Ask people at almost any company whether they use AI for work, and the honest answer is usually yes — whatever the official policy says. MIT researchers who studied enterprise AI adoption this year found that employees at more than 90% of the companies they surveyed were already using personal ChatGPT or Claude accounts for work, regularly, some of them every day — while only 40% of those companies had bought so much as one official AI subscription. The organisation didn’t approve any of it. It just didn’t get in the way fast enough to stop it.
That gap between what a business has sanctioned and what its people have already gone and done for themselves, is the real story behind a number that’s become almost a cliché this year: roughly 95% of generative AI pilots fail to deliver any measurable return, against $30 to 40 billion in combined enterprise spending. The same research found something more specific underneath that headline figure. Pilots built on generic, off-the-shelf tools reach implementation 83% of the time, because they ask almost nothing of the organisation around them. Custom pilots — the ones meant to touch a real process — mostly never get past the pilot stage at all.
What the pilots are actually funding
Look at where the $30–40 billion has gone and a pattern appears. Most of it has gone into things that generate enthusiasm rather than change: a chatbot here, a summarisation tool there, a demo for the board that photographs well in a slide deck. Each one is cheap to start and, when it doesn’t work, cheap to quietly abandon.
The harder work doesn’t get funded at the same rate: auditing whether the underlying data can actually be trusted, naming the political reason a broken process has never been fixed, admitting a failure that predates the technology by years. None of that demos well, none of it fits inside a single budget cycle, and none of it feels like progress in the moment — even though it’s the only work that would tell you whether a tool could ever have helped in the first place. So it keeps not happening, and the pilots keep multiplying instead.
The mirror, not the tool
AI doesn’t reveal what’s wrong with a business. It runs straight into it. A process that was already unclear stays unclear, just faster. Data that was already messy gets amplified rather than cleaned up, because nothing about deploying a model answers the question nobody asked about where that data came from or whether it was ever trustworthy in the first place.
The technology isn’t lying about the organisation it’s dropped into. It’s showing that organisation exactly what was already there — at a speed that makes the thing much harder to keep ignoring.
The alternative isn’t complicated
None of the harder work requires new technology. Deciding who owns a piece of data, and whether it can actually be trusted, is a decision an organisation could have made before any pilot started. Deciding who’s accountable when a model gets something wrong is a governance question, not an engineering one. Giving staff an approved tool that’s as fast and as unrestricted as the one they’ve already found for themselves is a procurement decision, not a research problem.
What makes this work hard isn’t complexity. It’s that it requires someone to admit, out loud, that the process was already broken before the model arrived — and admitting that is exactly what a shiny new pilot lets everyone avoid doing.
Who is left holding the gap
The employees quietly running work through their own AI accounts aren’t doing it to make a point. They’re doing it because the sanctioned tool doesn’t do the job and the deadline hasn’t moved. That leaves them carrying two things nobody chose to give them: the actual risk of putting company information through a personal, unmanaged account, and the actual judgement calls — what’s fine to paste in, what isn’t — that no governance process has ever talked them through.
None of that is a technology failure. It’s a business that outsourced a hard decision to whoever was closest to the deadline, then called the resulting mess an AI adoption problem.
A better model won’t fix what a business has never been willing to look at directly. It will just show that back, more clearly, and faster than before — and leave someone else to carry what it finds.
You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.
Neil Catton is the author of The Next Evolution, The Cognitive Crucible and The Shadow System - available on Amazon, and writes at the intersection of technology, ethics, and human purpose.
Really like this framing, The Next Evolution that highlights the glut of fragmented AI pilots that seeing industry talk about. I’d add that AI doesn’t necessarily have to wait until the foundations are fully enabled.
To scale AI at the speed of the business, you ultimately need the right control and data layers around it, but AI can also provide a relatively low-cost mirror into the size and shape of those underlying issues, while helping build those foundations.
That’s where I think an MVP approach is powerful: take a thin slice of the application, data and control layers, prove value, expose the gaps and then iterate.
Rather than a big-bang data/AI transformation before doing anything useful or a shallow use case pilot, use the first use cases to build the foundations at the same time. The AI becomes both the catalyst for change and a way of measuring how ready the organisation really is.