LEETConsulting
← Insights

2 min readLeet Consulting

The 95% problem: what the MIT GenAI report gets right

The headline says most generative AI pilots fail. The detail says something more useful: back-office automation pays, and a learning gap, not the model, is what stalls the rest.

  • Strategy
  • Production

The most shared AI statistic of the summer comes from MIT NANDA's report The GenAI Divide. As Fortune summarised it, about 5% of AI pilots achieve rapid revenue acceleration; the rest stall with little or no measurable impact on the P&L.

The 95% figure is from a preliminary research program and has been read more broadly than it should be. But three findings in the detail are worth more than the headline.

1. The problem is a learning gap, not the model

The report attributes the stall to a "learning gap": tools that do not adapt to the organisation's workflow, and organisations that do not adapt their workflow to the tools. We see the same thing. A pilot that answers questions in a chat window does not change how a claim gets processed. An agent that is wired into the claim system, with clear hand-offs, does.

2. The biggest returns are in the back office

More than half of generative AI budgets go to sales and marketing, yet the report found the biggest ROI in back-office automation. That fits our experience. Back-office work is repetitive, rule-bound and measurable, which is exactly what makes it automatable, and exactly why the savings show up in the numbers.

3. Partnerships beat solo builds

Purchasing from specialised vendors and building through partnerships succeeded about 67% of the time; internal builds succeeded about one-third as often. We read this less as "outsource everything" and more as "do not learn production AI on your most important workflow alone." The partner should leave the capability behind, not a dependency.

What we take from it

If you are planning your next AI initiative:

  • Start in operations, not in the chat window.
  • Integrate into the system of record from the first pilot.
  • Measure the P&L line the work is supposed to move, before you start.
  • Make knowledge transfer a deliverable, not a courtesy.

The 5% are not lucky. They picked work that can be measured, and then measured it.

Sources

  1. MIT report: 95% of generative AI pilots at companies are failing, Fortune, August 18, 2025.