AI won’t replace managers. It will mass-produce them. One year of daily agent work banks several years of a manager’s delegation reps - back-of-envelope
Odds are you were promoted with zero training into a job where feedback arrives a year late. Or you’re an individual contributor (IC) who almost never gets to practice delegating.
Nvidia’s chief, Microsoft execs, HBR, and the WEF all predict you’ll soon manage AI agents. Nobody has noticed the skill transfer runs the other way - agent work trains you for exactly the half of management nobody teaches.
82% of managers were never trained. They practice on you.
Management is learned mostly by experience and feedback. In the UK, 82% of managers took the job with no formal training - “accidental managers.” In a 17,000-leader sample, the average age at first leadership training was 42, years into the job for most. Practicing on live humans is the industry-standard onboarding.
A rep is concrete: define the task, define done, hand off with context, observe the result, adjust the brief. Anders Ericsson - the psychologist who spent his career studying how experts are made - called this recipe deliberate practice: repetition, immediate feedback, error correction. His blunt finding: expertise “could not be accounted for by a mere accumulation of experience.”
So: how many reps does the job give you, and how good is the feedback?
Management is a terrible classroom
Gallup’s panel of 16,442 managers puts the median span at 5-6 direct reports. Say each report gets 1-2 reviewable tasks a week: 5-6 reports × 1-2 tasks ≈ 5-10 delegations a week; × 46 weeks ≈ 250-500 delegation cycles a year (derived - nobody counts these). An IC, with zero direct reports, gets almost none: an occasional contractor brief or peer handoff, call it 10-25 cycles a year (assumption, argued from structure).
Now the feedback. Routine signal arrives in days, at the weekly 1:1. The formal signal is far slower: about half of workers get a formal review only once a year, and a quarter get one even less often. And hiring - the biggest delegation a manager ever makes - gives no verdict at all for the first year: a new hire takes 12+ months to get up to speed, so you can’t tell for a year whether your bet worked. Even then, results land confounded by markets, teammates, and luck. Hogarth calls this a wicked learning environment: feedback “poor, misleading, or even missing.” There, experience can teach false lessons and grade them as wisdom.
Scarce reps, slow signal, tangled attribution. Now run the same rep with an agent.
An agent is the same rep with the answer key
On the mechanics of delegation only, working with an agent exercises the identical elements (goal, context, definition of done, review) in a kinder classroom. In Anthropic’s ~400,000 logged Claude Code sessions, a typical delegation turn is one prompt setting off ~10 agent actions, and a session runs about four turns. Anyone running agents daily knows a turn finishes in minutes. Re-runs are near-free: issue one brief three different ways in an afternoon and diff the results. That diff is the answer key.
Two caveats. First, speed has a catch: Schmidt and Bjork found that when every attempt gets instant feedback, you start relying on the feedback instead of learning to judge the work yourself - you look better during practice than you really are. The teaching comes from attribution and repetition; speed just makes them cheap. Second, nobody grades your brief. It’s practice only if you compare the brief against the output and ask how the brief shaped the result - because a capable agent rescues a vague brief and hands you a false positive about your delegation skill. Neither domain is uniformly kind or wicked. Agents make kind practice cheap for bounded, gradable tasks, and the skills you train (defining outcomes, constraints, verification) outlive this month’s model.
Same rep, kinder conditions - so count the reps.
Run the numbers: several manager-years of reps in one year
DORA 2025 puts tech professionals’ median time working with AI at 2 hours a day. A realistic reviewable cycle is 10-20 minutes - at everyday reliability, agent task horizons run under an hour, and METR’s 320-minute horizon is a frontier maximum at 50% success, not an everyday cycle time. The arithmetic, one line at a time:
- 2 hours a day ÷ 10-20 minutes per cycle ≈ 6-12 cycles a day
- 6-12 cycles × 220 workdays ≈ 1,300-2,600 cycles a year
One counting rule applies to both sides, human and agent alike: count only delegated tasks whose result you actually review. Hallway check-ins don’t count, and neither does quick chat with an agent. Even if only half those two hours are true delegation cycles, ~650-1,300 still beats both human rows below. Back-of-envelope, as of mid-2026. Side by side:
| Role | Delegation reps per year |
|---|---|
| Daily agent user | ~1,300-2,600 |
| Manager (5-6 reports) | ~250-500 |
| Individual contributor | ~10-25 |
Divide the averages: ~1,950 ÷ ~375 ≈ 5× a manager’s reps, rising to ~10× at the generous end. Against an IC’s ~18 reps: ~100×. Which means a daily agent user banks several years of a manager’s delegation reps every year, a decade at the generous end. Opportunities, not skill: they become practice only when you review your briefs. Latency stacks on top: minutes versus days-to-a-year-plus, two to three orders of magnitude.
And this training window narrows: better agents rescue lazier briefs.
Before extrapolating, ask what this rep can train.
Half of management transfers. The other half never will.
Transfer research says skills move along shared elements; far transfer is rare. Agent work shares the coordination elements, so what transfers: brief-writing, context-giving, definitions of done, review discipline. Delegation judgment gets a weaker signal (an agent rarely pushes back or tells you the goal is wrong), yet the outcome still grades you: you see whether the goal was achieved.
The human half of management shares nothing with agent work. Kotter split management (coping with complexity) from leadership (motivating people); agent work is pure complexity-coping - no morale, no trust, no emotional labor. The claim stays at “can train”: outcome evidence doesn’t exist yet, and agent-fluent people may already have been systematic delegators.
Partial transfer is the science’s own prediction - and the same framework that predicts the gain predicts the trap. Applied to humans, fast-verification reflexes read as micromanagement: expecting perfect specs and instant compliance, then discarding and retrying. Without strong intuition about people, agent reps sharpen the wrong instinct. Mollick argues the reverse direction: management skill improves your AI use. The de-skilling worry lands elsewhere: delegation is the skill exercised here; the skill you handed off atrophies.
The first companies full of trained delegators
ICs historically bank 10-25 delegation reps a year; agent-native ICs run ~100× that. A 50-person company of daily agent users can compound org-wide delegation skill - clearer briefs and cleaner handoffs between humans too. My prediction has one condition: the freed time must flow back into human coordination. Otherwise agents become the coordination layer and people talk less. And every task sent to an agent skips a junior who would have learned by doing it. That ~100× is both the prize and the failure mode.
If you run a company, your people have been stacking management reps for months - at several manager-years per year. The question is whether anyone is reviewing the briefs.
Key takeaways
- A delegation rep: define → context → done → observe → adjust. Managers bank ~250-500 a year, ICs ~10-25, daily agent users ~1,300-2,600 (derived, back-of-envelope).
- The mechanism is a kind learning environment (near-free re-runs, reviewable results - not raw speed): one agent year ≈ several manager years of delegation reps, a decade at the generous end. Opportunities, not skill; reviewing your briefs converts them.
- Only the delegation half transfers - motivation, trust, and emotional labor have no agent analog - and applied carelessly, agent reflexes on people read as micromanagement.
- Agent-native companies can compound coordination skill nobody measures - unless agents become the coordination layer and the junior pipeline starves.
Where the numbers come from
All sources accessed 2026-07-18. Methodology note: every rep count in this piece is derived, not sourced - the arithmetic is shown in the text; agent-usage figures carried “as of mid-2026.”
Gallup, span of control (panel of 16,442 US managers, 2022-24) · Gallup, performance-review frequency · Gallup, onboarding ramp (12+ months) · DORA 2025, State of AI-assisted Software Development (n≈5,000) · Anthropic, How Claude Code is used in practice (~400k sessions) · METR, Time Horizon 1.1 (2026) · CMI/YouGov, Better Management 2023 (UK) · Zenger Folkman via HBR, 2012 · Ericsson 2008, deliberate practice · Hogarth, Lejarraga & Soyer 2015, kind vs. wicked learning environments · Schmidt & Bjork 1992 / Bjork 2018 · Barnett & Ceci 2002, taxonomy for far transfer · Kotter, What Leaders Really Do · Hochschild, The Managed Heart (1983) · Mollick, Management as AI superpower (2026) · HBR, Agent Managers (2026) · WEF (2025)
Leave a Reply