Last month, I sat in a room watching a CIO present their "AI Transformation Roadmap" to the executive team.
Real nice deck. Just lovely.
There were Gantt charts showing perfectly cascading timelines. Logos of expensive LLM vendors arrayed like fighter jets on an aircraft carrier. There was a slide — and I swear this is true — with a literal rocket ship on it, piercing through clouds labelled "Legacy Processes" toward a sun labelled "10X Productivity."
"We're deploying Copilot to 500 seats next Monday," the CIO said, practically glowing. "Conservative estimates put efficiency gains at twenty percent by Q3."
Everyone nodded. The CFO looked pleased. Someone asked about security protocols and got a satisfying answer full of words like "encrypted" and "compliant."
I smiled. I nodded. I drank the lukewarm coffee.
But my stomach was doing that familiar dropping thing. That sinking feeling when you're watching someone step on a rake in slow motion and you can't warn them because they're too excited about where they're walking to listen and they're just milliseconds away from splitting their nose open with a whack from a solid wooden handle.
Because I knew that in six months I'd be back in this room. The rocket ship slide would be gone. The mood would be funeral-grade sombre. And someone will be asking why they spent half a million bucks to help junior analysts write slightly more enthusiastic emails.
I wanted to stand up. I wanted to grab the CIO by the shoulders and say, "The AI isn't the answer. Software has never been the answer. You are the problem. Your org chart is the problem. The fact that Brenda in Compliance still prints emails IS THE PROBLEM."
But I didn't. I watched them buy the subscription. I let them schedule the all-hands kickoff. I saw them march toward the rake.
And now I'm writing this because I'm exhausted from having that same meeting.
The expensive business hallucination
We are living through what might be the greatest collective delusion in business since the dot-com boom convinced people that grocery delivery would change everything[^1]. (It did, eventually, but not in 1999, and not the way anyone thought.)
The narrative is seductive: Buy AI. Plug it into your workflows. Watch productivity go vertical. It's the same playbook we used for cloud migration, for SaaS adoption, for Slack. It's a technology upgrade. Flip the switch, pour champagne, update LinkedIn with "Exciting news!" (Please, don't do this.)
Except it's not working.
Look at the data. Despite billions of dollars being dumped into AI investment ($252 billion in 2024[^2]), global productivity growth is flat[^3]. Most enterprises are stuck in "pilot purgatory." (Translation: They built a cool demo for the board, got applause, and now it's gathering dust while everyone uses the AI tool to generate meeting summaries nobody reads.)
The disconnect is maddening because the technology is miraculous. I use it virtually every day. It writes code that would have taken me hours. It summarises Slack threads I'm too lazy to read. It crunches huge data-filled spreadsheets and finds insights that make me look smarter than I have any right to be.
So why does it fizzle the moment we try to scale it across an organisation?
Because we're treating a systemic intervention like a software update.
Coal miners and the socio-technical systems theory
A few years ago I was failing spectacularly to get a client to adopt an OKR system. Like, embarrassingly bad. The software was great. The training was solid. The business case was airtight. And yet, nobody used it.
I was complaining about this to a friend — an organisational psychologist — and she said, "Have you read the coal mining studies?"
Nope. I had not read the coal mining studies.
She sent me research from the Tavistock Institute in the 1950s[^4]. A bunch of British academics went into coal mines to study why new mining technology kept failing. Better drills, better conveyor systems, better lighting — all of it technically superior. All of it rejected by the miners.
The problem wasn't the tech. The problem was that the new systems broke the social bonds that kept miners alive. The old "shortwall" method forced men to work in pairs, watching each other's backs in the dark. Trust was everything. The new "longwall" method was more efficient on paper but isolated workers. Trust evaporated. Accidents went up. Productivity tanked.
The researchers called it Socio-Technical Systems theory. Organisations aren't just machines. They're marriages between a Technical System (the tools, the processes) and a Social System (the trust, the informal knowledge, the unwritten rules about who you ask for help when things break).
Optimise the Tech while ignoring the Social, and the system doesn't just fail. It revolts.
This is exactly what's happening with AI right now.
The Technical System is evolving at a pace that feels almost violent. We get model updates every week that make the previous generation look like pocket calculators. But the Social System? That evolves at the speed of human trust. Which is to say, glacially. With frequent reversals.
We've installed Ferrari engines inside Model T chassis and we're genuinely confused about why the car is shaking itself to pieces.
The thing nobody wants to say out loud
The reason your AI rollout is failing is because your people are scared shitless. They're sceptical. Resentful.
When the Social System feels threatened, it freezes. You announce a big AI initiative with language about "driving efficiency" and "eliminating mundane tasks." You think you're saying, "We're going to free you from boring work so you can do the fun stuff!"
What they hear is, "We're figuring out which of you are redundant."
So what happens? Shadow AI. Resistance disguised as "quality concerns." People using the tools for trivial nonsense to check the "I'm using AI" box while continuing to do the real work the old way because that's the thing that makes them feel valuable.
I worked with a creative agency last year that was bleeding money on concept development and storyboarding. We tried to introduce image generation tools — RunwayML mainly. The senior designers revolted. Not openly, because they're professionals. They just "couldn't get the style right." They spent hours tweaking prompts to prove that the AI couldn't match the human touch.
Was the tech capable? Absolutely. Was the culture ready? No way.
We had to stop. Full reset. We stopped talking about "efficiency" and started talking about "boredom." Framed the AI not as a replacement artist but as the intern who does the tedious prep work — the colour corrections, little background variations, the fifteen iterations of "can you make that part bigger?"
We had to make it psychologically safe. We had to say out loud, with witnesses: "Using AI is a sign that you're smart enough to focus on what actually matters. The people who don't use these tools are the ones we're worried about."
It worked. Not perfectly, but it worked.
Because until you address the fear, nothing else matters.
Although, the opposite problem can also exist. The people who love AI so much they use it for everything, whether it makes sense or not.
I have a close mate who's 100% guilty of this. He bragged to me that he spent just six hours building an agent to automate a task. A task he does once a week that takes maybe twelve minutes. He felt like a wizard. "I'm automating! Optimizing! I live in the future."
I thought he had wasted an afternoon.
This is what happens when you let the Technical System run wild without asking the basic question: Does this actually create value?
Companies that actually got it right
You know what Shopify and Duolingo have in common? Neither of them created a "Centre of Excellence" or hired a Chief AI Officer.
Instead, they did something simple. They got out of the way.
Shopify embedded AI experimentation into their existing "GSD" (Get Shit Done) process — their internal project management framework[^5]. Every project prototype phase now includes AI exploration. Not as a checkbox. As a genuine requirement to ask: "Could this be faster with AI?" They identified the early adopters, the developers already hacking together solutions on nights and weekends, and gave them a stage. They held internal summits where people demo'd their weird experiments. Some worked. Most didn't. Nobody got fired.
Duolingo went even simpler. "FriAIdays."[^6] Every Friday, anyone can spend time experimenting with AI tools. There's a $300 budget for API credits or subscriptions. The only rule: Try something. If it works, share it. If it fails, share that too.
This is what distributed innovation actually looks like. Not a formal program with steering committees and KPIs. Just space to tinker and permission to share.
The Social System thrives on peer learning. Someone tries something cool, it works, they show their friend. Now two people are faster. That friend shows their team. Suddenly a whole department operates differently and nobody had to sit through a mandatory webinar.
Zapier's approach is worth mentioning here — because it helps avoid the "AI for AI's sake" issue that my mate suffers from. When they went all-in on AI in 2023[^7] (Code Red, as they called it), they could have just tracked adoption rates. Instead, they focused on business impact[^8]. They track whether AI experiments transition into full workflows, not just whether people are playing with the tools. If it doesn't move a real metric — cycle time, quality scores, customer satisfaction — it doesn't survive.
They stopped asking "Can AI do this?" and started asking "Should AI do this?" That discipline is what separates AI theatre from AI impact.
It means not just asking "Can AI write this report?" but "Why do we write this report? Who reads it? If an AI writes it and an AI summarises it for the recipient, did any actual communication happen, or did we just build an expensive redundancy spiral?"
These three companies are leading at making the shift. But it only works if you kill the red tape. If every experiment requires three levels of approval and a vendor security review, nobody experiments.
They just wait for someone else to figure it out.
Leaders who get it
The companies winning right now aren't the ones with the biggest AI budgets or the fanciest models. Models are commodities. You can rent the smartest brain on the planet for twenty bucks a month.
The winners are the ones who figured out that AI is a relationship more than it is a tool.
That means hard conversations. About trust. About what work is actually valuable and what's just theatre. About whether the org chart makes any sense when information can flow instantly instead of through layers of management.
It means admitting you don't have the answers. It means promoting the quiet analyst who automated himself out of a job instead of firing him. It means killing your favourite project when the data shows it's not working. It means letting go of control.
The Socio-Technical leaders — the ones who can hold both the Technical System and the Social System in their heads at the same time — are rarer than they should be. Most executives are optimisers of one or the other. They're either tech utopians who think culture will magically adapt, or they're culture guardians who treat any technology change like a threat to the soul of the company.
The ones who thrive see that you can't optimise one without redesigning the other.
The question you should be asking
So here's where I'm supposed to give you a tidy framework. Five steps to AI transformation. A maturity model. A checklist.
I'm not going to do that.
Because the honest answer is messier than most consultants want to admit. There is no playbook. Every organisation is a different mix of technology, culture, politics, legacy systems, accumulated trauma from previous failed transformations, and emotional, irrational humans who show up every day with their fears and ambitions and unspoken resentments.
But there is one question that cuts through the noise.
It's NOT "what can this tool do?" That's the wrong question. That's the Technical System question. That's the question that leads to pilot purgatory and expensive licenses gathering dust.
The right question forces you to think about the Social System. About whether people will use it. Whether they'll trust it. Whether it makes their work more meaningful or more mundane. That question is:
"How does this tool change how people relate to their work?"
Most executives can't answer that question. That's why most AI transformations fail.
[^1]: See Webvan — A Strategic Collapse of Billion-Dollar Proportions, it's a good read. [^2]: Stanford HAI AI Index Report 2024. [^3]: OECD Economic Outlook, 2024 — Labour productivity growth across OECD countries averaged just 0.4% in 2024. [^4]: Trist, E. L. (1981). The evolution of socio-technical systems. In A. H. Van de Ven & W. F. Joyce (Eds.), Perspectives on organization design and behavior. John Wiley. [^5]: See Shopify CEO Toby Lutke's April 2025 memo. [^6]: "Duolingo CEO admits his controversial AI memo 'did not give enough context' and insists the company never laid off full-time employees", Fortune, August 18, 2025. [^7]: How Zapier rolled out AI org-wide: Our playbook to driving 89% adoption, November 2025. [^8]: What is AI transformation? How to build an AI transformation strategy for your business, November 2025.