The Most Cited AI Stat

Almost every day I’ve been on LinkedIn since August 2025, I’ve seen at least one post quoting the same number: 95% of AI projects fail. It comes from MIT Project NANDA’s report, The GenAI Divide: State of AI in Business 2025, dated July 2025 and picked up by Fortune that August under the headline “MIT report: 95% of generative AI pilots at companies are failing.” It’s close to a perfect modern example of reading the headline but not the paper, and everyone has an opinion on the headline.

In fairness, the 95% is in the report. The executive summary says that “95% of organizations are getting zero return” on $30–40 billion of enterprise investment. But almost nobody quoting it is quoting what it measures, and what it measures is more interesting than the headline. It isn’t evidence that AI doesn’t work. Read closely, the report is a careful description of how organizations are choosing to deploy it, and why that choice keeps failing.

What the Number Counts

The 95% comes from one kind of tool: custom or task-specific AI built into enterprise workflows. The report counts a deployment as successful when users or executives describe “a marked and sustained productivity and/or P&L impact,” measured about six months after the pilot (p. 7). General-purpose tools like ChatGPT and Copilot are counted separately, and roughly 40% of organizations report deploying them successfully.

So what does the 95% actually count? Notice what the chart is measuring: organizations, not pilots. Sixty percent of organizations evaluated a custom tool, 20% piloted one, and 5% got one into production with measurable impact. Among the organizations that actually ran a pilot, roughly one in four succeeded. That isn’t a failure rate. It’s a funnel.

The summary’s own phrasing, “just 5% of integrated AI pilots are extracting millions in value,” isn’t what its exhibit shows, and that sentence is the one that traveled.

The method matters too. The findings come from 52 structured interviews, 153 survey responses collected at four industry conferences, and a review of more than 300 public AI announcements. The authors call the numbers “directionally accurate based on individual interviews rather than official company reporting,” and they list their own limits: selection bias toward organizations willing to talk, success definitions that vary by company, and a six-month window that may be too short for complex systems (p. 24).

That’s an honest working paper. It isn’t a census of enterprise AI, and it was never presented as one.

It’s also worth noticing who wrote it. Project NANDA builds infrastructure for networked AI agents on top of MCP and A2A (p. 23), and the report’s conclusion is that “agentic AI and memory frameworks (like NANDA and MCP) will define which vendors help organizations cross the divide” (p. 18). That doesn’t make the findings wrong. It does mean the diagnosis and the cure come from the same lab, and the diagnosis is the part with evidence behind it.

The Learning Gap

The report’s central finding isn’t about model quality. It’s about memory. When users were asked what keeps them from trusting AI with core work, the top answers were that it doesn’t learn from their feedback, that it needs too much context every time, and that it can’t be customized to their workflow (pp. 12–13). The same users prefer ChatGPT to their company’s tools for quick work, mostly because “the answers are better.” For complex, multi-week work, they’d hand the task to a junior colleague instead of an AI by about nine to one.

Why would the same people prefer ChatGPT for quick work and refuse it for anything that matters? The two findings look contradictory until you see what they share. A consumer chatbot wins on quick tasks because the user supplies the context in the conversation and iterates until it’s right. It loses on complex work for the same reason: the user would have to supply all of that context every time.

A corporate lawyer in the report put it exactly. Her firm bought a $50,000 contract analysis tool, she kept drafting in ChatGPT, and she still wouldn’t trust either for sensitive work because “it doesn’t retain knowledge of client preferences or learn from previous edits. It repeats the same mistakes” (pp. 12–13).

What the lawyer is describing is a structural problem in how enterprise tasks get handed to models. A request like “review this contract” or “evaluate this claim” sounds like one question to the expert asking it. To a model starting from zero, it’s five or ten questions bundled together, each with definitions, thresholds and context the expert supplies without thinking. I wrote about how to find and close those gaps in Prompt Engineering Is Finding the Seams. The report arrives at the same place from the buyer’s side. Its users aren’t asking for a smarter model. They’re asking for one that already knows what they know.

Looking Where the Light Is

The report’s most underrated finding is about where the money goes. When executives were asked to allocate a hypothetical $100 of AI budget, sales and marketing took the largest share. The report isn’t consistent about how large. Its takeaway says 50%, and the paragraph beneath it says roughly 70% (p. 9). Either way, it’s the majority. Why there? The report is candid about it: “Sales and marketing dominate not only because of visibility, but because outcomes can be measured easily.”

There’s an old joke about a drunk searching for his keys under a streetlight. A policeman asks if he lost them there, and he says no, he lost them in the park, but this is where the light is. The philosopher Abraham Kaplan called it the principle of the drunkard’s search in 1964, and social scientists have used it ever since to describe research that measures what’s easy to measure rather than what matters. AI budgets have the same shape. Email response rates and demo volume are easy to show a board, so that’s where the money goes.

A VP of procurement in the report described the other side of it: “If I buy a tool to help my team work faster, how do I quantify that impact?” (p. 10).

It doesn’t help that sales and marketing are mostly natural-language communication, which is the work LLMs were built to do in the first place. Most of a business isn’t pure communication. It runs on domain language, regulated processes and handoffs across several disciplines at once, and the report found the stronger returns there. The back-office deployments it documented were eliminating $2–10 million a year in outsourced customer service and document processing, cutting external creative and agency spend by 30%, and saving $1 million a year on outsourced risk checks in financial services (p. 21). The light is on the front office. The keys are in the back.

The same finding quietly answers the jobs question. The report found that AI-driven workforce reductions so far concentrate in customer support and administrative processing, at roughly 5–20% of that work, much of it previously outsourced, and that most of the measurable savings came from cancelling outside contracts rather than cutting staff (p. 21). Those are the most standardized and most outsourced functions in a company. In healthcare, energy and advanced industries, most executives told the researchers they expected no hiring reductions over the next five years. In technology and media, more than 80% expected to hire less within two years (p. 21). Displacement is real, but it has been narrow, and it has followed the work that was already closest to a script.

The Quiet Rehiring

The cuts that made headlines tell the same story from the other direction. Klarna said its AI assistant was doing the work of 700 customer service agents, then started hiring people again in May 2025 because, in its CEO’s words, cost had been “a too predominant evaluation factor” and “what you end up having is lower quality” (Entrepreneur). Salesforce cut its support organization from about 9,000 people to 5,000 “because I need less heads” (CNBC). Amazon’s CEO wrote in June 2025 that AI would mean fewer corporate jobs, then said of the 14,000 cuts that followed that they were “not even really AI-driven, not right now at least” (Fortune).

McKinsey’s data shows how far expectations ran ahead of the work. In its 2025 survey, 32% of respondents at companies using AI expected it to shrink their workforce by 3% or more over the following year. When McKinsey asked again a year later, only 14% said AI had contributed to any decline in headcount (McKinsey).

Some of the cuts that did happen are being walked back. In October 2025, Forrester predicted that half of AI-attributed layoffs would be quietly reversed, with the jobs coming back offshore or at lower wages, and it named the pattern underneath: “AI washing,” attributing cuts made for financial reasons to AI that wasn’t ready to do the work (Forrester). Gartner went further on customer service specifically, predicting that by 2027, half of the organizations that planned to significantly reduce their service workforce because of AI would abandon those plans (Gartner).

None of this surprised me, because it’s the argument I made in The Wrong Layer of Abstraction. AI operates on tasks. A job is a bundle of tasks plus the judgment, accountability and context around them. When a company cuts the job because a model can do some of its tasks, the rest of the job doesn’t disappear. Somebody still has to handle the edge case, own the customer who’s angry, and answer for the decision.

So the work comes back, under a different title, in a different country or at a lower salary, and usually without a press release. The quiet rehiring is the wrong layer of abstraction showing up on the income statement.

Why Buying Beats Building

The report’s fifth myth is that the best enterprises build their own tools. In its sample, tools bought from and co-developed with outside partners reached deployment about 67% of the time, against 33% for internal builds, and employees used the external tools nearly twice as much (pp. 19–20). The authors are careful to say this is a correlation. Companies that buy may simply be better at buying.

Why would a vendor succeed twice as often as a company’s own team? I think there’s a structural reason behind it anyway. The scarce input in these systems isn’t the model, which everyone rents from the same few labs. It’s curated domain knowledge: the definitions, edge cases, rubrics and context that the lawyer wanted the tool to remember. That takes years and experts to build.

A vendor can spread that cost across hundreds of customers, while an internal team pays all of it for one. When an internal build skips that work, which most do, it ends up as a thin layer over the same model the users already have on their phones, and they can tell. The report’s description of struggling tools fits: the ones built on “complex internal logic, opaque decision support, or optimization based on proprietary heuristics” (p. 16).

That’s also why we chose not to market GARDE1 as AI-first. Buyers had been burned on wrappers and burned on quality. Investors had been burned on wrappers and burned on quality. The word “AI” had become a reason to be skeptical, so we led with the problem and the result instead, and with PWTL, which makes every decision cite its evidence and show its reasoning.

The report also shows how much the buying decision turns on trust rather than on the product. When executives described how they choose vendors, a vendor they already trust came first, ahead of deep understanding of their workflow (p. 15). One procurement head said it plainly: “We’re more likely to wait for our existing partner to add AI than gamble on a startup.” That’s the distribution moat I wrote about in IP and Novelty in the Age of AI, showing up in survey data. It’s also a problem every AI startup has caused for itself. Almost none of them publish their results, even against a fixed internal test set. If the only evidence a buyer ever sees is a demo, trusting the incumbent is the rational choice.

The Resistance Question

The barrier executives and users cited most often was “unwillingness to adopt new tools,” well ahead of model quality.

It’s tempting to read that as ordinary change fatigue, but the rest of the report makes that hard to believe. Nobody in this story is refusing AI.

Workers at over 90% of the companies surveyed used personal AI tools for work regularly, while only 40% of those companies had bought an official subscription (p. 8). Microsoft’s 2024 Work Trend Index found the same pattern at a much larger scale: 78% of people using AI at work were bringing their own tools.

Everyone can chat. Most people have been doing it for years. So what are they resisting? A particular tool, and the report suggests why: “employees know what good AI feels like, making them less tolerant of static enterprise tools” (p. 11). An enterprise AI product isn’t being compared with the old process. It’s being compared with a $20 subscription the user already likes.

There’s a second layer, and here the report and other research disagree. The report says that “concerns about workforce impact were far less common than anticipated” (p. 16). Pew, surveying more than 5,000 U.S. workers directly, found that 52% feel worried about how AI will be used in the workplace, against 36% who feel hopeful, and only 6% expect it to bring them more job opportunities. Microsoft found that 52% of people using AI at work were reluctant to admit using it for their most important tasks.

The difference may be who was asked, and by whom. People rarely tell their employer’s researchers that they’re afraid of being replaced, or that they don’t want to help build what replaces them. My read is that some of what shows up as “unwillingness to adopt new tools” is exactly that, expressed in the only vocabulary a rollout meeting allows.

Two Methods, One Answer

McKinsey reached the same conclusion with a very different method. Its March 2025 survey tested 25 organizational attributes against reported impact on EBIT (earnings before interest and taxes) and found that redesigning workflows had the largest effect of any of them, though only 21% of respondents had done it. The November 2025 survey of nearly 2,000 respondents found that only 39% could point to any enterprise-level EBIT impact from AI, and that the roughly 6% it calls high performers were nearly three times as likely to have fundamentally redesigned workflows. By August 2026, nearly everyone was using AI, and the share getting real value from it hadn’t moved.

Neither study is definitive on its own. MIT’s rests on a few dozen interviews. McKinsey’s rests on survey answers, which sit near the bottom of any hierarchy of evidence: respondents describe their own companies, nobody audits the figures, and the people who answer are the people who chose to. Both are correlational.

But they were built differently, asked different people, and still point at the same cause. Organizations that redesign the work around the model see returns, and organizations that insert a static tool into an unchanged process mostly don’t. When two weak instruments agree, the agreement is the strongest evidence either of them has.

McKinsey adds one more detail that matters. At larger companies, the attribute most associated with EBIT impact was CEO oversight of AI governance. That fits the report’s barrier list, where a lack of executive sponsorship sits near the top, and it fits what I’ve seen.

Leadership understanding is the constraint that everything else depends on, and most boards won’t engage with what these systems actually do unless someone with hands-on experience brings them along. The best version I’ve seen is at some of the startups we work with, where the CEO meets with engineering managers every week to learn how the team actually uses AI. It’s a skip-level habit more executives should have.

The Big Four are a poor substitute. Their incentive is to sell services and partnerships, not to tell a client the simpler answer, and I watched a $6M PwC delivery fail to answer most of the questions it was built for. In October 2025, Deloitte agreed to partially refund the Australian government for a report containing citations to papers that don’t exist (Fortune).

What the 95% Is Really Measuring

Put the pieces together and the famous number reads differently. The 95% isn’t a measure of what AI can do. It’s a measure of a particular way of deploying it: static tools that don’t learn, placed into processes nobody redesigned, chosen where results are easy to show rather than where the value is, and bought from vendors who don’t know the workflow, or built by teams who skipped the domain work.

Each of those choices shows up in the report’s own data, and each has a counterpart in the organizations that did see returns. They started in the back office, where the outsourcing bills are. They picked one workflow with a baseline and one person who owned it through production. They broke the work into its parts before asking a model to do any of them. They bought or co-developed, and held vendors to business outcomes rather than benchmarks. They let line managers and power users, not a central lab, decide where AI went, and they got from pilot to production in about 90 days rather than nine months (pp. 7, 20).

The report’s title is the GenAI Divide, but the divide it documents isn’t between companies with good AI and bad AI. Nearly everyone has access to the same models. It’s between organizations willing to change how the work is done and organizations hoping a tool will do it for them. The 95% was never a verdict on AI. It’s a verdict on the 95% of organizations that left the work the way it was.

Sources

1. MIT NANDA, The GenAI Divide: State of AI in Business 2025, July 2025
2. Fortune, “MIT report: 95% of generative AI pilots at companies are failing,” August 2025
3. McKinsey, The state of AI: How organizations are rewiring to capture value, March 2025
4. McKinsey, The state of AI in 2025: Agents, innovation, and transformation, November 2025
5. McKinsey, The state of AI in 2026: On the road to ROI, August 2026
6. McKinsey, “AI job losses fall short of forecasts,” September 2026
7. Forrester, “Predictions 2026: The Workforce Muddles Through Ambient Disruption”
8. Gartner, “Gartner Predicts 50% of Organizations Will Abandon Plans to Reduce Customer Service Workforce Due to AI,” June 2025
9. Microsoft and LinkedIn, 2024 Work Trend Index
10. Pew Research Center, “U.S. Workers Are More Worried Than Hopeful About Future AI Use in the Workplace,” February 2025
11. Entrepreneur, “Klarna CEO Reverses Course By Hiring More Humans, Not AI,” May 2025
12. CNBC, “Salesforce CEO confirms 4,000 layoffs ‘because I need less heads’ with AI,” September 2025
13. Fortune, “CEO Andy Jassy says Amazon’s 14,000 layoffs weren’t about cutting costs or AI taking jobs: ‘It’s culture,’” November 2025
14. Fortune, “Deloitte was caught using AI in $290,000 report to help the Australian government crack down on welfare after a researcher flagged hallucinations,” October 2025
15. Wikipedia, Streetlight effect