Why most AI pilots never reach production, and what the studies actually measured.

A demo runs on one laptop, with one person watching, who is forgiving and knows which inputs work. Production runs unattended, on real data, for people who will not tolerate a wrong answer and will not report it either; they will simply stop using it. The demo is the part that is pleasant to build and pleasant to watch, which is why so much budget goes into it. Almost every stalled AI project we see is a good demo that nobody costed the distance from, and the distance is roughly four times the demo.

That distance is now measured. Since we first published this guide, three large studies and two UK government surveys have put numbers on how often the demo becomes the product, and the numbers are worse than most buyers assume. This revision sets them out, with their bases and their limits, and then does the practical part: the seven ways a pilot fails, what to measure before you start one, and a worked plan from pilot to production priced against published figures.

One caution before the statistics. The headline numbers come from studies of large American enterprises, and a twenty-person UK firm is not a Fortune 500 company: it has fewer systems to integrate, less data to clean, and no innovation budget to protect a pilot that is not working. Some of that makes production easier to reach, and some harder. Where the UK data differs, and it does, we say so.

Five figures are quoted everywhere on this subject, usually without their bases. Here they are with the bases attached, because a 95 per cent failure rate for one thing and a 12 per cent success rate for another are not the same claim. The first two rows describe large, mostly American enterprises and are the source of most of the headlines; the last three are UK government and statutory surveys, and they are the rows a UK buyer should weigh most.

Read on the source pages on 14 September 2026. The MIT and Gartner figures describe large enterprises and are not UK-specific; the DSIT, PwC and ONS figures are UK.

The MIT report is the one worth reading in full, because it separates two things the press coverage merged. General-purpose tools such as ChatGPT and Copilot are widely adopted: over 80 per cent of the organisations studied had explored or piloted them and nearly 40 per cent reported deployment. The report’s point is that these tools improve individual productivity, not the profit and loss account.

Enterprise-grade systems, custom or vendor-built, are a different story. Sixty per cent of organisations evaluated such tools, 20 per cent reached pilot stage, and 5 per cent reached production. The reasons the report gives are brittle workflows, a lack of contextual learning, and misalignment with day-to-day operations; it calls the underlying problem a learning gap, tools that do not learn, integrate poorly, or match workflows.

Two further findings matter for a small UK firm deciding who should build. In MIT’s sample, external partnerships with learning-capable, customised tools reached deployment about 67 per cent of the time, against about 33 per cent for internally built tools, a difference the authors say was consistent across interviewees though self-reported. And although half of AI budgets went to sales and marketing, the most dramatic cost savings the researchers documented came from back-office automation, with faster payback and clearer cost reductions.

The UK evidence is more encouraging than the American headlines, provided you read the right line. DSIT’s 3,500-business study found that three quarters of adopters, 75 per cent, reported improved workforce productivity. But 77 per cent had not yet seen a change in revenue, and just 12 per cent reported an increase. PwC’s survey of UK chief executives in January 2026 found the same shape at the top of the organisation: 21 per cent said AI had increased revenue in the previous twelve months, 74 per cent reported little or no change in revenue, and 60 per cent little or no change in costs. Only 45 per cent said their organisation had a clearly defined roadmap for AI.

Gartner’s own survey of 822 business leaders, run in late 2023, found early adopters reporting on average a 15.8 per cent revenue increase, 15.2 per cent cost savings and a 22.6 per cent productivity improvement. Those are averages among people who had already got something working, which is the point: the returns exist, they are concentrated in the minority that reached production, and the distribution is what a buyer should plan around.

The ONS adds the depth. Of the 35 per cent of businesses with ten or more staff that use AI, only one in ten uses it extensively. Most UK adoption is a tool in one person’s browser, not a system in a process. A pilot that stays a tool has not failed; it has simply not started the journey this guide is about.

The four blockers in our original guide still bite in the order we gave them. The published studies add three more, and for a small UK company each of the seven has a particular shape that the enterprise literature misses: fewer systems but older ones, less data but messier, and an owner who is usually the founder. Together they are the checklist we run on any stalled pilot, and most stalled pilots fail on at least three of them at once.

A pilot without a baseline cannot succeed, because nobody can tell whether it did. These are the numbers to write down before the first line of code, and the reason each one matters. They take a week to gather, most of it spent timing the task as it is done today, and they are the same numbers the production business case will need, so nothing is wasted. If the supplier cannot help you fill in this table in the scoping session, that is the answer to whether they should build it.

An illustration, not a client account. A 25-person distributor keys about 2,000 supplier invoices a month into its accounts package. Timed over a week, each takes about six minutes, so the task is roughly 200 hours a month. At a loaded cost of £15 an hour, an assumption you should replace with your own, that is £3,000 a month, or £36,000 a year. That baseline is the whole business case; everything else is arithmetic.

The pilot is an audit. On our ladder that is £2,500 to £6,000, fixed, one to two weeks, credited against any build. It tests extraction on 200 real invoices including the scanned and the handwritten, agrees an accuracy threshold (say 98 per cent of line totals correct, with everything below a confidence score routed to a person), confirms the accounts package has an API, and prices the running cost. Suppose the audit concludes that two thirds of invoices can be posted without a person touching them.

Production is a single-process build at £4,000 to £10,000. The running cost, at today’s published prices, is small enough to state. Reading 2,000 invoices a month at roughly 3,000 input tokens and 500 output tokens each is 6 million input and 1 million output tokens. On OpenAI’s gpt-5-mini ($0.25 and $2.00 per million) that is $3.50 a month; on Anthropic’s Sonnet 5 ($2 and $10) it is $22; even on the most expensive model on any of the three pricing pages, Anthropic’s Fable 5.1 at $10 and $50, it is $110. Hosting and document storage will cost more than the model.

Against a saving of about £2,000 a month from the two thirds automated, a build at the midpoint of the band pays back in roughly four months, and the year-one total, audit credited, is about £8,000 plus a few hundred pounds to run. The point of the arithmetic is not the answer, which will differ for every business, but that every input to it was known before anything was built, so the decision to proceed was a decision rather than a hope.

Do not restart it. Scope the production version specifically: which system it must integrate with, what it costs per user or per item, what accuracy is acceptable and what happens when it falls short, and who owns it afterwards. Fill in the table above for the pilot you already have; most of the blanks will explain why it stopped, and the exercise takes an afternoon rather than a budget.

Then treat the pilot as what it was, a sales tool that won the decision internally, rather than as a foundation to build on. Rebuilding the working part is usually faster than adapting a prototype that was never meant to carry load. MIT’s 5 per cent are not the organisations with the best demos; they are the ones that treated the demo as the start of the costing rather than the end of it.

Every figure above was read on its source page on 14 September 2026. The MIT NANDA report was read in the published PDF; its authors describe the pilot and production figures as directionally accurate and based on interviews rather than company reporting, and we have quoted them with that caveat. Gartner’s figures are from its 29 July 2024 press release, including the 822-leader survey it cites. DSIT’s figures are from the AI Adoption Research report published 13 February 2026, with fieldwork from February to May 2025. PwC’s are from the UK pages of its 29th CEO Survey, January 2026. ONS figures are from the July 2026 article on AI in UK businesses.

Model prices are from the OpenAI and Anthropic pricing pages on the day of writing and are in US dollars excluding VAT. The worked example uses stated assumptions, marked as such in the text, and our own published price ladder; it is not a client account and the hourly cost in it should be replaced with your own. We will revise this page as the sources change, and the date at the top is the date of the last check.