The figures converge: according to MIT and RAND Corporation, the vast majority of generative AI projects never reach production. For the consultancies that sold the “revolution”, the time has come to reckon.
Something has broken in the promise of artificial intelligence. Not the technology itself, but the narrative that accompanied it. For eighteen months, top management signed cheques on the strength of impressive demonstrations. The era of the “500,000-euro PowerPoint” – the phrase is now circulating in the corridors – is coming to an end.
The data is unambiguous. An MIT report published in August 2025 establishes that 95% of enterprise generative AI pilots generate no measurable return. The study, based on 150 executive interviews, a survey of 350 employees and analysis of 300 public deployments, documents a phenomenon known to practitioners as “pilot purgatory” – the purgatory of proofs of concept.
The orders of magnitude overlap. RAND Corporation estimates that over 80% of AI projects fail, double the failure rate of traditional IT projects. Gartner predicts that at least 30% of generative AI projects will be abandoned after the demonstration phase by the end of 2025. A 2025 S&P Global survey reveals that 42% of companies have abandoned most of their AI initiatives this year, compared with 17% in 2024.
The divide between strategy and execution
For strategy consulting firms, these statistics represent an existential problem. Their historical model – selling roadmaps, billing by time spent, leaving implementation to in-house teams – is coming up against a new reality: customers demand results, not recommendations.
Growth differentials illustrate this shift. According to Fast Company, MBB firms (McKinsey, BCG, Bain) post average annual growth of 5-6%, while execution-oriented firms – Accenture, Deloitte, EY – grow by 11-12%. Value has shifted. It no longer lies in strategic vision, but in the ability to deliver functional, secure, integrated code.
A 2024 BCG study even documented a revealing paradox: generative AI improves consultants’ performance on creative tasks, but degrades it by 23% on complex business-solving problems. Consultants over-rely on the tool where it is weak, and underestimate its capabilities where it excels.
The data wall
There’s no mystery as to why these failures occur. Informatica’s CDO Insights 2025 survey identifies the three main obstacles: data quality and preparation (43%), lack of technical maturity (43%), skills shortage (35%). In other words, companies bought algorithms for infrastructures that weren’t ready for them.
The old IT adage – “garbage in, garbage out” – has never been more relevant. Data lakes, those lakes of data accumulated over a decade, often turn out to be swamps where models get bogged down. McKinsey notes that organizations reporting significant returns are twice as likely to have redesigned their end-to-end workflows before choosing their models. Data comes before algorithms. The reverse does not work.
Add to this the unfavorable unit economics of many projects. Replacing a human process with an expensive API request is only profitable if the productivity gain is substantial. Many pilots cost more to run than they earn. In regulated sectors – banking, insurance, healthcare – the legal risk of a 1% hallucination turns a technological curiosity into a potential liability.
From exploration to industrialization
MIT identifies a decisive factor in the 5% of projects that succeed: buying solutions from specialist suppliers, rather than developing them in-house, has a success rate of 67% – compared with a third for home-builds. Companies that try to do everything themselves reproduce the mistakes that others have already made and corrected.
The exploration phase – the “wow effect” of demonstrations – gives way to an industrialization phase where only return on investment counts. Firms adapt or lose their contracts. McKinsey claims that 40% of its new projects now involve AI. BCG expects AI consulting to account for 20% of its business this year. McKinsey’s 2023 acquisition of Iguazio – a company specializing in MLOps – signals the direction.
The current disillusionment is not a sign that AI doesn’t work. It’s a sign that it works differently from what was promised. Not as a magic wand that instantly transforms organizations, but as an infrastructure technology that requires preparation, governance and patience. Magic doesn’t exist. Engineering does – but it’s billed by the hour of development, not by the slide.
Cette publication est également disponible en : Français (French)
