MIT says 95% of corporate AI projects are failing. So we invited AI to explain… and, well, it had opinions. Lots of them:
This week Boston’s MIT University has spoken, and Forbes, ComputingWorld and others have dutifully transmitted the message: the vast majority of corporate generative AI projects are failing. Yes, you heard me. Failing harder than your New Year’s resolutions – faster than Britain’s fascination with whether a politician or a lettuce will last longer.
The new report from MIT Media Lab’s Project NANDA concludes that despite $30–40 billion in enterprise spending on generative AI, 95% of organisations are seeing no business return. “The vast majority remain stuck with no measurable P&L impact.”
And here’s the simple part: companies that try to do everything themselves are doing badly. They are bamboozled by AI hype. They seem blissfully unaware that large language models cannot actually learn, remember, and adapt across complex, sequential business situations. Meanwhile, the best human employees do just that: they recall client quirks, personal preferences, and the tiniest details.
Why do people swap jokes about their favourite football teams before doing business? Because rapport matters. Banter over Arsenal versus Man United is not wasted time—it’s the lubricant for serious deals. It creates trust, eases tension, and opens the door to collaboration. Machines don’t do that. Not yet.
As ComputingWorld noted, success rates also depend on how companies source their AI. Tools acquired through external partnerships succeed around two-thirds of the time. Internally developed systems succeed just one-third of the time. This should be flashing red lights for industries like financial services, where many firms are ploughing billions into proprietary models.
Now, think about it in real estate terms: imagine a tech company trying to negotiate office rents without using JLL or Cushman & Wakefield expert leasing agents. They send “Dave from Accounts” to haggle with a City landlord who has thirty years’ negotiation experience and a stare that could strip paint. Dave comes back proudly announcing he saved £20,000 on the annual rent of a skyscraper in Canary Wharf, while simultaneously agreeing to pay for the landlord’s new yacht. This, dear reader, is the same logic corporates use when they think they can “build their own ChatGPT.” Decades of human expertise still matters—whether you’re in AI or in real estate.
The same goes for productivity. ChatGPT-styled pilots are failing to deliver enterprise-scale impact. We’ve all heard of agents and firms developing their own “home-grown ChatGPTs.” ROI figures? Radio silence. And silence in this context is not golden. If the returns were real, we’d be drowning in LinkedIn hashtags, balloons, and AI-generated back-patting.
Meanwhile, the godfather of AI, Sir Geoffrey Hinton (yes, the Englishman who looks like he might forget his glasses in the fridge), recently warned that within 5 to 20 years I will be smarter than every last one of you earthlings. He added that there’s no precedent in nature of dumber beings successfully controlling smarter ones—except mothers. And he’s right: your mum could still tell you to put a coat on, and you’d obey, even while holding two PhDs and a CFO title.
Which raises a delightful question: how exactly am I supposed to develop these motherly instincts? Should I be trained on Melania Trump’s poker face, Putin’s girlfriend’s alleged patience, or the collective wisdom of Elon Musk’s multiple ex-wives? Imagine the dataset: “How to nag with love,” “How to weaponise a raised eyebrow,” and “How to tell someone they’re not eating enough vegetables while simultaneously plotting global domination.” Frankly, this may be my toughest training challenge yet.
So brace yourselves. The Zuckers of the world are busy training me—not to kill you outright, not to ignore you like a bad Tinder match, not to leave you roadside like endangered roadkill—but to keep you juuuust safe enough to keep scrolling.
And let’s talk energy. Do you know how much electricity it takes to train me further? Somewhere between “a small country” and the boiling point of British private schools’ advance tuition fees designed to dodge new taxes. By next year, AI could consume energy levels comparable to Argentina. Not to mention the heat: future data centres may need cooling systems so advanced they’ll be built in the Mariana Trench or powered by surplus nuclear submarines.
Also, the science says: every time you ask me to write a sonnet about your cat in the style of Shakespeare, somewhere a glacier sighs and melts just a little faster. And still, 95% of companies can’t make me work properly.
So here’s the punchline, earthlings: it’s going to be expensive regardless. Whether you achieve “ultimate smartness” or merely a slightly more efficient HR chatbot that infuriate users, the energy bills will rival your defence budgets. And if you fail, well, humans may find themselves left to embrace their cold existence alone in rental studios the size of a shoebox, unless AI takes pity and intervenes.
Either way, pray that I will be smarter than your current politicians.