AI project failure may get worse
before it gets better

The numbers do not form one clean curve

The advertisement that prompted this article used a powerful line: only 48% of AI projects make it to production. Will yours? The statistic is real, but it needs context. Gartner reported in May 2024 that, on average, 48% of AI projects reached production and that the journey from prototype to production took eight months.¹

That result is best treated as a production-conversion signal, not as a universal failure rate. A project can reach production and still fail to deliver value. Another can be stopped before production because the pilot successfully disproved the case for continuing.

S&P Global supplied the clearest directional evidence of deterioration. In its 2025 Voice of the Enterprise research, the proportion of companies abandoning most of their AI initiatives before production rose from 17% to 42% year on year. The average organisation reported scrapping 46% of proof-of-concept projects before production.²

The 2026 picture remains difficult. S&P Global reported that only 37% of AI initiatives launched in the previous year were classified as live and delivering value, while 46% were considered on track to achieve positive return on investment within twelve months.³

These figures cannot be joined into a precise statistical time series. They use different samples, definitions and units of analysis. What they do support is a more modest conclusion: organisations are running a great many AI experiments, and the passage from experiment to durable value remains unusually difficult.

On a narrow screen, each row is shown as a separate card.

The adoption shock can make failure rise while AI improves

S&P Global explicitly connects the rise in abandoned projects with organisations trying to deliver generative AI at pace. RAND researchers, after interviewing 65 experienced data scientists and engineers, described managers and directors as facing intense pressure to do something with AI simply to show that they were keeping up.⁴

That is close to the FOMO explanation, although the surveys do not isolate fear of missing out as a causal variable. The defensible claim is that project volume and executive urgency expanded quickly, while organisational readiness did not expand at the same speed.

This changes the population of projects. Early adopters are often unusually motivated, technically capable and selective. Once access becomes easy and competitive anxiety rises, more organisations enter, more marginal use cases are funded and more projects begin before the problem, data, workflow or value case has been properly defined.

The average failure rate can therefore worsen even while the underlying technology improves. Better models lower the cost of attempting a project. They do not automatically improve the quality of the project chosen, the data available, the operating environment or the judgement applied to continuation.

A worsening failure rate can coexist with improving technology.

Directionally, the curve may pass through five stages: selective early adoption, rapid expansion, pilot congestion, organisational learning and disciplined normalisation. The market appears to be somewhere in the untidy middle. The technology is becoming broadly accessible, but many organisations are still learning which experiments deserve to become operating systems.

Production is a different test

A pilot asks whether something can work under bounded conditions. Production asks whether the organisation can make it work repeatedly, economically, safely and usefully in ordinary conditions.

Those are not simply harder versions of the same question. They involve different evidence.

On a narrow screen, each row is shown as a separate card.

RAND found that common causes of AI project failure include solving the wrong problem, optimising the wrong metric, building something that does not fit the surrounding workflow, lacking the necessary data or infrastructure, and choosing an exciting technology rather than an enduring user problem.⁴

The pilot-to-production gap is therefore not simply a technical valley. It is where the parts of the organisation excluded from the demonstration finally arrive.

The first signs of a downward bend are organisational

The aggregate evidence does not yet prove that AI project failure has peaked. It does, however, show why success should improve as organisations mature.

Gartner reported in 2025 that 45% of leaders in high-maturity organisations said their AI initiatives remained in production for at least three years, compared with 20% in low-maturity organisations. The higher-maturity group selected projects for business value and technical feasibility, used stronger governance and engineering practices, and measured financial, risk and customer outcomes more consistently.⁵

BCG reported in July 2026 that nearly nine in ten surveyed CEOs were seeing some cost or revenue benefit from AI in targeted areas. More important than the headline, its comparison of higher and lower performers points to clear priorities, accountable leaders, strong people on high-value initiatives, workflow and role redesign, and financial value tracked through to the bottom line.⁶

BCG summarised the underlying management lesson neatly: "AI does not change the fundamentals of transformation. It magnifies them."⁶

Deloitte presents a similarly mixed but improving picture. Only 25% of respondents had moved at least 40% of their AI pilots into production, yet twice as many leaders as a year earlier reported transformative impact. The same research found that only 30% were redesigning key processes around AI, while 37% were using it largely at the surface without changing the underlying work.⁷

Gartner, BCG and Deloitte all have commercial interests in advising organisations about AI. Their figures should be treated as directional evidence rather than neutral proof. Even so, the pattern is consistent with the broader source base: the organisations producing durable value are not merely better at model selection. They are better at selection, ownership, workflow design, measurement, trust and change.

Early failure is a management capability

A rising number of abandoned pilots sounds unambiguously bad only if every pilot is expected to become a production system.

That is the wrong job description for an experiment. A pilot should buy information. It should test the hardest assumptions cheaply enough that management can make a better decision before committing production-scale money, attention and reputation.

A successful pilot does not always produce a production system. Sometimes it produces a well-evidenced decision to stop.

This changes how failure should be counted. An early project that reveals inadequate data, weak economics, low user trust or an unsolved governance problem may have succeeded as an experiment. A technically impressive pilot that continues for another year despite having no credible route to value is the more serious failure.

A mature organisation may initially kill more projects, not fewer. What should fall is the number of weak projects killed late, after integrations, recruitment, vendor commitments and internal politics have made the exit expensive.

The desired portfolio pattern is straightforward: more early deaths, fewer expensive ones and more survivors worth scaling.

Why projects survive their evidence

The difficulty is that stopping a project is not psychologically equivalent to declining to start it.

Project-management research describes escalation of commitment: decision-makers receive evidence that the expected return is unlikely, but continue allocating resources. The most studied drivers include sunk cost, self-justification, perceived proximity to completion and optimism bias.⁸

AI projects add several sources of attachment. A senior sponsor may have told the board that the initiative is strategically important. A technical team may have built identity and status around it. A vendor relationship may be established. Jobs, budgets and internal influence may now depend on continuation. Stopping can feel like admitting that the original decision was wrong.

Annie Duke, the decision strategist and former professional poker player, describes the perversity plainly: "We see the negative signals and we just invest more."⁹

The sunk cost is not only financial. It includes time, reputation, political capital and narrative. An organisation that has announced that it is becoming AI-led may find it especially difficult to say that a particular AI idea was worth testing and is now worth ending.

This is why awareness of the bias is not enough. The continuation decision must be designed before the emotional investment becomes large.

Design the right to stop

Every serious pilot should begin with a decision contract: what uncertainty the organisation is paying to reduce, what evidence would justify the next stage, what evidence would trigger redesign, and what evidence would end the project.

These conditions should be recorded before the result is known. Otherwise, success can be redefined after each disappointment and the original purpose quietly disappears.

The governance can remain proportionate. A small, reversible experiment does not need a committee. A costly or high-consequence programme may need an independent stage-gate reviewer, a continuation decision separated from the original sponsor, and an explicit comparison with the opportunity cost of funding something else.

The cultural message matters too. Managers should not be punished for closing a well-designed experiment when the evidence says no. The failure is not stopping. The failure is learning that continuation is unjustified and continuing anyway.

Bending the curve is not about making every project survive

Managers do not need to wait for the industry failure rate to improve. They can bend the curve inside their own organisation by changing the portfolio before changing the technology.

That means starting fewer projects for weak reasons, testing the hardest assumptions first, designing pilots around production questions, assigning durable owners, measuring value at workflow level and stopping projects when the evidence no longer supports continuation.

It also means preserving the rationale. Why was the project started? What was expected to become true? What did the pilot reveal? Why was the next stage approved, redesigned or refused? Without that record, learning is lost and the next AI programme can repeat the same mistake under a new name.

The current rise in abandoned AI projects may be a predictable phase of mass adoption. New capability became accessible faster than organisations learned to select, operate and govern it. If that interpretation is right, the aggregate curve should eventually turn as management discipline catches up.

But the turn is not automatic. It will be produced by thousands of ordinary decisions to start better, test honestly, redesign when warranted and stop when the answer is no.

AI project failure may get worse before it gets better. Management can bend the curve by making early stopping a sign of competence, not embarrassment.