Three years ago I would not have believed how much these tools would extend my own work. I read organisations for a living: structuring problems, building arguments, finding the load-bearing thing in a mess of detail. Pointed at that, AI has changed my capacity and my speed to a degree I still find faintly startling.
Pointed at visual design, where my expertise is close to nil, the same tools produce work that is confidently, fluently poor. My last attempt at an AI-assisted client graphic landed like a piece of furniture dropped from a height.
The tool had not failed. It had amplified exactly what I brought to the task, which in that domain is very little. A designer colleague, using the identical model, produces in minutes what would have taken her a day. When I try the same, I amplify mediocrity at scale. The tool is identical. The person using it is everything.
This is the most consistent observation I have from working inside client organisations: the variance in AI output quality between people in the same building, using the same tools, on the same kind of work, is far larger than anyone expects going in. And it is not explained by technical sophistication. The best prompt engineer in the building is rarely the person producing the most valuable AI-assisted work.
What explains it is depth of domain expertise, and then something more specific on top of that. Whether the person knows where their expertise ends.
Part of why this catches people out is that the entry experience is so misleading. At the trivial end of the range, AI assistance works for everyone. A shopping list, a training plan, troubleshooting your boiler. These tasks require no expertise from the user, the cost of error is low, and the value, while real, is modest. Nothing about amplification applies down there, which is precisely why the experience of using AI for trivial things is such a poor guide to what happens when the stakes rise.
That is where plenty of executives form their conviction about AI. The chief executive who had a remarkable weekend conversation with a chatbot about his golf swing and arrived on Monday convinced the technology could transform the supply chain. The tool that impressed him was operating under entirely different conditions from the tool he was about to fund.
Beyond the trivial, a different rule takes over. AI amplifies what you already have. It does not manufacture what you do not.
Which creates a trap with an unusual shape. The tool always produces something. The something always looks finished. And it will never once indicate that you are the wrong person to be evaluating it. In almost every other professional situation, working outside your competence announces itself eventually. You get stuck, or you produce something visibly thin, or somebody tells you. Here you get a polished artefact and a feeling of progress. The absence of friction is the problem.
So the discipline that matters is not prompting. It is declining to reach for the tool in domains where you have no way of judging what comes back. I am guilty as charged myself, and I still catch myself doing it, which is roughly how I know it is hard.
I should name the obvious objection, because a consultant whose strength is analysis arguing that AI rewards people who know their strengths is suspiciously convenient. Two things give me some confidence the principle generalises. The mechanism is not mysterious: these are prediction systems whose output quality is highly sensitive to the framing, the context and the ability to recognise a wrong answer, and expertise is exactly what supplies those. And the pattern repeats across every organisation I have worked in, at every level of seniority, in domains far from my own. High confidence in the mechanism. Moderate confidence in how far it travels beyond the kinds of knowledge work I can observe directly.
Something to try this week. Draw your own competence map. Two columns: work where you can reliably tell whether the output is any good, and work where you cannot. Use the tool freely in the first column. In the second, either don't, or route the output past somebody who lives in the first.
Then the uncomfortable version, which is where the actual learning is. Take a recent piece of your own AI-assisted work to a colleague who knows that domain properly, and ask them which column they would have filed it under.