Why do most AI projects fail?
Short answer: almost never the model. Projects die in messy data, integration into systems never designed for it, workflows people do not adopt, and a missing last mile of evals and guardrails.
S&P Global found 42 percent of organizations abandoned most of their AI initiatives in 2025, up from 17 percent, scrapping 46 percent of proof-of-concepts before production (S&P Global Market Intelligence, 2025).
Almost never the model. It's data, integration, workflow, and the last mile to production.
›transcript
- Why most AI projects never ship.
- Over 80% of AI projects fail (RAND).
- Only about 5% see real returns (MIT NANDA).
- It's rarely the model. It's everything around it.
- Run the free, interactive course at heimlandr.io/ai-reality-map.
// the deep dive
The demos work. What kills the project is the boring middle: messy data nobody wants to clean, integration into systems never designed for it, a workflow people will not actually adopt, and a last mile of evals and guardrails that turns a clever prototype into something you can trust in front of a customer. The numbers around this are stark and worth reading carefully. MIT's Project NANDA reported that only about 5 percent of integrated enterprise AI pilots showed measurable profit-and-loss impact (the widely quoted "95 percent failed" framing is contested, and applies most cleanly to custom internal builds). RAND estimates more than 80 percent of AI projects fail, roughly twice the rate of non-AI IT projects. McKinsey found 88 percent of organizations now use AI somewhere, yet only about 39 percent report any enterprise-level EBIT impact. When a vendor's plan skips the boring middle, that is the tell.
// chapter faq
What percentage of AI projects fail?
RAND estimates more than 80 percent, roughly twice the rate of non-AI IT projects. S&P Global found 42 percent of organizations abandoned most of their AI initiatives in 2025, scrapping 46 percent of proof-of-concepts before production.
Is the "95 percent of AI pilots fail" statistic true?
It is a contested framing of MIT's finding that only about 5 percent of integrated enterprise pilots showed measurable profit-and-loss impact, and it applies most cleanly to custom internal builds. The honest version is bad enough: most projects produce no measurable return.
What kills AI projects most often?
The boring middle: data quality, integration and adoption. The demo works; what fails is wiring it into real systems, getting people to change how they work, and building the evals that make it trustworthy in front of a customer.
Every figure in this chapter is sourced. The full source list lives on the main map. Open the map
This is one chapter of ten. The whole course is free.
The full map has the interactive tools, the 8 minute audio edition, the live layer and every source. And if you want it run against your own reality, that call is free too.