There’s a pitcher going around conference rooms and boardrooms right now. It’s branded with words like “transformative,” “revolutionary,” and “the most important invention in human history.” That last one comes from Sam Altman, who happens to be selling the stuff.

The AI industry has mixed up a batch of promises so intoxicating that companies are gulping it down at a rate of $581 billion in 2025 alone, with global AI spending projected to hit $1.5 trillion the same year.

Trouble is, when punch gets served at that velocity, nobody stops to read the ingredients.

The 95% Problem

MIT’s Project NANDA released a study in mid-2025 that should have been a splash of cold water across every C-suite in the world. After examining over 300 AI deployments, conducting 150 interviews, and surveying 350 employees, the researchers found that 95% of generative AI pilots are producing zero measurable return. Not low return. Not “we’ll see results eventually.” Zero.

Enterprises poured an estimated $30 to $40 billion into generative AI initiatives, and nineteen out of twenty of those projects delivered nothing you could take to a CFO with a straight face.

S&P Global’s data was worse on momentum. 42% of companies abandoned most of their AI initiatives in 2025, up from just 17% the year before. The average organisation scrapped nearly half of its AI proof-of-concepts before they ever reached production.

That’s not a correction. That’s a stampede toward the exit.

And the people writing the cheques? Less than 30% of AI leaders reported that their CEOs were satisfied with the return on AI investment. Meanwhile, 92% of surveyed executives told McKinsey they planned to increase AI spending over the next three years.

So the overwhelming majority of leaders are unhappy with results, and plan to spend more anyway. That’s not strategy. That’s the Kool-AI talking.

When the AI Meets the Drive-Thru

The numbers tell one story. The real-world failures tell a funnier and more damning one.

McDonald’s spent three years testing an AI-powered drive-thru ordering system built with IBM. The vision was elegant. Artificial intelligence takes your order, speeds up the line, cuts labour costs. The reality went viral on TikTok. The bot misheard customers, confused orders between lanes, added hundreds of dollars’ worth of unwanted McNuggets to a single order, mistook ice cream for ketchup and butter, and helpfully topped a dessert with bacon nobody asked for.

After becoming one of the internet’s favourite jokes, McDonald’s pulled the plug in July 2024, removing the technology from all 100-plus test restaurants. The system now has its own entry in the Museum of Failure. Literally.

In one case, attorneys representing MyPillow CEO Mike Lindell submitted a court filing containing more than two dozen errors, including citations to cases that simply don’t exist. The judge fined each lawyer $3,000. In another, a law firm was hit with $31,100 in sanctions for nine hallucinated citations. A Denver attorney, caught denying he’d used AI, accepted a 90-day suspension after investigators found he’d texted a paralegal about the fabrications, admitting that “like an idiot” he hadn’t checked the work.

Stanford researchers tested general-purpose AI models on legal queries and found hallucination rates between 58% and 82%. Even specialised legal AI tools built on retrieval-augmented generation, the technology marketed as the cure for hallucinations, still make things up in roughly one out of every six queries.

These aren’t edge cases. They’re the product performing exactly as designed in the field.

The Cult Mechanics

What makes the AI hype cycle different from previous tech bubbles isn’t the technology. It’s the psychology. The Kool-AI works because of three forces operating at the same time.

The first is FOMO at institutional scale. When your competitors announce an AI strategy, any AI strategy, the pressure to match them becomes existential. A Gallup poll found that only 15% of employees report their workplaces have communicated a clear AI strategy. But having a strategy and having a good one are very different things, and most companies are choosing speed over sense. They’re buying the pitcher before they know what’s in it.

The second is the demo-to-deployment gap. AI demonstrations are extraordinarily impressive. A chatbot that writes poetry, summarises documents, and generates code in seconds makes for a compelling boardroom presentation. But deploying that same technology into messy, real-world workflows with inconsistent data, legacy systems, and humans who don’t behave like test subjects, is a fundamentally different problem. Gartner found that 57% of organisations acknowledge their data isn’t even AI-ready. You can’t build a cathedral on sand, but plenty of companies are trying.

The third is the sunk cost spiral. Once you’ve committed millions to an AI initiative, announced it to shareholders, hired a Head of AI, and restructured teams around it, admitting failure becomes almost impossible. So you spend more. You launch a second pilot. You rebrand the first failure as “phase one learnings.” The RAND Corporation studied AI project failures and found that the biggest killer wasn’t technical. It was misaligned incentives and the absence of end-user input in design. The people building the systems weren’t talking to the people who’d have to use them.

What’s Actually in the Cup

None of this means AI is useless. That would be as lazy as saying it’s magic.

The same MIT study that found 95% failure also found that the 5% that succeeded shared clear patterns. They targeted back-office automation rather than flashy customer-facing features. They bought specialised tools from domain-specific vendors rather than trying to build everything in-house. Vendor-led projects succeeded roughly 67% of the time, compared to about 33% for internal builds. They set measurable goals before writing a single line of code. And perhaps most importantly, they let the people closest to the work (line managers and power users) drive adoption from the bottom up, rather than imposing AI from the top down.

The hype is settling. The companies that will matter are the ones doing the unglamorous work of cleaning their data, integrating AI into specific workflows, and measuring outcomes with the same rigour they’d apply to any other investment.