A few weeks after I published “The Frictionless Trap”, a colleague stopped me at the coffee machine. He was describing a fatigue he couldn’t quite locate. Getting genuine value from AI — Copilot for emails, ChatGPT for working through problems — but drowning in the volume of responses coming back.
We traced the pattern. People around him were using AI to write thorough emails and reports. He was using AI to compress them before reading, because there wasn’t enough bandwidth otherwise. We were spending tokens inflating content, then spending more to compress it back down, the core information transfer unchanged. Only the AI providers were winning.
My first instinct was: stop using it. But the inefficiency wasn’t really the problem.
That question is what this article is about.
I’m not writing this as a sceptic. I spend my working life helping organisations apply AI and I believe in what it can do. The question I keep coming back to is whether we’re thinking carefully enough about where.
A few weeks after I published “The Frictionless Trap”, a colleague stopped me at the coffee machine. He was describing a fatigue he couldn’t quite locate. Getting genuine value from AI — Copilot for emails, ChatGPT for working through problems — but drowning in the volume of responses coming back.
We traced the pattern. People around him were using AI to write thorough emails and reports. He was using AI to compress them before reading, because there wasn’t enough bandwidth otherwise. We were spending tokens inflating content, then spending more to compress it back down, the core information transfer unchanged. Only the AI providers were winning.
My first instinct was: stop using it. But the inefficiency wasn’t really the problem.
That question is what this article is about.
I’m not writing this as a sceptic. I spend my working life helping organisations apply AI and I believe in what it can do. The question I keep coming back to is whether we’re thinking carefully enough about where.
The early anxiety was that AI would eliminate entry-level roles — the foundational work done faster and cheaper, leaving no space for people who learn by doing. Then the story shifted: AI as a harness, a way for someone earlier in a domain to engage more equally with people who’ve been in it longer. Both stories are about output. Neither asks what people are actually developing along the way.
What the ladder actually does
Michael Polanyi, philosopher and chemist, spent much of his career on a deceptively simple question: how do people actually come to know things? His answer: we know more than we can tell. The surgeon’s hands know things that can’t be put in a manual. The chef who tastes a dish and knows, without checking anything, that it needs another minute — that knowledge lives in accumulated experience. It has to be earned through doing things, getting them wrong, and doing them again.
As Amazon’s CEO Andy Jassy once put it: “There is no compression algorithm for experience. You can’t learn certain lessons without going through the curve.”
Polanyi called this tacit knowledge. For most of human history it moved through apprenticeship — you worked alongside someone who had it, made mistakes in front of them, and gradually built the internal reference points that let you recognise what good actually felt and looked like. We mechanised physical work, then cognitive work. Each time, people moved into more generative roles and learned through the same process of doing and struggling alongside others who had already done it.
What is different now is that we risk becoming apprentices to AI before we have the skills to evaluate what it is teaching us. The tool has no mastery to transmit — only patterns extracted from people who did. If you have never done the work yourself, you cannot tell the difference.
How the ladder breaks
In any knowledge domain, expertise develops through doing the hard version of the work — the version that resists, that builds the feel for what’s right and what’s off in ways that approving someone else’s output simply can’t replicate.
AI breaks this by absorbing the foundational work entirely. Someone who might have spent months on the hard analytical version of a problem can now produce polished synthesis in an afternoon. The struggle that would have built their calibration has been skipped — not by design, just because the tool was there and the output looked convincing. And the person who used to create those conditions for others, by assigning that work and watching someone push through it, no longer has the same work to give. The development that used to be a byproduct of getting things done stops.
What you end up with is people across every field producing the appearance of competence while the foundation quietly stays unbuilt — in legal reasoning, clinical judgment, research synthesis, strategic thinking. We are all going to feel confident in our own ignorance. That is the Emperor’s New Clothes version of AI adoption — and unlike the original story, there may not be a child left in the room who can see it.
What strategic AI leadership actually looks like
The organisations that will get the most from AI are not the ones that apply it most broadly. They’re the ones that know what to protect.
A few concrete things that matter:
Map the work before automating it. Not all foundational work needs to be preserved — some is genuinely commodity. But some is where people build the calibration they’ll need to evaluate AI output later. That distinction belongs in transformation planning from day one, not as an afterthought.
Build a skills strategy alongside your AI strategy. Operational leaders should be perform Strategic Workforce Planning and be asking: which skills are needed now, which need to be deliberately developed as AI takes on adjacent tasks, and which can reasonably be deprecated? Most organisations have an AI roadmap. Few have a parallel human skills roadmap. They belong together.
Rebuild apprenticeship deliberately. The informal transmission of expertise that used to happen through delegation no longer happens automatically. Identify who carries tacit knowledge, what experiences build it, and how those get created even when AI could produce the output faster.
If the capability lives in the AI and not in the people, you don’t have a capable organisation. You have a dependency!
Go back to the coffee machine conversation.
The missing skill isn’t prompt engineering. It’s the ability to write a message clear enough that it doesn’t need to be compressed, and to read with enough judgment to notice when the compression has dropped something. Both develop through practice — through the actual work of writing and reading, of making your thinking readable to another person and working to understand theirs.
The harness and the broken ladder run together — in the same person, on the same workflow. Someone earlier in a domain uses AI to produce a legal brief that looks argued, a research synthesis that reads as considered, a clinical summary with the right structure. The harness delivers. But the process of actually arguing the reasoning, of sitting with ambiguous data until something resolved, of making the structural decisions that would have built calibration — that work was routed around. They were harnessed before they were apprenticed.
What surfaces later isn’t an inability to produce. It’s an inability to evaluate — to catch the plausible error, the dropped assumption, the conclusion that sounds right but isn’t. That judgment is built through exactly the foundational work the harness replaced.
The leaders who build something durable from AI won’t be the ones who found the most things to automate. They’ll be the ones who thought carefully about which work belongs to the tool and which belongs to the person.
The Frictionless Trap asked what this costs individually. This is the same question for anyone leading a transformation. What are you optimising for, and is competence on the list?
This article follows The Frictionless Trap (June 2026).