Why so many AI projects fail: the numbers, then the seven avoidable mistakes

Let us start with the numbers that circulate in every board deck, with their exact sources and their limits, because a number without a methodology is a slogan.

The most spectacular comes from MIT: the NANDA project’s report “The GenAI Divide: State of AI in Business 2025” concludes that 95% of organizations get no measurable return from their generative AI initiatives, despite investments estimated at $30 to $40 billion. A point of honesty: it is a preliminary report, based on a review of more than 300 public initiatives, 52 organizations interviewed and 153 executives surveyed, and its authors themselves describe the figures as indicative. Gartner, for its part, predicted in 2024 that at least 30% of generative AI projects would be abandoned after proof of concept; by 2026 the firm finds that reality has outrun the forecast, with more than 50% abandoned by the end of 2025.

Two further measurements complete the picture. S&P Global Market Intelligence’s 2025 survey of more than 1,000 decision-makers: 42% of companies had abandoned most of their AI initiatives, against 17% the year before, and 46% of proofs of concept die before production. And RAND’s study, based on interviews with 65 experienced data scientists: more than 80% of AI projects fail, roughly double the rate of IT projects without AI.

“We need AI” is not a business problem. Projects that start from the technology (because a competitor has it, because the board asked for it) end as elegant demonstrations nobody uses. The projects that survive start from a quantifiable pain: quotes that take three weeks, know-how heading into retirement, document searches that eat up whole days.

The test is simple: if you cannot complete the sentence “this project succeeds if [indicator] moves from [current value] to [target],” the project is not ready. It is brutal, and it saves tens of thousands of dollars.

The most common cause of failure cited by the practitioners RAND interviewed is data: not findable, scattered, contradictory, or simply absent. In an SME the symptom is familiar: the documents exist, but in three versions, across five network folders, with access rights nobody understands any more.

An AI assistant that indexes disorder answers with disorder, confidently. Hence our rule: the data self-assessment is done before choosing a tool or a vendor, and its result dictates the timeline. The good news: that diagnostic can be funded, as our overview of Quebec support sets out.

S&P Global measures that nearly one proof of concept in two dies before production. The most mundane cause: nobody defined what “succeeding” meant, so the pilot floats, neither good enough to deploy nor bad enough to stop. It eventually dies of budget exhaustion, taking the credibility of the next project with it.

The antidote is mechanical: a baseline measured before deployment, one indicator per use case and a verdict date. Our method for measuring return gives the full sequence; the essential fits in a line: a pilot with no success criterion is not a pilot, it is an expense.

A perfect tool nobody uses produces exactly zero. Adoption cannot be decreed: it is earned with a tool that cites its sources (trust), short but real training (competence) and first use cases that genuinely help day to day (interest). A useful reminder from the research: the study published in the Quarterly Journal of Economics by Brynjolfsson and colleagues shows that the largest gains go to the least experienced employees; training broadly, not just the champions, changes the project’s return.

Neglecting adoption has a cousin: ignoring shadow AI, the consumer tools your employees are already using with no oversight. An official project that arrives without an acceptable use policy leaves that risk untouched, and training, it bears repeating, is funded up to 75% by Services Québec.

The central finding of the MIT report is in its word “divide”: the organizations stuck on the wrong side are overwhelmingly those that deployed generic tools without connecting them to the company’s processes and data. The tool impresses in the demo, then fails on the trade vocabulary, the in-house templates and the special cases that make up your day.

The fix is not necessarily “build everything custom”: it is connecting the tool to your documents and your rules (the RAG approach explained in our glossary for executives), and choosing the deployment according to how sensitive the data is, as we compare in Local AI or ChatGPT Enterprise.

The scene repeats itself: the pilot works, enthusiasm rises, then someone asks where the personal information is going, and everything stops. A privacy impact assessment carried out after the fact can invalidate the entire architecture: unassessed transfers outside Quebec, missing consents, automated decisions with no safeguards.

Done at scoping, that same assessment costs a few days and strengthens the application, including with the funding bodies. Late compliance is a cause of failure; early compliance is an accelerator. This is not virtue, it is project economics.

The monumental program that promises to transform the whole company in eighteen months stacks all the previous risks at once: heterogeneous data, diluted adoption, impossible measurement, a budget exhausted before the first value arrives. When it fails, it burns AI’s credibility for years.

The opposite discipline has proved itself in the field: short waves, the simplest first, a decision point at every milestone, and the right to stop. That is the mechanism described in our field lessons, and it is what puts a project on the right side of the statistics at the top of this article.

Before you commit a dollar, run your project through the seven mistakes: a quantifiable problem, an inventory of the data, a dated success criterion, an adoption plan, a real connection to your documents, a privacy impact assessment at scoping, and a modest first wave. A project that ticks all seven boxes looks nothing like the ones the studies are counting.

If you want an outside eye on your work, the scoping call is there for that, and the 12 questions to put to any consulting firm, ourselves included, complete the armour. Our interest is aligned with yours: a client who fails costs us more in reputation than a mandate we turned down.