The MIT NANDA report won't age well — and that's the problem

The MIT Project NANDA report — “The GenAI Divide: State of AI in Business 2025” — dropped about a week ago, and the headline already escaped containment.

95% of generative AI pilots fail to deliver measurable business impact.

It is everywhere now. Fortune ran it on August 18.¹ LinkedIn carousels by lunch. Every consultancy with a marketing team will be quoting it for the next two quarters. And the number is real — based on 150 leader interviews, 350 employee surveys, and analysis of 300 public deployments. The methodology holds.

But the framing — that AI is failing companies — is going to age badly. The report does not actually say that. Read past the executive summary and the story gets sharper.

What the report actually says

The MIT researchers identified four structural patterns that separate the 5% from the 95%. Three of them have nothing to do with the AI:

1. External vendor tools succeed twice as often as internal builds. Not because the engineers are worse — because building production AI is a different craft than building a working demo.
2. AI investment is heavily biased toward sales and marketing, despite operations and finance workflows showing better ROI.
3. Mid-market companies move faster than enterprises. Top performers averaged 90 days from pilot to full implementation. Enterprise teams with dedicated AI staffs reported the lowest pilot-to-scale rates.

The fourth finding is the one nobody is quoting on LinkedIn but is the most useful:

The core barrier to scaling is not infrastructure, regulation, or talent. It is learning. Most GenAI systems do not retain feedback, adapt to context, or improve over time.

The systems that fail are static. They generate a draft. They answer a query. They never get better, because nobody built them to. The ones that succeed have what the report calls “memory and workflow adaptation” — they remember what worked, learn from corrections, and integrate with the surrounding business processes.

That is an engineering problem with a known answer. It is not a fundamental limit of the technology.

Why this will age badly

The “95% fail” stat is going to get cited for two years by people who never read past page three. It will become the lazy executive’s excuse for not starting an AI project (“did you see that MIT report?”) and the lazy consultancy’s pitch for selling them strategy decks instead of working systems.

Both are wrong. The MIT data shows the opposite of what the headline implies: the 5% who succeed are doing identifiable, repeatable things. The 95% are not failing because AI doesn’t work. They are failing because they tried to build AI projects without anyone on the team who had built one before, picked workflows where the value was unclear from the start, and skipped the unglamorous data plumbing work that determines whether anything actually ships.

Aditya Challapally, the lead author, told Fortune what the successful startups do:

“They pick one pain point, execute well, and partner smartly.”

That is the entire playbook. Pick one. Execute well. Partner with someone who has done it before. The companies that follow this in 2025 will be looking at very different P&Ls in 2026 than the ones who keep funding committee-designed “AI strategies.”

What this means right now

If you are at a company considering an AI project, the actionable read of the MIT report is:

• Pick one workflow with a measurable before-and-after. If you cannot describe it in one sentence, you do not have a project.
• Buy the operational layer, build the company-specific layer. Trying to recreate evaluation harnesses, cost controls, governance, and orchestration from scratch is how 95% of pilots end up in the 95%.
• Invest in data integration before model selection. The model is the most replaceable part of the stack. The data plumbing is the part that determines whether anything works at all.
• Aim for 90 days to production, not nine months. The longer the pilot runs, the more likely it is to never ship.

The MIT report is the most useful piece of AI research published this year. It is also going to be the most misused. The companies that read it carefully — and the ones that pick partners who have read it carefully — will be on the right side of the divide it describes.

---

Footnotes

1. Sheryl Estrada, “MIT report: 95% of generative AI pilots at companies are failing,” Fortune CFO Daily, August 18, 2025. Full report: MIT Project NANDA, “The GenAI Divide: State of AI in Business 2025,” July 2025. ↩