You gave your team the same AI everyone raves about. Why did the results come back flat?

The demo was dazzling. The AI answered everything, drafted in seconds, made the room lean in. So you rolled it out, expecting the same lift across the business. Then the numbers came back, and they were flat. In a few places they were worse than before. The tool clearly works, you have watched it work, so the result makes no sense. You are left with a quiet, nagging question that the vendor cannot answer. Why does the same AI that dazzles in the demo do so little in my operation?

Here is the answer, and it changes where you look. The value was never sitting inside the machine, waiting to be switched on. It lives in the arrangement between your people and the tool, in how the two are set to work together. Get that arrangement right and the same AI transforms the work. Get it wrong, and the best model in the world gives you nothing, or less than nothing.

Why does the same AI help one team and sink another?

Because the tool is only half of the pair, and the half that carries the meaning is the human one. Look at what happens when researchers test this carefully. Take a 2024 randomised trial in JAMA Network Open. Doctors given GPT-4 alongside their usual resources did no better at diagnosis than doctors without it. Yet the AI on its own scored higher than the doctors did. The same model. The same task. The gain vanished, because it was bolted onto the old workflow with no thought given to how doctor and machine should actually work together.

The pattern repeats across very different settings. In a field experiment with 758 consultants, using AI lifted the quality of suitable work by around 40%. Yet on tasks that fell outside the tool's reliable range, the people using it did worse than those working alone. In a separate study of sales experts, the same AI raised results when it was tailored to how each person thinks, and pushed results below the no-AI baseline when it was imposed without regard for it. The lesson is blunt. The technology did not decide the outcome. The arrangement did. It is the same truth behind why so many AI efforts stall on the human side rather than the technical one.

What is the human actually doing in a good arrangement?

Holding the thing together. This is where it helps to name what AI is and is not. It is extraordinarily capable, and it understands nothing. It has no meaning of its own, no stake in the outcome, no sense of when it has quietly drifted off course. So when you work with it, that job falls entirely to you. You set the intent, notice when the output slides away from it, re-anchor, and redirect. Context drifts. Coherence flattens. A human has to keep restoring it, because nothing else in the arrangement will.

That work has a name in the way I think about leadership: it is holding coherence in the field, the steady human presence that keeps a fast, fluent, meaning-blind tool pointed at what actually matters. It is not a temporary flaw that better models will fix. It is a permanent feature of working with something that processes without understanding. And it is precisely the part that a naive rollout strips out, in the name of efficiency, without noticing that it has thrown away the source of the value. It is the deeper reason you cannot out-hustle the machine, and have to upgrade the human instead.

So what should a leader do differently?

Stop asking the small questions and start asking the real one. Leaders arrive asking which model to buy, whether it is clever enough yet, and which roles it can replace. Those questions keep your attention fixed on the machine. The question that actually governs your return is different: is the arrangement designed so that our people's judgement is preserved, structured and compounded, rather than bypassed? The bottleneck is no longer the technology. It is the arrangement around it. Work it in this order:

1. Design the interaction, do not bolt it on. Decide deliberately how person and tool hand work back and forth, rather than dropping AI onto the old workflow and hoping.
2. Keep the human directing. Structure the work so your people set the intent and stay able to notice drift and intervene, especially where judgement matters most.
3. Match the tool to the task, and know its edges. AI lifts some work and quietly degrades other work. Know which is which, and arrange accordingly.
4. Measure capability, not just cost. Track whether your people are getting sharper alongside the tool or slowly hollowing out. That number tells you if the arrangement is working.
5. Fit the arrangement to how your people actually think. The same tool imposed uniformly underperforms; shaped around real work and real people, it compounds.

What does this make possible?

An organisation where AI compounds what your people can do, instead of quietly flattening it. When the arrangement is designed with care, the tool carries the load it is good at, your people spend their attention on judgement and meaning, and the pairing produces work that neither could reach alone. Capability grows rather than wastes away. That is a stronger position than a business that bought the tools and changed nothing about how its people work.

Picture your operation a year on. The AI is everywhere, and it has not replaced your people's thinking; it has raised it. The rollout that once came back flat now lifts the work, because you designed the arrangement instead of hoping the machine would carry it. That is the advantage most will miss while they argue about how clever the model is: the value was never only in the machine, and never only in the person. It was always in the human capacity to hold it all together, which is exactly where your edge now lives.

About the author

Thomas Green

British technology futurist, AI keynote speaker and advisor. Thirty years across enterprise technology and AI strategy, helping leaders navigate the future of work. The futurist who died.