AI adoption is not a rollout

About ten years ago I (eventually…see the end) bought a book for 50p from the kind of charity shop that keeps its business titles on the bottom shelf, spines out, where nobody looks. I was deep in a particular obsession at the time. I wanted to understand how you actually design an organisation so that real innovation happens inside it, rather than getting talked about in off-sites and then quietly dying to never be talked about again. I still remember fondly all my chats with a close colleague Tom, who to this day we can get into these topics for hours.

So we’d been reading about constraints. The counterintuitive stuff. The idea that creativity needs limits, that a blank page kills more ideas than it frees, that the right box makes people think harder than no box at all. We’d been reading about moonshots and forcing functions, about teams told to make something ten times better rather than ten percent better precisely because ten percent lets you keep your existing assumptions and ten times does not. I had an process consultants’ view of the problem at the time. Find the right inputs, apply the right process, get the output.

The book was Creative Ecologies by John Howkins. It was the subtitle that drew me in…

“Where Thinking Is A Proper Job”

I expected it to give me another lever to pull. It did something more annoying than that. It told me my mental model was wrong.

Howkins’ argument cut straight across the instinct I’d brought to the whole question. You cannot build innovation the way you build a process, he said. You can’t install it. You can only grow it, and you grow it by getting the conditions right and then living with whatever those conditions produce, which is never quite what you planned. I half-agreed, found it a bit unsatisfying as a thing to actually do on a Monday, and put it on a shelf.

It stayed with me anyway. Through every role I’ve had. And now, watching company after company try to absorb AI, I keep reaching back for it again. Because it describes what is happening in front of me far better than anything written in the last few years.

What adoption actually looks like

Walk into almost any organisation right now and look closely at how AI is really being used, not how the all-hands slide says it’s being used, and you find the same thing. It is uneven, it is patchy, and it will not hold still.

A handful of people have quietly rewired how they work. They’ve gone past the novelty and built AI into the texture of an ordinary day, and they can now do things that genuinely took them hours before, or things they frankly couldn’t do at all. TBH I was doing 4 jobs at once recently, and that was only possible because of AI fundamentally allowing me to do things I couldn’t do before. Some of those people pass on what they’ve found. Most don’t. Not out of any meanness. There’s just no mechanism that asks them to and no particular reason it would occur to them.

Then there’s a much larger group who are stuck. Some of them don’t know where to begin and are slightly embarrassed about it, so they say nothing.

But it’s ok as they will be shown how to write a good prompt in a workshop tomorrow…

Others have done the obedient, literal thing. They opened the company-approved tool, used it for the task on their desk, got a mediocre result, and concluded the whole thing is overhyped. Quite often the tool was wrong for that particular job and nobody told them, because telling, or more importantly, showing them, was nobody’s job.

You can usually find the wrong-tool problem in about five minutes if you ask, or even one minute if you take a look. Someone is pasting a long document into a chat window, asking for a summary, getting something bland and generic back, and sharing it to someone who is pasting into a chat to work out what it was in the first place to send a chat-generated response back to the person that sent it (ad infinitum). The real issue is that they reached for the one thing IT put in front of them and it was never built for the outcome they really need and it’s highly unlikely they’ve been given the time, tools and guidance to be allowed to fundamentally rethink out they achieve that outcome. Meanwhile, three floors down, someone has worked out exactly how to do that same task well, maybe in their own time, and has told nobody, because nobody asked and there was nowhere obvious to say it. Both of those people work for you. Neither of them knows the other exists.

And the whole picture moves. The person who is ahead in March is behind by June, because someone two desks over found a better way of working and they never heard about it. There is no stable map of who is good at this. There’s a shifting, living thing that looks different every time you check.

Now look at how most companies respond to that. They pick a platform. They write an acceptable-use policy. They book some training. They buy a block of licences. They track adoption, by which they mean logins and seats and “monthly active users.” and the increasing usage, maybe even proudly with a leaderboard. They put a number on a slide and call it progress.

This is the rollout model, and for an enormous range of things may be correct. It treats the organisation as a machine: define the new component, install it, confirm it’s running, move on.

The problem is that AI is not a component you install. It doesn’t change what people do so much as what they are capable of, and capability does not travel through a rollout plan. It travels through people, socially, unevenly, and on a timetable nobody controls.

You can mandate that everyone has access to the tool. You cannot mandate that they become good at thinking with it.

Those are different categories of change, and we keep using the method for the first one on the second one and wondering why the slide says 87% and the work looks the same.

The book

Here is what Howkins actually did, stripped of the academic packaging as I would never think of myself as an academic even if I enjoy these sort of books (I don’t think a certificate in Jazz counts…)

He took the word ecology literally. Ecology is the study of how living things relate to their environment. He applied that to creativity and innovation, and his claim was that creative output doesn’t come from gifted individuals working alone, and it doesn’t come from structure pushed down from the top. It comes from conditions. He kept returning to four of them: diversity, change, learning, and adaptation.

His definition of a creative ecology was, roughly, a habitat where different sorts of people use ideas to make new ideas, and where others back that effort even when they don’t fully understand it. The thing that counts, he said, is the relationships and the activity, not the infrastructure.

You don’t get there by buying better equipment. You get there by the right people doing the right kind of work next to each other.

He argued a creative ecology needs three things to exist at all: talent, freedom, and markets. And his central, awkward point, the one that makes the book genuinely hard to act on, was that you cannot command any of it. You can’t order people to be creative. You can only tend the conditions and watch what grows.

Ten years ago I read that as a sharp idea about org design that didn’t quite tell me what to do. How do I create with the skills I had at the time create a culture of innovation? A diagnosis without much of a prescription. I’m reading it now as something closer to a field guide for the single biggest change moving through every business I walk into or talk to anyone about.

What it gets right about AI

Lay the principles against what’s actually happening and they line up almost uncomfortably well.

Take diversity. The thing that looks like chaos, the finance analyst and the warehouse supervisor and the junior in marketing all poking at AI in completely different ways for completely different jobs, is not a mess waiting to be standardised. It is the engine. The variety is precisely where the useful discoveries come from, because no central team would ever have thought to try half of what those people are trying. The strong managerial instinct here is to consolidate. One approved tool, one approved use case, one tidy story for the board. That instinct is the thing most likely to kill the ecology, because monocultures are efficient and brittle, and the moment the approved use case stops being the interesting one you’ve designed yourself a dead end.

Then there’s adaptation rather than implementation. There is no finished state. The tools change every few weeks. What was impressive in spring is ordinary by autumn. A rollout has an end. You’ve deployed it, you’re done, you close the project. An ecology never closes. A product doesn’t have an end. If you are treating AI adoption as a project with a completion date in it, you have already misread the situation.

This is genuinely hard for organisations, because organisations like closure. They like the project that ends, the box that gets ticked, the thing you can finally stop thinking about. AI offers none of that, and the discomfort of that open-endedness is part of why so many leaders reach for the rollout shape in the first place. It at least promises an ending. The ending is a fiction, but it’s a comforting one, and a comforting fiction will beat an uncomfortable truth in most meetings.

Then the compounding, which is the bit Howkins cared about most. Ideas making ideas. One person automates a slice of their reporting. A colleague sees it and realises the same move works on something else entirely. A third person watches that and asks a question that takes the whole thing somewhere neither of the first two would have reached on their own. That chain is worth more than any single clever use of the tool. But it only fires if the ecology is connected. In most companies it isn’t, so the chain snaps at the first link and the discovery dies in the head of the person who made it, useful to exactly one human being.

And place, which Howkins insisted still mattered even in a networked world, and he was right. The fastest learning happens when someone can lean over and show you the actual thing on their actual screen. A channel called #ai-tips is not the same and everyone knows it’s not the same, because they’ve all got one and none of them read it.

Put those four together and you get a fairly brutal conclusion. The single biggest determinant of how much value your company gets out of AI is not the tool you choose, the budget you set, or the policy you write. It’s whether the conditions let capability spread or trap it where it lands. We could be talking generally about innovation and culture, but lets stick to AI adoption for now. Most companies are spending almost all of their attention on the first three things and almost none on the fourth, which is close to the exact inverse of where that attention would actually do some good.

What it gets wrong

This is where I personally wanted more from the book, and it’s the part I’d ask you to sit with, because it’s the part that turns all of this from an interesting lens into something you can actually do.

Howkins makes an ecology sound very nearly self-organising. Set the conditions, step back, let nature take its course. It’s a beautiful idea and it is where the book quietly stops being useful, because an ecology left completely alone does not reliably head towards health. It heads towards whatever is most aggressive.

In a real organisation that means specific, recognisable things. The loud, confident person doing the wrong thing at volume becomes the default, because confidence spreads through a company faster than correctness ever does. The person quietly hoarding what they’ve learned stays ahead of everyone else, which teaches the rest of the building that hoarding is the smart play. One team’s preferred tool becomes the de facto standard, not because anyone decided it was best but because it got there first and nobody pushed back. None of that is the conditions failing. That is what the conditions produce when no one is paying attention.

Anyone who has actually kept a garden knows the difference between a garden and a patch of ground you’ve stopped looking at, and the difference is not the seeds. It’s attention. Someone deciding, week after week, what to encourage and what to pull out before it takes over everything else. Left to itself, a garden does not tend towards the plants you wanted. It tends towards the ones that are best at taking over. That holds for an organisation under quarterly pressure at least as firmly as it holds for soil.

In a company, that someone is a leader. And more often than not it’s the product and technology leadership, because they are sitting at the exact junction where people, tools and commercial outcomes meet, which is the only vantage point from which the whole ecology is even visible. This is the job Howkins didn’t write about. Fair enough. He was describing creative scenes and cities, not an organisation with a P&L and a board asking what the AI plan is by Thursday. The conditions still matter, enormously. They are just not sufficient on their own, and pretending they are is how you end up with brambles.

I should say I’m not the first person to land here, and I’d rather tell you that than pretend the idea is mine which it obviously isn’t There’s a whole field, complexity theory, built on the difference between problems you can analyse your way through and problems you can’t. Dave Snowden’s work and his Cynefin Framework which I discovered as an Agile Coach (yes I know…but I hopefully was one of the good ones) is the part of it I keep coming back to. His basic move is to separate the situations where an expert can study the thing and hand you the right answer from the situations where cause and effect only line up in hindsight, so the only honest way forward is to try something small, watch what actually happens, and respond to that rather than to your plan. AI adoption is squarely the second kind. The rollout treats it as the first. That’s the whole mistake in one sentence.

What’s useful about it is that it gives the gardener something to actually do. You don’t steer a system like this by setting targets and chasing outcomes, because you can’t control the outcomes. You steer it by managing constraints. Some constraints dictate: the rule, the policy, the approved tool. Some just shape the space without dictating: a boundary, a rough heuristic, fifteen minutes at the front of a meeting. The job is mostly the second kind. Which, for me, was the moment two old obsessions turned out to be one obsession, because that is exactly what I’d been chasing years earlier with constraints and moonshots, without knowing it had a name.

Tending

So what does tending actually involve, if it isn’t writing a policy and isn’t standing back and trusting the process?

Go and look, first. (Lean practitioners reading this, are now screaming ‘go to the gemba’ like they have been for years, quite rightly too: Go See, Ask Why, and Show Respect.) Most leaders genuinely do not know who in their organisation has changed how they work, because they’re measuring the wrong thing. Seats, usage and logins tell you who has access and maybe using it. And not even how they are using it. They tell you nothing about who has changed. The only way to find out is unglamorous: watch the work, and talk to people about how they’re doing it. Who is producing things now that they couldn’t three months ago? What is spreading from desk to desk without anyone’s sign-off? That unsanctioned, slightly-against-the-rules activity is usually where the real adoption lives, and most governance instincts are quietly designed to stamp it out before anyone learns from it.

Move what people find, second. Capability stuck in one person’s head is worth nothing to the company, however good it is. The job is to get it out and around. Not by asking people to write it up in a system no one opens. By putting the person who cracked something in front of the people who haven’t, doing it again next week, and making that a normal and slightly celebrated thing rather than a favour you had to ask for. The unit that actually spreads is one person showing another person. It has never once been a PDF, or unfortuntaely even for AI, an .md file.

The cheapest version of this I’ve seen work is fifteen minutes bolted onto the front of a meeting that already exists. One person shows the single most useful thing they did in the last fortnight, live, on their own screen, and takes two questions on it. No deck, no write-up, no programme with a name. The cost is fifteen minutes. The effect, repeated over a few months, is that the gap between your fastest people and everyone else stops widening and quietly starts to close, which is the only direction that gap has ever needed to move.

Cover the people taking the risks, third. Everyone experimenting is exposed. They might burn time, look foolish in front of colleagues, or break a process someone more senior cares about. If that exposure isn’t covered by someone with the authority to cover it, two predictable things happen. The cautious ones stop, sensibly. And the bold ones carry on underground, where you can no longer see what they’re learning or where they are doing it which brings it’s own problems with ‘grey work’ and security adn data risks. Howkins listed freedom as a condition for the whole thing to exist. Inside an organisation freedom is not in the air for free. It is granted and defended by a named person, or it isn’t there at all.

And have a view, fourth, which is the part the gardening metaphor most wants you to dodge because gardening sounds so gentle. It isn’t, entirely.

Some of what’s growing is the confident, wrong use of the wrong tool, and left to itself it spreads as happily as anything good does.

Part of tending is judgment. Backing the experiments that are clearly going somewhere, delivering outcomes that weren’t possible previously and steering people off the ones that aren’t, including the very common dead end of reaching for the approved tool to do a job it’s bad at. (If I see one more basic workflow that is just a few steps of automation that is now a fully agentic flow with LLM, I think I’m going to go mad). That isn’t control. It’s direction. There’s a real difference, and the difference is the whole skill.

A change of verb

The reframe I’d leave you with is small and it’s a change of verb.

You are not running an AI rollout. You are tending an AI ecology.

Rollout is finite, mechanical, and mostly about access. Tending is continuous, social, opinionated, and about capability. One of them produces a slide that says 87% adoption and a vague sense that you’ve done the thing. The other produces a company that can demonstrably do more than it could six months ago, which, when the noise dies down, was always going to be the only number that mattered.

This is the same gap I spend most of my working life pointing at, in a new outfit. A company can be extremely busy with “insert thing”, in this case, AI. Licences bought, policy drafted, training booked, steering group convened, slide updated. And it can have changed almost nothing about what it is actually capable of producing. Busy is not the same as effective here any more than it is anywhere else.

This new more photogenic version is not cost-free. While you’re reporting adoption, the competitors who are tending instead of rolling out are compounding, quietly, in the way Howkins described. That gap doesn’t show up in a quarter. It shows up over a year or two, and by the time it does it isn’t a gap any more. It’s a moat, and you’re on the wrong side of it. Or more likely, drowning in it.

The book is still on my shelf. The lines I underlined ten years ago, about conditions beating control, read completely differently now, because back then it was a theory about how you might design a place for innovation, and today it’s a flat description of the largest change moving through every business I walk into. Howkins got the ecology right. He just left out the gardener.

Right now, the gardener is the job. (And please, no AI Gardener titles on LinkedIn. I’m begging you.)

The most valuable thing you can do this year isn’t choosing the right tool. It’s walking the floor and actually noticing. Who’s close to flourishing and needs a little more light and a little more cover. Who’s been quietly held back by a process nobody’s looked at in two years. And who, without telling a soul, has already outgrown the whole setup you built for them.

That last person is the one I’d go and find first, because they’re usually the most valuable thing in the building and the easiest in the world to miss. The same way the most useful book I ever bought was the one nobody else had bothered to take off the bottom shelf. Fifty English pence. Which I didn’t have in cash, only card which wasn’t taken… So even had to come back another day to buy it! Spine out. Ten years of being quietly right, waiting for the rest of us to catch up.

Your version of that book is a person. They’re somewhere in your org chart, a few rows down, doing things this year that wasn’t even possible last year, and nobody senior has noticed yet. Go and find them. Then spend the rest of your time building the conditions that let everyone else become them.

That’s the job. It doesn’t end, and it was never going to fit on the slide.