Why 95% of AI Projects Fail (and How to Land in the 5%)

In 2025, MIT went and checked how enterprise AI was actually going, all thirty to forty billion dollars of it, and found that about 95% of projects deliver no measurable business value. Ninety-five percent. The tempting read is to blame the technology: the models hallucinate, the tools aren’t “enterprise-ready,” give it a year, wait for GPT-6. Comforting, too, because if the problem is the tech, the fix is to wait, and waiting asks nothing of you.

But the study gives the game away: the winners and the losers were mostly running the same models. The tech showed up for both sides. So the projects aren’t dying in the code. They’re dying in the first conversation, before anyone has typed a prompt, and once you know the tell, you can hear it inside two minutes: the person can tell you everything about the AI and almost nothing about the problem. They know which model they want, they’ve picked the orchestration framework, they’ve got opinions about vector databases. Ask what the output is for and the sentences get noticeably shorter. Every single time.

Here’s the conversation I keep having. A founder wants AI to do lead generation. Everybody wants that one, and I get it: hypothetically, it’s beautiful. The vision is always some version of the same movie: personalized emails going out by the thousands while you sleep, meetings just appearing on the calendar like magic. So I ask the question I always ask: does any part of your funnel work right now? Have you ever, even once, sent the emails by hand and booked a single meeting?

And there’s this pause.

Because the answer is no. They’re asking me to help them automate a process that has never worked one time. I don’t say it harshly, because I’ve been on the other side of that pause.

But automation only does one favor for a broken process: it helps it fail faster, at scale, on a subscription.

You’ve been on the receiving end of where this goes. The email that opens with “I came across your profile and was blown away,” addressed to a first name that is very obviously a merge field. That’s what it looks like when a funnel that never worked gets a machine strapped to it. The dream was ten thousand sends a day. The result is ten thousand new ways to get marked as spam, daily, automatically.

Walk it backwards and the failure was loaded in at the start. It began with a decision to use AI. The board asked about the AI strategy, a competitor put the word in a press release, a founder read a thread on a plane and landed with a mission. Nobody picked a problem. They picked a technology and went hunting for a place to put it, which is the entire decision running in reverse. Then the place they picked got chosen for the wrong reason: a demo looked incredible. A clip of an agent booking its own flights, an app built from one sentence, and everyone’s eyes go wide like the first time you saw the T-1000 walk through those bars. Nobody was measuring impact. They were measuring cool. And cool, I say this with love, has never once shown up on a P&L.

From there the rest follows like dominoes. Because the dream is “AI does it all,” they wire up the entire pipeline at once: no decomposition, no checking whether each step works on its own, because breaking a problem into pieces is boring and they came for magic. Theoretically, it all works. What they get is a pipeline that fails somewhere in the middle, and nobody can say where. Was it the scrape? The prompt? The model having a weird Tuesday? Unknowable. They built a black box and now they’re mad at the box. And when output comes out the other end, it sounds right. There’s volume. A room full of people nods. Nobody asks whether it’s any use, because way back at the start, nobody ever said what “useful” was supposed to mean.

Then, on the calls that are already in trouble, there’s a question I ask that stops things cold, and it’s two words: then what?

The AI wrote the thing. Then what? Where does it go? Who picks it up? What happens next, and does that next step exist? People get so locked in on the AI step — the impressive step, the demo step — that they forget the output has to land somewhere. It has to enter a real process, get used by a real person, trigger a real next move, or it’s just very expensive exhaust. This is bigger than “keep a human in the loop.” It’s the whole loop: how this one new piece fits into the tools, people, and handoffs you already have. I’ve watched a team build a beautiful system that generated competitive-intel reports nobody had asked for, on a schedule nobody checked, into a folder nobody opened. Flawless execution. The answer to “then what” was a shrug, so the shrug is what it produced.

The people who land in the 5%

Here’s what I notice about the ones who pull it off. We already know what they don’t have — better models; same ChatGPT, same Claude as the other 95. What’s different is the order of operations. They run the chain forwards.

They start from a problem that was annoying them before AI was in the picture. Not “where can we use AI” — more like “this eats eleven hours a week and I hate it.” Sometimes they conclude AI isn’t even the right fix, and that conclusion saves them six months of building a product that never needed to exist. Passing on a bad build is a result. It doesn’t demo well, but it’s a result.

They’re strangely resistant to impressive demos. The question they ask is colder: if this worked perfectly, what changes? If the honest answer is “not much, but it would look amazing,” they walk. No hard feelings, just no. That immunity is rarer than it sounds, because the entire industry is engineered to defeat it.

They have a feel for the shape of the technology — where it’s reliable, where it’s a coin flip, where it comes apart the moment real data touches it. The same model that summarizes a fifty-page contract without breaking a sweat will add up the numbers inside that contract and hand you a wrong total with complete confidence — no flag, no hedge, just vibes and a wrong answer. The people who win know which of those two jobs they’re giving it.

They match the tool to the job, because AI isn’t one thing. Drafting an email, refactoring a codebase, pulling live data off a web page, running a hands-off multi-step task — four different jobs, four different tools, even when the same model powers all four. Meanwhile the person forcing a serious data job through a chat window is wondering why it falls apart at row one hundred, and the person who stood up an elaborate autonomous agent is slowly realizing a ten-line Python script would have done it perfectly, forever, for free. That one stings. I know because I’ve been that person, and the script is still running.

They build in pieces and check each piece. Slower at the start, dramatically faster at the end, because when it breaks — and it breaks — they know exactly where. No black box, no mystery.

They put a person exactly where a mistake is expensive, and nowhere else. A wrong word in a draft nobody sent yet costs nothing. A wrong number in an invoice that already went out the door costs plenty. Human at the second one, machine alone at the first. They’re not removing people from the loop. They’re placing them.

And they follow the chain all the way out to where the output gets used — because they asked “then what” at the start, they know the AI producing something is the middle of the job, not the end of it.

Underneath all of it, they keep asking two questions the other 95% skip: does this matter, and does this work. That’s the whole game. It isn’t clever. It’s just relentless, and relentless is unglamorous, which is exactly why it’s rare.

The unglamorous part

So before the next AI project, a few honest questions. What’s the problem, and what changes if it’s solved? Is this a job AI is good at, or a job it’s being forced into? What’s the right tool for this specific task — not the best model in the abstract? Has anyone run this by hand, once, and watched it work? And where does a person need to be standing when it’s wrong?

If those don’t have answers yet, the project isn’t ready, and no amount of model horsepower rescues a project that started in the wrong place.

That, more or less, is what my next book is about — the questions worth asking before you build anything with AI, and how to spot the one project worth your time in a lineup of ten that only look good in a demo. But you don’t need the book to start. Everything above is available today, free, to anyone willing to do the boring part first.

The 5% aren’t smarter, and they don’t have better technology. They did the boring work first, and almost nobody wants to, because the boring work has no demo day.

But that’s the whole opening, right there, hiding in plain sight like a cheat code somebody printed in the manual. Most people skip the manual. Read the manual. Let’s go.