I wrote recently about Desire Paths (the worn track across the grass that appears because the paved route is too much of a faff for actual people); the argument was that a workaround is feedback. What people actually do, with or without official sanction, is definitely something to consider, and maybe the smart thing is to roll with the punches, and enable the work-arounds, rather than fight against them.

So then if you look at The GenAI Divide from Project NANDA at MIT Media Lab (you might have seen it quoted, because it is where the “95% of AI projects fail” line comes from, which is quotable but not quite correct - more on that below). There is something else in the same report, that isn’t being quoted, and I think is at least as interesting, maybe more (section 3.3 if you want to read it in full).

Roughly 40% of the companies they looked at had bought an AI subscription. Employees at over 90% of the companies they surveyed were using AI for work anyway, on personal accounts, in a lot of cases without IT knowing anything about it. Those people reported getting more out of their own tools than out of their employer’s official programme, which was typically still sitting in pilot.

Very much “Desire Path”, or to put another way, while you were planning a Formal Event in the Ballroom, people couldn’t wait and started their own, much more wild and fun party in the kitchen.

If you weren’t sure where your AI adoption programme was … the party has maybe started, and likely they didn’t ask you, on purpose (with your rules, keep the music down, don’t drink too much, have fun but stay safe, and other sensible stuff, you square!).

So it’s all gone rogue?

Think about what it takes to get to that point where people are using their own “home brew” AI. I STRONGLY suspect many of you who bother to read this have played with AI, installed AI components and/or paid AI subscriptions, applied that to “personal projects” … and have preference(s) (I know I do).

But what this is saying is that actually most people have decided that the fastest way to do their job was to open a personal account, on their own time, with no training, no rollout, no comms plan and no steering group! Great, up to a point, as no plan, implies no, or limited guardrails (data security, no use-case, no cost/benefit analysis).

I heard a 2nd hand story from a CTO recently (and we both think it’s probably apocryphal), that a developer asked <insert coding AI here> to redevelop an SME’s entire backend on Friday, came in Monday and it had! But it had cost >£200k in tokens.

Lock it all down then?

Yes you have to be careful, but also this is absolutely gold requirements-gathering, and it costs you nothing. When they spend their own effort to get around your process, they have told you precisely which capability was worth having and roughly what it was worth to them … and when they vote with their own tokens and by extension cash … you have to take them seriously!

The report’s own numbers make the same point from the other direction. The tools that got adopted were the general-purpose ones people could simply pick up and use. The custom, vendor-pitched enterprise builds were the ones that stalled, and the users interviewed described them as brittle, overengineered, or badly matched to how the work actually gets done (how shocking, all classic symptoms of disconnected programmes AI or otherwise, but that’s another post!).

Which means the official programme was busy solving a problem the organisation did not have, while the problem it did have got solved for £20 a month by people who never told anyone (because they could get their work done better, and so much quicker they could have a full hour for lunch every day and disappear at 3pm on a Friday).

The bit worth being careful about

Since I have quoted this report, I should say what it is. It is version 0.1, labelled preliminary findings. It draws on interviews with 52 organisations and survey responses from 153 senior leaders gathered at four conferences. The authors say plainly that their figures are “directionally accurate based on individual interviews rather than official company reporting”. It is not a peer-reviewed paper. And MIT took it down: the PDF was still being served from the NANDA site early on the afternoon of 18 August 2025 and was returning a 404 by that evening, within hours of the coverage going viral. It has never appeared on the group’s publications list. The link above goes to an archived copy of MIT’s own file, because that is the only place it still exists.

That matters for the famous number. The report says 95% of organisations reported no measurable return. It does not say 95% of pilots fail. Its own chart shows general-purpose AI tools going from 50% piloted to 40% implemented, and it notes chatbot pilot-to-implementation rates of around 83%. The 5% figure comes from one narrow category, custom embedded tools, where 20% reached pilot and 5% reached production.

But, I think this seems to confirm what we all kind of know … which is most organisations are not getting a measurable return, and most custom builds are struggling, while the off-the-shelf tools people actually chose are landing fine. That is a considerably less dramatic sentence, and it happens to be the one supported by the data.

What I would go and do

If you are accountable for an AI programme, I would want to know one thing before the next steering group. What are people in this organisation already using, and what are they using it for?

Not so you can stop them. So you can read the answer. You are holding a live experiment that ran for a year, across your actual workforce, on your actual work, with no budget and no business case, and it has produced a clear result. Almost nobody is looking at it, because it happened somewhere the governance cannot see.

Go and look at the path. It has been worn into the grass for a while now.

There is a harder question sitting underneath this one, which is what you do about it when the data going through those personal accounts belongs to patients, account holders, your staff or customers.

That one needs its own post!