Why do most AI projects fail?

If your AI pilot went nowhere, the useful thing to know first is that this is the ordinary outcome, not a verdict on your team.

MIT's Project NANDA published The GenAI Divide: State of AI in Business 2025 — 52 organisations interviewed, 153 senior leaders surveyed, more than 300 public AI initiatives reviewed. The headline finding: 95% of organisations got zero measurable P&L return from their generative AI investment.

That number gets quoted everywhere. The parts that actually help you are further in.

Failure one: it was never measured

The most common shape is a pilot that produced something impressive, was well received, and had no number attached to it at either end. Nobody wrote down what the process cost before. So when somebody senior asks what it saved, there is no answer — and a thing with no answer does not get funded into production.

Worth being blunt about: if you cannot state what the workflow costs you today in hours or dollars, no result can be a success, because there is nothing for it to be better than.

Failure two: it was pointed at the wrong department

The research found that 50–70% of corporate generative AI budget goes to sales and marketing. That is where the visible, demonstrable wins are — a chatbot, a campaign generator, something you can show a board.

The documented returns were somewhere else entirely. In the back office:

• $2–10M a year from eliminating business-process-outsourcing spend
• a 30% cut in external agency spend
• $1M a year saved on outsourced risk management at one financial services firm

Unglamorous work: document handling, intake, reconciliation, report production. Nobody demos it at a conference. It is where the money was.

Failure three: it was built entirely in-house

The same study measured how often pilots reached production, split by who did the work:

Twice the success rate. And on speed, the top-quartile external engagements went from pilot to production in roughly 90 days, against nine months or more internally.

We are an outside partner, so treat that as interested testimony and go read the report. But the mechanism is not mysterious: internal teams do this alongside their existing jobs, without having done it before, and there is rarely one person whose only measure of success is that it shipped.

Failure four: it was a demo

One CIO in the study put it better than we could:

"We've seen dozens of demos this year. Maybe one or two are genuinely useful. The rest are wrappers or science projects."

A demo shows that something is possible. It says nothing about whether it is worth doing, whether it survives real data, or what happens on the inputs nobody thought of. The gap between those is where the 95% lives.

Checking your own project

Four questions, and they take ten minutes:

1. What did this process cost before we started? A number, not a feeling. If there isn't one, that is the first problem.
2. Which department is it in? If it is in sales or marketing, you are in the crowded half.
3. Who owns getting it to production, as their actual job rather than as an extra?
4. What would make us switch it off? A project with no failure condition tends not to end — it just stops being mentioned.

---

Source: MIT Project NANDA, "The GenAI Divide: State of AI in Business 2025." Some press coverage misstates the sample size; the report itself is the thing to cite.