Most AI initiatives now fail not at the idea stage, and not even at the pilot stage. They fail at the point where someone has to prove the number, and there is no number to prove.
How many AI projects fail?
MIT's The GenAI Divide: State of AI in Business 2025, out of the Media Lab's Project NANDA, is the sharpest number available: against an estimated $30–40 billion of enterprise spending, roughly 95% of organisations saw no measurable return. It drew on 300 disclosed AI initiatives, 150 leadership interviews and 350 employee surveys.
RAND's 2024 report on the root causes of AI project failure is the other one worth reading, though it is worth being precise about what it says: RAND notes that by some estimates more than 80% of AI projects fail, roughly twice the rate of non-AI IT projects. That 80% is RAND summarising existing estimates, not a figure RAND measured. Their own contribution is qualitative — 65 structured interviews with data scientists and engineers about why projects die.
We have deliberately cut several sharper-sounding percentages that were previously in this post. They came from secondary write-ups we could not trace to a primary source, and a number we cannot stand behind has no business on a page arguing that you should only trust numbers you can stand behind.
What do the survivors do differently?
It is not the model, the vendor, or the budget. It is that the survivors decide what “worked” means in numbers, in advance, and baseline against it before anything is built. Every other project retrofits a justification after the fact, which is exactly when the number turns out not to exist.
What does baselining look like?
Pick the handful of numbers the business already tracks, or should: hours, quote turnaround, error and rework rate, cash days, on-time delivery. Measure them before anything changes. Then measure again after. If a stage will not move one of those numbers, it does not get built, and you find that out before spending on it, not after.