After ChatGPT, Which Managerial Skills Did Firms Stop Asking for First? ChatGPT 이후 기업은 어떤 관리 능력부터 덜 찾게 되었나

Skill Deprioritization: Reorganizing in the Age of Generative Artificial Intelligence | Management ScienceHow does generative artificial intelligence (GenAI) reshape the skills that organizations seek as they adapt to a new general-purpose technology? GenAI effectively retrieves data, performs analysis…🔗 doi.org/10.1287/mnsc.2025.01859Skill Deprioritization

Devigili, M., Yilmaz, E. D., Gaba, V., & Greve, H. (2026). Skill Deprioritization: Reorganizing in the Age of Generative Artificial Intelligence. Management Science, 0(0).

✔️ Introduction and Summary

This issue covers a paper that empirically examines which managerial skills firms stopped looking for once generative AI arrived.

“AI will take our jobs” is a familiar line by now, but it doesn’t answer the question of exactly which skills a company cuts first when it hires less, or why those particular ones. The authors call this phenomenon skill deprioritization — a systematic reduction in firms’ demand for human skills that GenAI can effectively substitute for. Unlike existing research, which has mostly focused on augmentation (does AI make a person better at their job) or reskilling (what new skills does a person need), this paper asks a different question: what does the organization no longer need?

What matters about this definition is that what the paper measures is job postings — not headcount, and not layoffs. Take an example: recruiter postings used to routinely list “familiar with maps” as a requirement for sales or delivery roles, but once navigation apps became standard, that line quietly disappeared from postings. It’s not that companies let go of the people who once knew this skill — it’s that they stopped requiring it when hiring anew. That’s the kind of change this paper is about.

The authors combine organization theory’s classification of managerial work with queuing theory from industrial engineering to predict which skills will shrink first, then treat ChatGPT’s release as a natural experiment and use difference-in-differences to test that prediction against 11.15 million U.S. publicly listed firms’ job postings.

The headline finding: after ChatGPT, firms in industries with high GenAI exposure cut job postings requiring skills that break down and sequence work by 23%, relative to firms in low-exposure industries. Postings requiring skills that convey information (roughly 18–21%), observe and evaluate performance (roughly 23%), and handle operational problems (roughly 19%) also declined. Postings requiring the skill to hire and place people, to design reward systems, or to resolve conflict, on the other hand, showed no statistically significant change.

✔️ Research Content and Logic

Where existing research stood

Existing research on AI and work has largely operated at the level of the task — can AI substitute for a given task, how quickly and accurately can a person do it with AI’s help. The authors argue a different lens is needed. A general-purpose technology only pays off once the organization itself is redesigned around it. According to David (1990), simply installing electric motors in a factory didn’t raise productivity — the gains only showed up once the factory floor was redesigned. Applying that logic directly, looking only at how much faster GenAI makes an individual task tells you nothing about change at the organizational level. The lens the authors chose for observing that change is the firm’s hiring demand.

How to break down what a manager does

This paper classifies work not by its content, but by Puranam’s (2018) problems of organizing framework — what needs solving when multiple people work together. It splits broadly into division of labor and integration of effort, which further break into five branches and ten sub-items.

Category	Sub-item	What it means
Task division	—	Breaking a goal into executable pieces and sequencing and prioritizing them
Task allocation	Staffing / Mapping	Deciding who gets which piece
Information provision	Coordination / Synchronous communication / Asynchronous communication	Making the information people need flow
Reward distribution	Monitoring / Incentives	Motivating people to act in line with goals
Exception management	Conflict resolution / Operational exception management	Handling things that didn’t go as expected

The authors give this level of granularity for two reasons: it enables more concrete predictions — something like “GenAI will affect operational exceptions but not conflict resolution” becomes possible — and, on the measurement side, it lets each category be defined without overlapping the others.

Predicting with queuing theory

Now the moment to predict which of these ten items shrinks first. The tool the authors borrow for this is queuing theory — an industrial-engineering tool for problems of waiting in line, where knowing how fast customers arrive, how long one staff member takes to process each one, and how irregular that processing time is lets you calculate the expected wait time. Glynn et al. (2020) applied this to organizations, treating a firm’s problems (decisions requiring judgment) as customers and the people who process them as the staff behind the counter.

The simplest prediction is that one person processing problems faster means the same work gets done with fewer people. But the authors say this effect applies roughly equally across all ten categories, so it can’t tell us which category is relatively more vulnerable. So they offer three more concrete predictions.

By contrast, the case is weak for task allocation (staffing and mapping), incentive design, and conflict resolution — even once a person is using AI, the nature of selecting the right person and motivating them doesn’t itself change. A second-order effect could still shrink these, but only if the organization’s structure itself gets redesigned, and that takes time to experiment with, learn from, and implement — too long to expect it to show up in this paper’s observation window right after ChatGPT’s launch.

Why this counts as “reorganizing” — the span-of-control story

The “Reorganizing” in the paper’s title points to why a decline in demand for a few skills connects to the firm redesigning itself. The concept that bridges this gap is span of control — the number of subordinates a manager directly supervises. Span of control is set by (i) how often a decision requiring the manager’s approval comes up, and (ii) how variable the time to process one decision is — and it’s (ii) that this paper focuses on. Queuing theory says expected wait time falls not only when processing gets shorter, but simply when processing time becomes more predictable. As GenAI takes over a share of coordination work, the variability in a manager’s remaining workload falls, and that slack lets one manager take on more subordinates.

✔️ Data and Methodology

What was collected, and from where

The sample is 1,820 U.S. publicly listed firms, 40,465 firm-month observations. Three data sources were combined. Lightcast, a database that collects U.S. job postings daily and tags each posting with the skills it requires, supplies all of this paper’s dependent variables. Compustat provides quarterly financial data, and Revelio provides monthly headcount data by firm; both were used as control variables.

The observation window is ±12 months around ChatGPT’s public release on November 30, 2022 (pre-period: December 2021–November 2022; post-period: December 2022–November 2023). The pre-period needs to be long enough to establish a comparison baseline but recent enough to be unaffected by earlier AI developments; the post-period needs to give firms enough time to react without being so long that other events get mixed in.

How “high-exposure industries” was defined

This paper’s treated group is not “firms that actually adopted AI” — a point that’s easy to get confused about. What the authors use is AI Occupational Exposure (AIOE), a measure from Felten et al. (2021/2023). The U.S. Department of Labor’s O*NET database maps how more than 800 occupations connect to 52 abilities; multiplying this by a survey-based matrix of “how exposed is each of these 52 abilities to a given AI capability” produces an occupation-level exposure score. This paper picks out exposure to the “language modeling” AI capability specifically.

Here, k is an occupation, j is one of the 52 abilities, A is how exposed ability j is to language-modeling AI, and L and I capture how important ability j is to occupation k and how frequently it’s used. To roll this occupation-level value up to the industry level, the authors used Lightcast to tally monthly job postings by four-digit NAICS industry and occupation, found each occupation’s share within its industry and weighted accordingly, then averaged the monthly values over the pre-ChatGPT period to arrive at each industry’s exposure. Industries whose exposure exceeded the sample average were classified as treated, yielding 966 treated firms and 854 control firms — a near-even split. To put it precisely, what this paper’s criterion sorts on is not “did this firm adopt AI” but “how exposed is this firm’s occupational mix to language models.”

Sorting 33,832 skills in job postings into ten categories

Lightcast’s skill taxonomy (version 9.25) contains 33,832 individual skills, each with a roughly 70-word description. The authors went through a multi-step process to sort these into the ten categories above: a topic model (LDA) first narrowed the field to 1,765 candidate skills tied to the organizing categories; the authors read these directly and assigned labels, then checked agreement against a research assistant’s independent classification (Cohen’s kappa of 0.79). This classification standard was then taught to three GPT-4o agents and scaled up to the full 33,832 skills (93% accuracy, 89% F1), and a final human check confirmed 657 skills and 719 skill-label connections. Applying this classification produced 11,154,915 processed job postings; since one posting can span multiple categories, the counts across categories are not mutually exclusive.

The math of the queuing model

Let’s work out the span-of-control mechanism mentioned in “Research Content and Logic” above as an equation. Suppose problems arrive at a manager at an average frequency of λ, and processing time follows a distribution with mean 1/μ and variance σ². If GenAI takes over a share φ of coordination work, the remaining variance is:

For a queue of this form (M/G/1), the expected wait time is:

What’s worth noting is that σ² sits in the numerator: it isn’t only that processing time gets shorter — a fall in the variability of processing time by itself shortens the wait. As φ grows, wait time falls, and that slack lets the manager take on more subordinates. The additional headcount the authors work out is:

Here Δn is the number of additional subordinates that can be added without extending wait time, and λ̄ is the average arrival rate generated by one subordinate. The larger the original variability in processing time, and the closer the manager was already to saturation, the bigger this effect.

How the estimation was done

The authors used difference-in-differences (DiD): subtracting the change experienced by firms in low-exposure industries from the change experienced by firms in high-exposure industries, before versus after ChatGPT. This cancels out any factor that acted equally on all firms over the period, such as the business cycle or the overall hiring market.

Here, y is the count of firm i’s postings in month t requiring the given skill; Treated is 1 if the firm is in a high-exposure industry; PostGPT is 1 from December 2022 onward.

Estimation used Poisson Pseudo-Maximum Likelihood (PPML), because it can handle count variables with many zero values without a problematic transformation, it gives consistent estimates even when the data doesn’t strictly follow a Poisson distribution, and it holds up even when variance isn’t constant. Standard errors were clustered at the firm level.

Control variables are size (log total assets), ROA, leverage, R&D over total assets, an R&D dummy, capital expenditure over total assets, and workforce change rate — included because a firm going through large-scale layoffs might post fewer jobs overall for reasons unrelated to GenAI.

✔️ Findings

What skills were firms originally looking for?

First, the baseline: average postings per firm-month.

Item	Avg. postings
Task division	89.3
Information provision (synchronous)	75.5
Exception management (operational)	63.6
Information provision (asynchronous)	60.1
Information provision (coordination)	52.9
Reward distribution (monitoring)	24.6
Exception management (conflict)	13.7
Task allocation (mapping)	13.0
Task allocation (staffing)	5.1
Reward distribution (incentives)	1.4

Postings related to task division dominate overwhelmingly, while postings related to incentive design are barely posted at all. Fifty-three percent of sample firms were classified as belonging to high-exposure industries; the average firm has a log total assets of 7.87, a leverage ratio of 0.30, and 42% report R&D expenditure.

1 Baseline results — what shrank

The coefficients below are on a log scale and shouldn’t be read as-is; the actual rate of change is exp(β) − 1.

Item	Coefficient	% change	Significance
Task division	−0.261	−23.0%	p<0.001
Reward distribution (monitoring)	−0.264	−23.2%	p<0.001
Information provision (asynchronous)	−0.241	−21.4%	p<0.01
Information provision (coordination)	−0.240	−21.3%	p<0.01
Exception management (operational)	−0.211	−19.0%	p<0.01
Information provision (synchronous)	−0.204	−18.5%	p<0.05
Exception management (conflict)	−0.224	—	Not significant
Reward distribution (incentives)	−0.150	—	Not significant
Task allocation (mapping)	−0.097	—	Not significant
Task allocation (staffing)	+0.162	—	Not significant

All three of the theory’s predictions were confirmed in the data. Task division, and all three sub-items of information provision (coordination, synchronous, asynchronous), declined, and operational exception management — which was predicted to follow along with task division — declined as well. Monitoring within reward distribution declined too. By contrast, there was no significant change in task allocation (both staffing and mapping), incentives, or conflict resolution.

That said, conflict resolution needs a careful read. The coefficient itself (−0.224) is a similar magnitude to the other categories that did come out significant, but its standard error at 0.131 was large enough to miss the significance threshold — closer to “the data can’t tell us whether there was a change” than “there was no change.” Adding control variables leaves the results largely intact.

2 Did parallel trends hold?

Difference-in-differences rests on the assumption that “the two groups would have moved in parallel had ChatGPT not happened.” The authors broke out the pre-treatment period by month and tested whether the pre-treatment coefficients were all jointly zero, and this assumption was not rejected for most items. Task allocation (staffing) is the exception: the two groups were already moving differently before treatment, so the earlier “staffing showed no change” result rests on weaker footing than the other items.

3 What if exposure is treated as continuous, and GPT-3.5 is split from GPT-4?

Estimating with the industry’s continuous exposure value instead of the treated/control binary kept the results pointing in the same direction overall. Only information provision (synchronous, real-time communication) lost significance — a signal that this sub-item wobbles repeatedly across the robustness checks that follow as well.

Splitting the treatment period into the GPT-3.5 period (December 2022–February 2023) and the GPT-4 period (from March 2023 onward) produces a more interesting pattern.

Item	GPT-3.5 period	GPT-4 period
Reward distribution (monitoring)	−0.164 (not sig.)	−0.291 (p<0.001)
Exception management (operational)	−0.150 (not sig.)	−0.230 (p<0.01)
Information provision (asynchronous)	−0.220 (p<0.05)	−0.250 (p<0.01)
Task division	−0.262 (p<0.01)	−0.260 (p<0.001)
Information provision (coordination)	−0.272 (p<0.05)	−0.225 (p<0.01)

Monitoring and operational exception management are not significant during GPT-3.5 but become markedly stronger during GPT-4. This is read as a sign that actual improvement in AI capability, not novelty or hype, moved firms’ hiring behavior.

4 Placebo tests and falsification tests

A placebo test repeating the same analysis at fake event dates (6, 9, and 12 months before the actual release) produced no significant results at 9 or 12 months prior, and only one small, weakly significant coefficient at 6 months prior. Analyzing GitHub Copilot’s release the same way — a tool with a narrow, developer-focused user base — showed no comparable effect, as expected. Switching to a broader AI exposure measure not limited to language models produced similar results for task division, monitoring, and operational exception, but revealed a new significant effect for task allocation (mapping), which is left as a topic for future research. All three of these checks support that the original results aren’t due to chance.

5 Re-running with different estimation methods, classifications, and samples

The authors re-confirmed the results with a sensitivity analysis on violations of the parallel-trends assumption, a different estimator (ETWFE), keyword-based reclassification, excluding tech firms, redefining the treated group by a nonroutine-cognitive-occupation threshold, and matching similar firms via Coarsened Exact Matching (CEM), among other approaches.

Four items held significance across every check: task division, information provision (asynchronous), monitoring, and operational exception management — the paper’s most robust findings. By contrast, information provision (synchronous) and part of monitoring wobbled in significance depending on the check, and the conclusion for conflict resolution shifted depending on how the treated group was defined.

One result worth noting: collapsing the ten sub-items back into the five broad categories erases the internal differences within reward distribution and exception management (such as monitoring shrinking while incentives didn’t). This is exactly why the paper broke things down into ten categories in the first place.

6 Supplementary analysis — was this just an overall cut to hiring?

Controlling for total hiring demand and re-estimating leaves the declines in coordination and task division intact, showing these skills were pushed out relatively more independent of any overall cut to hiring. Comparing managerial postings against nonmanagerial ones, the higher a firm’s exposure, the more its share of managerial postings shrank — consistent with the span-of-control logic described above. At the same time, postings requiring GenAI-related skills rose sharply, led by high-exposure firms, which is itself evidence that the treated and control groups were correctly separated.

✔️ Conclusion

Over the year following ChatGPT’s release, U.S. publicly listed firms in high-exposure industries cut postings requiring skills that break down and organize work (−23%), convey information (−18–21%), observe performance (−23%), and handle operational problems (−19%), relative to firms in low-exposure industries. Postings requiring the skill to hire and place people, to design reward systems, or to resolve conflict did not change in a statistically meaningful way.

The significance of this result is that the prediction was built before the data was even examined. The authors modeled the organization as a queue, predicted that “functions where AI reduces the variability of processing time will see demand fall first,” and only then checked the data — and the actual result matched that prediction’s direction. The authors don’t claim firms are behaving fully optimally, but this paper shows the queuing model is a useful first approximation for explaining the initial reaction.

The limits the authors themselves note: first, treatment was defined at the industry level, not the individual firm, so differences in how fast firms within the same industry actually adopt AI go uncaptured — and the more that actual AI users are mixed into the control group, the smaller the gap between treated and control, so if anything this likely understates rather than overstates the true effect. Second, the observation window covers only 12 months after ChatGPT’s release, so it wasn’t possible to observe firms actually redesigning their organizational structure. What’s confirmed here is the first-order reaction visible in job postings; the second-order effect of organizational structure itself changing is expected to unfold over the coming years.

Even so, this paper’s contribution is clear. Until now, “AI will take our jobs” was mostly discussed at the level of occupation titles. This paper shifts that lens down one level, to the kinds of managerial work done within a job — meaning a question less like “which job should I choose” and more like “which kind of judgment within that job should I take responsibility for” may become the more pertinent one for anyone planning a career. This paper lays the groundwork for the next step: re-measuring that second-order effect using the same method, at the level of the firm.

✔️ References

• Queuing theory — the industrial-engineering tool for problems of waiting in line. The backbone of this paper’s theory.
• Difference-in-differences — this paper’s primary estimation method.
• O*NET — the U.S. Department of Labor’s occupation-ability database. The starting point for measuring exposure.
• Felten, E. W., Raj, M., & Seamans, R. (2021). Occupational, industry, and geographic exposure to artificial intelligence: A novel dataset and its potential uses. Strategic Management Journal, 42(12), 2195–2217. — the original paper behind AI Occupational Exposure (AIOE).
• Latent Dirichlet Allocation (LDA) — the topic model used to narrow down skill candidates.
• Cohen’s kappa — the metric for how much two people’s classifications agree.
• Lightcast — the job-postings database.
• Atlassian, “AI-powered project management” — the product page introducing Jira’s built-in generative AI feature, describing how it automatically breaks a plain-language goal down into subtasks and dependencies. (Source for the task division example.)
• TechCircle, “11 companies using GPT-4 in consumer products” (March 16, 2023) — an article covering 11 companies that applied GPT-4 to their products, including Stripe’s use of it internally to scan documents and code and flag error patterns. (Source for the operational exception management example.)
• FedScoop, “State Department encouraging workers to use ChatGPT” — an article reporting that the U.S. State Department deployed an internal GenAI chatbot for 1,000 staff in 2024, used for summarizing policy documents and translation. (Source for the information provision example.)
• SHRM, “How HR is using generative AI for performance management” — an article covering the HR analytics tool Culture Amp’s experiment using ChatGPT to compile Slack kudos, customer comments, and peer feedback into a performance summary. (Source for the monitoring example.)