In August 2025 a single number landed on the tech agenda like a bomb: 95% of companies get no measurable return from their generative AI investments. The source was "The GenAI Divide", a report from MIT's NANDA project.
The number was so striking that it got talked about more than the report itself, and some of that talk was wrong. In this post I look at what the report actually says, what it doesn't, and what the small winning minority does differently.
First, a reading note: what does 95% measure?
Reading this number as "AI doesn't work" would be a mistake. The report says organisations' rate of adopting AI is high, but their rate of transforming the business with it is low. The 95% means that most pilot projects stall without leaving a measurable mark on the company's profit and loss.
AI tools don't work. Employees don't use them. Investing in this technology is a waste.
Most companies can't embed AI into their workflows in a way that learns and adapts, so the pilots never show up in the P&L.
The method matters too. According to the report, the study is based on a review of more than 300 publicly disclosed AI initiatives, structured interviews with 52 organisations, and a survey of 153 senior leaders at four industry conferences. Some news coverage quoted different numbers, and the report was presented as preliminary work. The sample is small and the definition is debatable. Still, the direction of the findings matches the "stuck in pilot" picture Oxford saw at financial institutions almost exactly.
The AI nobody asked for, used anyway
To me, the most interesting finding in the report is that while official AI projects crawl along, employees are already using AI heavily.
This picture says two things. First, people do get value from AI; otherwise they wouldn't open their own accounts to use it. Second, the tool the organisation buys or builds is less useful than the chatbot in the employee's hand. There's a security angle too: company data flowing from personal accounts into public tools is the easiest route to the secret and data leaks I described in the first post.
The real problem: the learning gap
The report locates the cause of this divide not in model quality or regulation, but in a learning gap. General-purpose chatbots are excellent for drafting. But they fail in critical workflows: they don't remember context, they don't learn from feedback, they don't adapt to the company's processes. A tool you have to brief from scratch every time can't become part of daily work.
The money goes to the wrong place
The report's budget analysis shows a separate mismatch: more than half of generative AI budgets go to sales and marketing tools. Yet the highest return is in back-office automation, which gets much less attention: applications that cut outsourcing, agency costs and repetitive operational work.
The reason is understandable. Sales and marketing impact is easy to show and looks good in a presentation. The back office looks boring. But that's exactly where the measurable savings are: an outsourcing contract that shrinks, an invoice that disappears.
Buy or build?
As a developer, this was the finding that made me think the most.
At first glance it reads as "don't build it yourself, buy it". I read it differently. In-house projects usually start by picking a model and then get stuck on integration, data preparation and maintenance; that's the counterpart of the Oxford report's finding that "60–70% of the spending has nothing to do with AI". A focused vendor, on the other hand, has already solved those things many times for a single problem. So the question isn't "buy or build"; it's working with someone who has solved one problem in depth.
The effect on the workforce: a quiet shift
According to the report, the impact isn't coming as mass layoffs. Companies simply aren't refilling vacated positions. The change is concentrated in customer support and administrative work, especially in roles that had previously been outsourced. That's consistent with the back-office finding again: the most measurable saving is on the invoice paid to outsiders.
What are the winning 5% doing differently?
The report's lead author, Aditya Challapally, puts the common ground of the successful cases down to three things: picking one pain point, executing it well, and partnering smartly. According to the report, some young startups using this approach went from zero to 20 million dollars in revenue within a year.
- A general-purpose tool, a general goal
- Budget goes to visible work: sales and marketing
- Building everything in-house from scratch
- Systems without memory that don't learn
- Success metric: "is it being used?"
- One pain point, in depth
- Budget goes to measurable work: the back office
- A focused vendor or partnership
- Learning systems embedded in the workflow
- Success metric: "which invoice got smaller?"
The lesson I take from it
I read this report as a checklist for anyone building products. When I design an AI feature, I ask myself three questions:
- Which single job does it do? "An assistant" isn't a product; "a system that prepares the weekly progress report for parents" is.
- Does it remember the user's corrections? If it doesn't, a week later the user goes back to their personal chatbot.
- How will I measure success? Not by usage counts, but by the time saved, the cost reduced or the revenue gained itself.
Perhaps the most hopeful message of the report is this: the divide comes not from the technology, but from the approach. And an approach can be changed.
Sources
- MIT NANDA, A. Challapally et al., The GenAI Divide: State of AI in Business 2025, July 2025 (preliminary report).
- Fortune, MIT report: 95% of generative AI pilots at companies are failing, August 2025.
- Virtualization Review, MIT Report Finds Most AI Business Investments Fail, Reveals 'GenAI Divide', August 2025.
- Yahoo Finance, The shadow AI economy is booming, 2025.
- W. Kilminster, P. Özcan, AI Preparedness in Established Financial Services Firms, Oxford, 2025.