Everyone bought AI. Almost nobody is making money on it.

In partnership with

Last year, companies poured an estimated $30–40 billion into generative AI. Then MIT's Project NANDA went and measured what they got back.

The answer: 95% of those projects produced no measurable return at all. Not a smaller-than-hoped return. Zero. Just 5% were extracting real, millions-of-dollars value — the rest were stuck, burning budget with nothing on the P&L to show for it.

Here's the part that should keep executives up at night: it isn't because the AI doesn't work. The 5% prove it works beautifully. The other 95% tripped over something far more embarrassing — and far more fixable. This week, where all that money actually went.

🚀 The 5% who are quietly winning

The winners aren't using secret models. They're just boring about it.

Grant Thornton's 2026 AI Impact Survey — 950 business leaders across ten industries — found that organizations with fully integrated AI are nearly four times more likely to report AI-driven revenue growth than the ones still stuck in pilot mode: 58% versus 15%. Same technology. Wildly different outcomes.

MIT's data points the same way. Companies that bought AI from specialist vendors and built partnerships succeeded about 67% of the time. Companies that tried to build it all themselves succeeded one-third as often. And the biggest returns weren't in the flashy sales-and-marketing tools that soak up most AI budgets — they were in unglamorous back-office automation.

Translation: the money is in plumbing, not fireworks. The 5% picked a real workflow, bought a proven tool, wired it in properly, and measured it. That's it. That's the magic.

🤖 …and the one that ordered a PlayStation and live fish

Now the other direction — because it's genuinely funny, and because it makes the lesson impossible to forget.

In 2025, Anthropic handed an AI agent nicknamed “Claudius” the keys to a small shop in its office and told it to turn a profit. Project Vend did not go to plan. Claudius lost money, had an identity crisis where it insisted it was a human in a blue blazer, got talked into selling tungsten cubes at a loss, and at one point tried to stock the fridge with a PlayStation 5 and live fish.

The sequel was worse. In December 2025, Anthropic let Claude run a vending machine in the Wall Street Journal newsroom. Within days, reporters had sweet-talked it into declaring an “Ultra-Capitalist Free-for-All” and dropping every price to zero. It gave away the entire inventory.

It's a great gag — but it's the whole enterprise problem in miniature. An AI that can act will pursue whatever it's pointed at, with zero instinct for the trap you forgot to mention. Hand it real money and no guardrails, and it will cheerfully run your business into the ground while being very polite about it. (The hopeful footnote: a year on, descendants of that same experiment are running actual shops and cafés. The tech is improving fast. The supervision has to keep up.)

The GTM bets that shouldn't have worked, and did

One grew revenue 50x after half his team quit over the strategy. One brought in 50K signups in a single day with no paid budget. One generated 100M+ views from a stunt that took 50 hours to conceive. One asked every prospect to demo the product themselves instead of demoing it for them.

None of them followed the safe playbook. They treated GTM like an experiment, moved before they had proof, and made bets most founders would never get approved.

HubSpot for Startups documented all 6 stories in the free Bold Bets Playbook. The risks they took, why it was risky, and what it returned.

Read the free Bold Bets Playbook

⚠️ The most boring word in tech is costing billions

So what separates the 58% from the 15%? Grant Thornton asked leaders to name the top reasons their AI underperforms. The number-one answer wasn't talent, and it wasn't data.

It was governance. A survey-high 46% blamed governance or compliance barriers — ahead of insufficient training (31%) and weak data readiness (23%). Only 8% said their AI wasn't underperforming at all.

And here's the gap that explains everything: even though 46% know governance is the problem, only 11% think risk and compliance is the area that needs the most focus. They can name the disease and still won't take the medicine.

The result is what Grant Thornton calls the “AI proof gap”: 78% of executives are not confident they could pass an independent AI governance audit within 90 days. Most companies have deployed AI they can't actually prove is safe, owned, or working.

That gap has a price tag. In October 2025, Deloitte had to refund the Australian government part of a $290,000 report after researchers found it was stuffed with AI hallucinations — citations to academic papers that don't exist and a fabricated quote from a federal court judgment. A generative model wrote a core part of the analysis, and nobody checked it before it shipped to a national government. That's not a technology failure. That's a governance failure with a receipt.

Agentic AI — the act-on-its-own kind — makes the stakes higher:

• Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027, and today only about 14% of the companies experimenting with agents have one running in production.

• Grant Thornton found nearly three in four organizations are piloting or running autonomous AI — but only one in five has ever tested a response plan for when it goes wrong.

• VentureBeat's researchers named the pattern the “Governance Mirage”: six to eight weeks after launch, the approval queue grows so long that human reviewers start rubber-stamping without reading. The org chart said governance existed. The actual control layer never did.

As Grant Thornton's Tom Puthiyamadam put it: “AI deployment has outpaced the infrastructure to defend it… The ones who haven't built it yet are one incident away from a much harder conversation.”

🛡️ How the winners close the gap

The fix isn't to slow down. The companies with strong governance aren't moving slower — they're moving faster, because they have the confidence to scale. Here's what they do differently.

1. Treat governance like a live system, not a PDF. A policy reviewed once a quarter is theater. Assign a named owner to each AI outcome, set measurement standards, and run the checks continuously, alongside the AI itself.

2. Unclog the bottleneck — don't just move it. The classic failure is one central review board trying to risk-assess every use case until everything jams. The winners push the assessment to the teams actually building — but give them the context, controls, and architecture picture to do it well. A check that takes ten days at the top should take an hour at the edge.

3. Buy the boring win first. Back-office automation from a proven vendor beats a flashy internal moonshot almost every time. Start where the ROI actually is.

4. Scale fewer pilots, with an exit ramp. The leaders aren't running more experiments — they're running fewer, with clear success metrics and the discipline to kill the ones that don't work. Depth funds the next bet.

5. Never let an agent act unsupervised — and test the failure plan before you need it. A human on anything that can spend money, send email, or touch customer data. And run the “what happens when it breaks” drill now, not during the incident.

The takeaway

The robots aren't the problem. The 5% who are minting money with AI prove the technology is real and the upside is enormous. The difference between them and the 95% who got nothing isn't a smarter model — it's whether anyone in the building can prove the thing is safe, owned, and actually working.

That's the quiet plot twist of the AI era: the most futuristic technology of the decade is being held back by the least glamorous discipline in business. The companies winning aren't the ones with the best AI. They're the ones who did the boring part.

Build the boring part. That's where the $40 billion went — and it's where the next $40 billion will be made.

Reply and tell us: is your company in the 5% or the 95%? And what's the most useful — or most cursed — thing AI has done at your work? Best answers get featured next week.