Everyone Is a Manager: AI Use as an Act of Management (CESifo WP 13030)
B · Useful, limited A formal model from management economists Benjamin Leyden (Cornell) and Daniela Scur (Monash), clear and well argued, but theory only: it collects no new data and its predictions are untested.
Summary based on the full text·Original: about 75 minutes to read
In short
CESifo Working Paper 13030 (September 2026) by Benjamin T. Leyden and Daniela Scur. A worker with an AI assistant stops executing and starts delegating, so using AI is a management task. The model splits that management into direction (specifying the work) and verification (catching errors). As tools get more reliable, direction pays more and verification pays less, so people keep building the first skill and lose the second until checking stops altogether.
Key numbers
- 2.1% to 25.6%share of Codex users delegating tasks that would take an experienced human over a workday, first half of 2026 (cited from OpenAI data)
- more than a quarterOpenAI staff who managed five or more concurrent agents in a single week in June 2026 (cited)
- 3options per task in the model: do it yourself, trust the AI output, or vet it
As reported by the source.
English·Web page·opens d.repec.org
What it reports
- With skills fixed, the return to AI falls with task skill and rises with each managerial ability. This reconciles studies where AI compressed performance gaps (it substitutes for task skill) with those where gaps widened (outcomes depend on what people do with similar output).
- As reliability rises, the leading skill shifts from verification to direction: surviving errors are rarer and subtler, so each catch is harder and worth less. Where rare errors are catastrophic, verification keeps paying.
- With learning by doing, direction keeps improving with use while verification decays once checking no longer pays, ending in a "deskilling trap" where the skill collapses.
- An "accountability wedge" (workers bear only part of the cost of errors they ship) speeds the collapse; it does not affect direction, which the worker is paid for in full. AI use builds "lopsided" managers.
- Firms may rationally let verification collapse; the case for preventing it weakens as tools improve.
- In a competitive labour market, strong verifiers sort into high-stakes jobs, which delays but does not stop the collapse.
How they got it
A theoretical model of a non-managerial knowledge worker choosing per task between Solo, Trust and Vet, with task skill, direction and verification as separate abilities and the AI defined by reliability and scope. It extends Garicano's knowledge-hierarchy framework and adds learning by doing, an accountability wedge and labour-market sorting. Empirical motivation comes from cited studies and OpenAI usage figures.
Limits
- No new data; all propositions are model results awaiting tests.
- Interpersonal management is set aside by design.
- The usage statistics in the introduction are cited from another source, not produced here.
- The mapping from real tools to "reliability" and "scope" is abstract.
Why it matters for companies
Rolling out AI assistants turns every individual contributor into a part-time manager, though none was hired or trained for it. Training should cover how to specify and decompose work, and separately how to check output, since the second skill is the one that erodes. Where errors are rare but costly (finance, safety, legal, production code), firms need deliberate review practice, accountability that reaches the person shipping, and staff whose job is checking. Where errors are cheap, letting checking lapse may be a reasonable trade.
Summary based on the full text. Read the original for the full argument and data.Read the original on d.repec.org