The real cost of failed AI projects: waste, time, credibility

The numbers on AI project failure all came out within a few months of each other, and lined up they compose a picture no investment committee would accept for any other class of spending. MIT Project NANDA (The GenAI Divide, July 2025) analysed over three hundred enterprise deployments: 95% of generative AI pilots produce no measurable impact on the P&L, against 30-40 billion dollars invested. RAND Corporation estimated back in 2024 that over 80% of AI projects miss their intended value, roughly double the rate of comparable traditional IT projects. S&P Global Market Intelligence (Voice of the Enterprise, 2025) recorded the sharpest figure: 42% of companies abandoned most of their AI initiatives, up from 17% the year before.

The most common reaction to these numbers is to file the failed pilot under "learning" and restart with the next one. The true accounting of failure, the one we see surface in our assessments, is much broader than the cancelled budget line, and it is worth breaking down, because each component has a different remedy.

Direct spend is the small part

The visible cost, licences, consulting fees, vendor days, test infrastructure, is what ends up in the post-mortem, when a post-mortem happens at all. In our experience it rarely represents more than a third of the total bill, and it is also the only component the organisation perceives as lost.

Above it sits the time of your best people. A serious AI pilot absorbs for months the most capable data engineers, the most experienced process owners, hours of committees and steering meetings. If the project dies, that qualified time does not come back, and its opportunity cost is rarely calculated: what would those same people have produced on an initiative with solid prerequisites? RAND's research lists misaligned objectives and inadequate data foundations among the primary causes, conditions knowable before starting, which makes the burned time an avoidable cost rather than an accident.

The damage that shows later: internal credibility

The most expensive component is also the least accounted for. Every failed pilot raises the price of the next one, because it consumes the scarcest resource of any transformation: people's willingness to believe once more. The third AI project presented to a management team that watched the first two sink starts with a trust discount no demo can recover, and the same goes for operational teams, who quickly learn to treat the latest initiative as a passing fashion to wait out.

The MIT report describes this effect from a revealing angle: while official pilots stall, in over 90% of companies employees use personal AI tools that work. Demand for AI stays high, trust in the organisation's ability to meet it collapses, and the gap between the two curves is the internal reputational cost of failure, the one that turns future projects into uphill climbs.

Failing well, failing badly

One clarification, before the wrong conclusions: a share of failures is physiological and even desirable, because AI remains a territory where validation happens by experimenting. The problem with the 95% is not that pilots close, it is how. An experiment with a clear hypothesis, a written success threshold and a contained budget that ends in a no has produced information at low cost, and that is research well spent. A project started without criteria, grown by inertia and abandoned out of exhaustion after eighteen months has produced only the bill above.

The difference between the two outcomes is decided almost entirely before kickoff, and it is a matter of measurable prerequisites: data ready to survive scale, a sponsor with signing power, business metrics defined by whoever owns the process, a mapped risk and compliance perimeter. The failure causes catalogued by RAND and MIT are largely diagnoses a serious assessment can formulate in a few weeks, at a cost that is a fraction of a single pilot.

The missing due diligence

If 95% of pilots deliver no return, the highest-yield lever is not improving the execution of the next one, it is selecting better upstream. Our AI Rating does exactly this internal due diligence work: it measures on a 0-5 scale the organisation's capacity across four dimensions and says, before the investment, which gaps would doom the project and what closing them costs. Companies arriving from a string of failures usually discover that their pilots were not wrong in the idea, they were premature in the prerequisites.

The bill for failed projects, in the end, can be read two ways: as the price paid for learning, or as the measure of how much starting to measure is worth. If your 2025 contains at least one pilot archived without a written why, half an hour at calendly.com/fabiolalli/zerofive helps keep 2026 from repeating it, or write to hello@zerofive.ai. What did the last AI project you closed really cost, counting everything?