looks, particularly in businesses between 200 and 2,000 people where function heads have real autonomy and no obligation to each other.

The ownership gap also explains why the same failed pilot gets repeated. Nobody carries the memory. The sponsor moves on, the team disperses, and eighteen months later a new vendor pitches the same use case to a different department, which runs the same pilot and hits the same wall.

Gap three: the data path to production does not exist

This is the one that kills pilots quietly, after everyone has agreed the technology works.

During the pilot, someone prepared the data. They pulled an extract, cleaned the duplicates, mapped the fields, and handed over something usable. It took them a week and nobody logged it as a cost. Production needs that to happen every day, automatically, without that person. If your CRM and your finance system and your operations platform cannot exchange data without a human exporting a CSV, the pilot has no route to production. It has a route to that person doing the export forever, which is not the same thing and is far more expensive than anyone models.

This is why the audit’s first Data question is worded the way it is: if you needed a clean, accurate list of your best 500 customers right now, how confident are you in what you would get? Not whether you could eventually produce one. What lands in your inbox this afternoon. Most leaders know the honest answer instantly, and the flinch is the finding.

The related question in the same pillar is blunter still. Can your key business systems share data with each other without someone manually exporting and importing spreadsheets? Answer that one accurately and you can predict which pilots in your portfolio will scale, before you fund them. I have set out the specific checks worth running in what data readiness checks should you run before adopting AI, and the cost of skipping them in dirty data and AI.

Gap four: the budget ends where the pilot does

Pilots are funded as experiments. Production is funded as an operating cost. Nobody makes that conversion at the right moment, which is before the pilot starts.

What happens instead: the pilot succeeds in month three, the team asks for production funding in month four, the budget cycle does not reopen until month nine, and the pilot infrastructure gets switched off in month five because nobody is paying for the licences. By month nine the team has been reassigned and the enthusiasm has gone.

Production costs are also structurally different from pilot costs, and the difference is rarely modelled. Integration work. Monitoring. Someone to handle exceptions when the output is wrong. Training for people who were not in the pilot group and are not volunteers. Change to the process itself, which is usually the largest line and the one that never appears in a vendor quote.

The honest framing is that a pilot budget should always contain a contingent production budget, with a trigger condition attached to the baseline from gap one. If the pilot beats the baseline by X, this money releases. That single sentence prevents most of the death-by-budget-cycle cases I have watched.

Which brings us to the uncomfortable one. Writer and Workplace Intelligence found that 75% of executives admit their company’s AI strategy is “more for show” than actual internal guidance, and 48% now call AI adoption a massive disappointment, up from 34% the previous year. A strategy that is for show does not produce a production budget, because it was never meant to reach production. It was meant to be mentioned on an earnings call.

Mapping the four gaps to the six pillars

Each failure belongs to a pillar, and that is not a tidy coincidence. It is why the audit is built the way it is.

No stated baseline is a Strategy and Process failure. Process, because you cannot baseline something you have not documented. Strategy, because it means nobody defined what winning looked like in commercial terms.

No named owner is a Governance failure, with a People component. Governance asks who is accountable. People asks whether your team has the capability and appetite to run the thing once the enthusiasts move on.

No data path to production is Data and Technology. The 500-customer question is the Data half. The systems-sharing question is the Technology half. Fail both and nothing scales, regardless of how good the model is.

No budget past the pilot is Strategy and Governance. It is the clearest tell that AI sits outside your planning cycle rather than inside it.

Six pillars, five questions each, thirty questions, 7 minutes. You get a score out of 120, a band for each pillar, and a short synopsis. It is self-reported, which means it is only as honest as you are. That is a feature. The people who score themselves generously are the same people whose next pilot joins the 95%.

Before you fund another one, make the owner state the baseline in a sentence. If they cannot, you have saved yourself a write-off and found something more useful than a pilot: a list of what to fix first.