I Think We Got the AI Arrow Backwards

For the first few years of generative AI I talked happily about AI-empowered professionals. Give capable people AI and they become more capable: faster researchers, better writers, more productive consultants, sharper marketers.

I’ve changed my mind about which way that sentence should run.

Not because it’s wrong. AI really does make people more capable. But the sentence starts in the wrong place. It starts with the technology arriving and asks what it does to us. The more useful question is what we hand it before it does anything at all.

Take the demo everyone has already run. Five friends each open the same AI assistant — Claude or ChatGPT, say — and ask it to plan a reunion trip to Rome. They end up with five different plans, ranging from genuinely good to unusable. Much of that difference will sit on the human side of the exchange: what they asked, what they told it, what they challenged, what they accepted.

The phrase human-empowered AI is decoration unless I can say what a person contributes that the system cannot simply supply for itself. At least four things come to mind.

The first is the problem. Not the request, the problem. The model receives what you typed, which is rarely the thing you actually need solved. Someone has to make that conversion, and it isn’t a prompting technique.

The second is context that the model cannot know unless you bring it: which constraints are real and which are negotiable, what somebody said in the room, why the obvious approach failed here two years ago.

The third is a standard. A model will produce a plausible answer to almost anything. Deciding whether a plausible answer is a good one requires a picture of quality you brought with you. If you don’t have that picture, you can keep accepting fluent work indefinitely.

And the fourth is accountability. When the output goes out under your name, you own it. That cannot be delegated to the thing that produced it.

AI can help us articulate the problem, surface relevant context and challenge our standards. But none of those can simply be assumed to come from the model. They are part of what makes its contribution useful in the first place.

“The power doesn’t start with the AI. It starts with what we bring to it.”

There is a complication I have to take seriously, because it cuts against me. The 2025 Quarterly Journal of Economics study Generative AI at Work followed 5,172 customer-support agents and found the strongest gains among the less experienced and lower-skilled. AI helped them adopt some of what the stronger performers already did.

Nobody in that study had to define the problem from scratch. The work was defined, the workflow was known, and the tool was dropped into it. It worked. So the honest version of my argument is narrower than the slogan: where the problem is already well framed and quality is already well defined, you can add AI to an existing process and create real value. That is not a minor exception. It is one of the clearest ways organisations can benefit from AI.

I see organisations making much the same mistake at scale. The seductive route is familiar: buy the tools, roll out Copilot, give everyone access, find a hundred use cases, run prompt training, add an agent here and a chatbot there, total up the hours saved and announce that the organisation is becoming AI-enabled.

None of that is wrong. But it can leave one rather important question unasked: should the process we’re accelerating still exist in this form at all?

Adding AI to a process can improve it. It can also help an outdated process produce its outdated result faster.

“AI is a terrible reason to put a turbocharger on a process that no longer deserves an engine.”

McKinsey’s State of AI in 2026 found that organisations reporting the greatest value from AI were far more likely to be fundamentally redesigning workflows rather than simply adding AI to existing ones. That doesn’t establish causality, but it is consistent with the distinction between AI adoption and AI transformation.

There is, however, a stronger objection to my argument. Sometimes you cannot know what genuinely better looks like until you’ve played with the technology. Nobody sat down and specified all the things a spreadsheet should eventually be used for before building one. Search, version control and countless other technologies found important uses because people first got access to a new capability and started messing around with it. AI is no different. People discover possibilities by using it.

So start with the outcome can become bad advice if it means don’t explore until you already know what you’re looking for. I don’t think those positions actually conflict. Exploration should be capability-led. Give people the tools, let them experiment, try things whose value you cannot yet predict and see what becomes possible.

But exploration and commitment are different activities. The mistake comes when nobody goes back and asks the outcome question before turning an experiment into a workflow, an agent, a platform, a headcount decision or a budget line. A promising experiment gets hardened into a permanent process before anyone checks whether that process deserves to exist.

“Explore from the capability. Commit from the outcome.”

Even exploration depends on something human. Somebody still has to recognise that what they have stumbled onto is interesting, useful or dangerous. Judgement has not disappeared. It has simply entered earlier in the process.

Once we move from exploration to actually changing how work gets done, I would stop beginning with Where can we use AI? I would begin with What are we trying to accomplish? For whom? Why does it matter? What would genuinely better look like?

Then comes the question of what the work requires. Where is human judgement essential? Where do context, empathy, expertise or accountability decide the outcome? Where can AI contribute speed, scale, synthesis or pattern recognition? What combination produces something better than either could produce alone?

Only then does it make sense to ask what the workflow should become — and what technology belongs in it.

It sounds obvious. Yet a great deal of AI adoption still runs the other way: technology, use cases, workflow and eventually, sometimes, purpose. That is how an organisation can become extremely proficient at using AI without becoming noticeably better at creating value with it.

And this gets more important as the technology improves, not less. Access will stop being what distinguishes anyone. The models will get better. The tools will spread. Today’s remarkable capability becomes tomorrow’s ordinary infrastructure.

What stays scarce is everything around it: the problems you choose, the questions you ask, the context you hold, the standards you set, the judgement you apply and the responsibility you take for the result.

I still believe in AI-empowered professionals. AI will make good professionals dramatically more capable, and it will let less experienced people do things they couldn’t do before. That’s the effect. Human-empowered AI is where the story starts.

We didn’t get the destination wrong. We got the arrow backwards.