Here is a number that should embarrass an entire industry: 95%. That is the share of enterprise AI projects delivering no measurable business return, according to a 2025 study out of MIT. Companies have poured tens of billions into generative AI, and the overwhelming majority have precisely nothing to show for it. The obvious question is why, and the answer is more uncomfortable than “the models are not good enough”.
This one started with a sharp essay from Angus Sewell, aka Angus the Nontechnical, whose argument is blunt and, I think, largely correct: your AI problem is really a training problem. You cannot buy your way out of it with more software or another agency. As he puts it, “don’t rent your team a brain”. Build one.
What MIT actually found
The MIT Media Lab report, “The GenAI Divide”, is worth reading properly. It surveyed and interviewed executives and analysed hundreds of public AI deployments, and its central finding is not that the technology is broken. It is that most organisations cannot integrate it. The report pins the blame on a “learning gap” that runs through both the tools and the people using them. Around 40% of companies have deployed AI tools; only about 5% have woven them into real workflows at scale. Everyone bought the shiny thing. Almost nobody learned to use it.
Why “buy training, not tech” is the right instinct
Angus’s prescription is to stop renting consultants and commissioning custom builds that rot the moment the contract ends, and instead spend the time teaching your existing staff to use, adapt and eventually build with AI themselves. This lines up neatly with the MIT data. A bespoke system nobody understands is a liability with a shelf life. A team that actually knows how to wield off-the-shelf tools compounds in value every week. One depreciates; the other appreciates.
The bit the vendors will never tell you
Here is the part worth saying out loud, because nobody selling into this market ever will. There is no fat margin in telling a company “your people already have the tools, they just need to learn them.” You cannot bill two hundred grand for that. The entire consulting-and-custom-build economy exists precisely because “train your own staff” is not a product you can mark up. So the industry sells complexity instead: platforms, integrations, bespoke agents, a permanent dependency. The 95% failure rate is not a bug in that business model. For the people doing the selling, it is closer to the point. You stay stuck, and you keep paying.
Where I would push back, gently
Training is the missing 95%, but it is not a magic wand, and it would be dishonest to pretend otherwise. Teaching your team to prompt brilliantly does nothing if your underlying data is a swamp, your processes are broken, or your incentives reward nobody for actually changing how they work. The learning gap MIT describes is organisational as much as individual. Training fixes the skills. It does not fix a company that does not really want to change. Do both, or you will just have well-trained people using AI to produce the same mediocre output faster.
Still, the core reframe holds, and it is the most useful thing you will hear about corporate AI this year. Stop treating AI adoption as a procurement problem to be solved with a chequebook, and start treating it as a skills problem to be solved with time and attention. The companies in that lonely 5% did not buy better robots. They taught their people. Credit to Angus for putting it so plainly.
The AI tool stack actually worth paying for
One email a week. The tools, models and moves that matter, minus the hype and with the horseshit filter set to maximum. Free.
Did you know: the same MIT research found that the successful 5% were not the ones with the fanciest models. They tended to pick tools aimed at one specific workflow, and, crucially, let the people who actually do the work lead the rollout, rather than handing it to a central “innovation” team who would never have to live with the result.