95% of AI pilots fail? What the MIT report really showed and what the successful companies do

Where the 95% came from

The report is called “The GenAI Divide: State of AI in Business 2025”, and it comes from Project NANDA at the MIT Media Lab. Four authors signed it: Aditya Challapally, Chris Pease, Ramesh Raskar and Pradyumna Chari. They call it preliminary findings. Fieldwork ran from January to June 2025. The authors interviewed people from 52 organisations, collected 153 survey responses from senior leaders at four industry conferences and went through more than 300 AI projects that companies had made public.

Fortune ran a story on it on 18 August 2025. The headline said 95% of generative AI pilots at companies are failing, and it spread within hours. Three days later Jeremy Kahn, also at Fortune, wrote that the report had spooked investors. Nvidia, CoreWeave, Microsoft and Alphabet all fell noticeably that week, although Kahn also pointed to Sam Altman’s comments about an AI start-up bubble.

What was actually measured

Two kinds of tools get very different treatment in the report. Over 80% of organisations had tried or piloted ChatGPT, Copilot or similar assistants, and close to 40% said they had rolled them out. Those help individual people. The P&L barely notices. Then there are tools fitted to a specific process, bought from a vendor or built in-house. 60% of organisations looked at one. 20% got as far as a pilot. 5% had one in production with measurable impact.

That last step had a high bar: past the pilot stage, with a marked and lasting effect on productivity or the P&L within six months.

Why the number is misleading

It didn’t take long for people to poke holes in it. Rob Wiblin went through the maths on the 80,000 Hours podcast. The 5% is a share of every organisation in the study, and most of them never ran a workflow-specific pilot at all. Among those that did, about one in four made it. The episode also notes that Fortune got the sample wrong: 150 interviews and 350 surveyed employees in the article, 52 and 153 in the report.

Paul Roetzer of the Marketing AI Institute thought six months and a P&L test were far too narrow, because they leave out time saved or costs cut. R. Scott Raynovich of Futuriom complained that the report never shows the data behind the 95%. The authors admit their interview-based figures are directional. It wasn’t peer-reviewed. For a while, the only way to read it was to hand over your details through a form.

So no, the study did not show that 95% of AI projects fail. It showed something narrower. In a small sample, few companies got a clear result inside six months from tools wired into specific processes. That’s worth knowing, but it is a much smaller claim.

The interesting part is why pilots stall

Read past the number and the authors’ explanation is worth your time. They rarely blame the models or regulation. They blame how AI gets rolled out.

“The core barrier to scaling is not infrastructure, regulation, or talent. It is learning.”

Most generative AI systems, they write, don’t remember feedback, don’t adapt to a company’s context and don’t get better over time. Add fragile workflows and tools that don’t match the way a team really works. For small jobs, people are fine with AI. Seventy percent would let it draft an email. Give them complex work that runs for weeks and nine out of ten pick a human. What’s missing is memory and continuity, which has a lot to do with how a tool is plugged into the process.

Budgets go to marketing, savings show up elsewhere

One exercise in the report asked executives to split an imaginary 100 dollars of AI budget between departments. About half landed in sales and marketing. The authors think that’s because results there are easy to measure: demo counts and email response times look good in a board deck. The biggest savings they saw were in the back office, including a 30% drop in spending on outside creative and content agencies.

None of this proves AI doesn’t pay off in marketing. In McKinsey’s August 2026 survey, companies most often credit marketing and sales for the extra revenue AI brings in. Both can be true. In both studies, results appear when AI changes the way a whole team works.

What the successful minority does

McKinsey asked 1,719 people in 97 countries. 37% say AI has had at least some effect on EBIT, roughly the same as last year. The real high performers, who put at least 5% of EBIT down to AI and call its impact significant, are about 6% of the sample. Nearly three-quarters of them have fundamentally redesigned workflows because of AI. Among everyone else it’s one in four. They are also twice as likely to have a set way of measuring what their AI initiatives deliver.

The MIT report sees the same thing from another angle. The companies that get there want tools built for one particular process. They judge them by what changes in daily operations; benchmark scores matter less. Most start at the edge of a process and work their way in. The ideas often come from team leads and heavy users rather than a central AI lab.

Zalando redesigned the process, CNET added AI to the old one

Zalando took the first route. In an interview the company published in May 2026, Matthias Haase, who heads content solutions, said 90% of the marketing content on its platform is now generated by AI. Product pages and ads outside the platform aren’t included. A production cycle used to take six to eight weeks. Now it takes a few days, and the goal for this year is under 24 hours from spotting a trend to having it live.

That speed comes from a rebuilt process. Trends come out of search and sales data, AI makes a first version, and the creative team decides on taste and the final look. In August 2026 the company described its next move. Human reviewers can’t reliably catch the small changes AI makes, so Zalando has put an automated check into the pipeline. It compares each generated image with the original product photos and flags anything that changes the colour, the shape or the details.

CNET took the other route. Back in November 2022, the CNET Money team began trying out an AI tool built in-house to write simple finance explainers. Editors wrote outlines and edited the drafts, which means AI was dropped into the editorial process as it already was. They published 77 short pieces, around 1% of the site’s output. One was rightly called out for errors, and an audit turned up more. The Verge counted corrections on 41 of the 77. CNET paused the tool until it and the editorial process could prevent errors, and added extra checks.

Same lesson in both. A quality check built for human mistakes won’t catch the ones AI makes. Zalando redesigned it alongside production. CNET had to do it afterwards.

What this means for your company

Planning an AI pilot in marketing or design? Keep the 95% at arm’s length, but hold on to the explanation. Before any tool gets chosen, write down how long it takes today for a brief to turn into finished assets and how many correction rounds that involves. Then pick one job that keeps coming back. Adapting a campaign to new formats works well, and the AI should live inside the tools your team already uses. Put the brand into writing, with rules, templates and examples, so the tools keep it from one session to the next. And agree who checks what the AI produces before it goes out.

That is where our AI design workflows service starts too: a process map first, then one process where AI gives your teams time back fastest. For how to write a brand down so tools don’t lose it, see our post on teaching AI your brand. For how speeding up one team pushes the bottleneck further along, see our post on the bottleneck after AI adoption.