Part 2 of five. "AI, leadership and the human structures of work" is a series on what AI actually changes about leading people, and why those changes are choices, not inevitabilities.
A senior engineer I know approved a database migration plan last year. It was clear, well-reasoned, confidently written, and it would have taken the production system down, because one step assumed a lock the database did not actually provide. The plan came from an AI. The engineer was good. He caught it on the second read, not the first, because it read exactly like the correct plans he approves every day. That gap, between how right it looked and how wrong it was, is the whole subject of this piece.
The first piece argued that AI's cost to people's wellbeing is a leadership choice, not a technology outcome. This one is about a quieter loss: authority. When the model is a better specialist than most of your team on a Tuesday, and confidently, plausibly wrong on Wednesday, who is the expert in the room?
Start with the finding that should unsettle anyone putting AI into skilled work. In 2023, researchers at Harvard Business School and Boston Consulting Group ran a field experiment with several hundred management consultants. On tasks that sat inside the AI's capability, the consultants using it produced work rated around 40% higher in quality, finished roughly 25% faster, and completed more of it. On a task designed to sit just outside that capability, the pattern flipped: the ones using AI were more likely to reach the wrong answer than the ones without it. Not a little wrong. Fluently, confidently wrong, because the model produced something that looked exactly as authoritative as its correct work.
They called it the jagged frontier. AI is brilliant and useless in a pattern you cannot see from the outside, and the two sit right next to each other. The same tool that makes your team look expert on one task quietly makes them look expert while being wrong on the adjacent one. The line usually attributed to Mark Twain fits it better than anything written since: it ain't what you don't know that gets you into trouble, it's what you know for sure that just ain't so. AI is a confident-wrong machine, and confidence is the one signal humans are worst at ignoring.
"Is AI the specialist now?" is the wrong question
Watching a model out-perform your people on a narrow task, it is tempting to conclude that the AI is now the specialist and the humans are drifting into generalists who wear many hats. That gets the shape of it backwards.
A specialist you can rely on. Their expertise has edges you understand. You know roughly what they know, and where their competence stops, which is exactly what lets you trust them. The jagged frontier means AI has no reliable edges. It is a generalist that performs like a specialist in unpredictable patches. Treating it as "the authority" is not delegation to an expert. It is trusting a colleague who is sometimes brilliant, sometimes bluffing, and never tells you which.
A second body of evidence sharpens this. A 2025 meta-analysis in Psychological Bulletin reconciled years of contradictory findings about whether people trust or distrust AI, and the answer was: it depends on the task. People tend to appreciate AI on objective, capability-heavy work and resist it on subjective, personal work. Which means a blanket policy, "use AI" or "don't trust AI", is wrong almost by definition, because the right answer changes task to task, sometimes sentence to sentence.
So the authority in the room is not the model. And it is no longer, on its own, the person who knows the most about the subject. It is the person who knows where the model can be trusted and where it cannot. That is a genuinely new competence, and it is not the same as domain expertise. You can be the strongest engineer on the team and still wave through a plausible, wrong answer, because knowing a domain and knowing the shape of a model's blind spots are different skills. My engineer happened to have both. Not everyone will.
Garry Kasparov, who lost to a computer and then spent years studying human-machine teams, distilled it into a law: a weaker player with a better process, working with a machine, beats a stronger player with a worse process. The edge was never the human or the machine on its own. It was the process binding them, the judgment about who does what and when to override. That process is the authority now, and it lives in a person, not the model.
The part the productivity numbers hide
There is a second finding in that study every leader should hold onto. The consultants who gained the most from AI were the ones who started out weakest; the bottom half of performers improved far more than the top. AI is a leveller. It pulls the floor up.
Read quickly, that sounds like good news, and in the short term it is. Read slowly, it carries a warning. If AI makes your least experienced people produce work that looks senior, you lose the signal you used to manage by. You can no longer read competence from output, because the output has been levelled. The junior who genuinely understands and the junior who prompted well now hand you the same document. The tell you have relied on your whole career, good work means a good developer, quietly stops being true.
The counter-argument, taken seriously
The obvious objection: the models keep getting better, so will the frontier not just fill in and this whole problem solve itself? No, and it is worth being precise about why. The frontier moves, but it does not stop being jagged. Every jump in capability opens a new set of adjacent tasks the model now attempts and gets subtly wrong, because it attempts them with exactly the same confidence it brings to the ones it has mastered. A more capable model is a more capable confident-wrong machine on the new edge. And the signal-loss problem, the fact that you can no longer read competence from levelled output, does not improve as the model improves. It gets worse.
What this asks of a leader
Stop asking your people to trust or distrust AI as a blanket policy. Both are wrong on a jagged frontier. Ask them instead to build, and to show you, calibrated judgment: where they lean on the model, where they check it, and how they know the difference. Reward the engineer who catches the plausible, wrong answer over the one who ships fastest. Make "here is where this could be wrong, and here is how I checked" a first-class contribution rather than friction.
And notice what this quietly costs. Every task you hand entirely to the model is a task your people stop practising. The expertise that let my engineer catch the migration bug was built by doing that work himself, badly and then well, for years, the same work you are now tempted to automate away. Which raises a question worth leaving open: if we stop building expertise in our people, where does the next generation of people who can supervise the machine come from? Hold that one. The series comes back to it, and the answer is not comfortable.
For now, the nearer question. If judgment about the machine is the new authority, how do you actually run a team when half its "roles" are no longer people but agents you spin up on demand? That is next week.
The series: AI, leadership and the human structures of work
- The psychological cost of AI is a leadership choice, not a technology outcome
- Who's the authority now? Leading in the age of the jagged generalist
- Managing a team of agents: leadership when roles become software
- Org design for AI: why your Centre of Excellence becomes a bottleneck
- Cutting juniors is a choice, not an AI inevitability
You are reading part 2. Links added as each publishes.
Written by Richard Atkins.
Sources: Dell'Acqua et al. (2023), "Navigating the Jagged Technological Frontier" (Harvard/BCG) — https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4573321 · Mollick, "Centaurs and Cyborgs on the Jagged Frontier" — https://www.oneusefulthing.org/p/centaurs-and-cyborgs-on-the-jagged · Qin et al. (2025), "AI aversion or appreciation? A capability-personalization framework", Psychological Bulletin · Garry Kasparov, on process in human-machine teams (Kasparov's Law).
Top comments (0)