What does an AI-capable manager do differently?

An AI-capable manager is not the one who uses the tools most, and the evidence on this site is that the manager who pushes usage hardest is often running the team that gets faster and worse. What the capable manager does is allocate. Which work the machine may draft, which it may decide, which stays human, and which the team must keep doing unaided because that is where its judgement is made. Six habits follow from that, and each can be traced to a measured result rather than to a conference slide.

One: allocate the task before the tool#

Dell'Acqua and colleagues gave 758 consultants tasks inside and just outside a model's competence. Inside the frontier, the assisted consultants were markedly better and faster. Outside it, they were worse than consultants with no AI at all, and could not tell which side of the line they were on. The tool does not announce the frontier. The manager's first job is to say, task by task, where it sits, and that is an allocation a manager can make without understanding a single parameter of the model. It is the team-level version of the decision list at how AI decision rights should be allocated.

Two: keep some work unaided#

Shen and Tamkin had developers learn a new tool with and without AI assistance; the assisted group scored 50 per cent on a later test against 67 per cent for those who worked by hand. The output during the task looked fine in both arms. The manager who keeps a share of the work unaided, and rotates it, is protecting the team's ability to check the machine at the cost of some speed today. The manager who does not is spending that ability without knowing, which is capability debt at the team level.

Three: measure disagreement#

Kim and colleagues found that people with access to a system that was right half the time agreed with it 81 per cent of the time and scored lower than people with no system at all. A team that never disagrees with the machine is not reviewing it. The capable manager counts: of the outputs reviewed this month, how many did somebody change, and why. The number is cheap to collect and almost nobody collects it, and it alone tells a manager whether the review step is real.

Four: protect the rungs#

Brynjolfsson's customer-support study found the productivity gains concentrated in the least experienced agents, which is good news for output and a warning for development, because the agents were being handed the answers the experienced ones had learned by working out. Brynjolfsson's later work on employment found early-career workers in AI-exposed occupations about 19 per cent below trend. A manager decides which junior tasks are training and which are toil, and keeps the first even when the machine could do them. The argument in full is at the missing rungs.

Five: read the source when it matters#

Steyvers and colleagues found that a model's own confidence separated its right answers from its wrong ones far better than the confidence its readers felt, which means the person reading a summary is worse at knowing when to doubt it than the machine that wrote it. On a decision that matters, the capable manager reads the underlying document, and asks the team to, which costs an hour and is the difference between a manager who signs and one who decides. The individual version is at should I let AI summarise everything I read.

Six: treat confidence as a symptom#

Xiong and colleagues measured how far models' stated confidence matched their accuracy and found large gaps, clustering at round numbers between 80 and 100 per cent. Fluent, confident output is the default register of these systems and carries no information about correctness. The capable manager's reflex on a confident answer is to ask what would make it wrong, and to build that question into the team's habit, which is the fastest way to teach the failure mode without a training course.

What nobody has measured#

No study observes managers and relates these habits to team outcomes. Each habit is traced here to an experiment on individuals or on tasks, and the claim that a manager who adopts all six runs a better team is an inference from those, not a finding. The experiments also predate the current models, and the frontier they measured has moved, though not the fact that there is one.

Monday#

Write the team's task list and mark each: draft, decide, never, and learn-by-hand. Pick the unaided share and rotate it. Start counting disagreements. Then the manager is doing, at the scale of one team, what AI leadership asks of the organisation, and the two meet in the middle.

Key sources

• Dell'Acqua, F. et al. (2023). Navigating the Jagged Technological Frontier. Graded entry.
• Shen, J. H. and Tamkin, A. (2026). How AI Impacts Skill Formation. Graded entry.
• Kim, S. et al. (2024). I'm Not Sure, But...: Examining the Impact of Large Language Models' Uncertainty Expression. Graded entry.
• Brynjolfsson, E., Li, D. and Raymond, L. (2023). Generative AI at Work. Graded entry.
• Brynjolfsson, E. et al. (2026). Canaries in the Coal Mine? Graded entry.
• Steyvers, M. et al. (2025). What large language models know and what people think they know. Graded entry.
• Xiong, M. et al. (2024). Can LLMs Express Their Uncertainty? Graded entry.

Related SuperSkills research#

On briefing a team when AI drafts everything, what AI does to a team. On the review practice that makes disagreement visible, the shared prompt review. On automation bias, what is automation bias. On the organisational version of these six habits, AI leadership.

About this research#

Rahim Hirji is the author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company. He has run, grown, bought and advised businesses with AI in them. Findings are attributed to the studies and statements that produced them and kept separate from the interpretation. This is a living reference, reviewed and updated as significant new evidence appears.

How this research works  ·  Reviewed quarterly  ·  Found an error? Tell me and it is corrected on the page.