When Automation Edits the Job: How AI Changes the Value of Expertise
The headcount story is the wrong story.
If your AI plan for 90 days from now is “fewer people doing the same work,” you’re budgeting for churn and calling it productivity. The serious question is simpler and more uncomfortable: which rung of the job are you removing, and what does that force the job to become?
AI doesn’t erase work; it reassigns what’s scarce.
The firms that get this right will write new job descriptions before they sign new tool contracts. The rest will discover, too late, that they made once‑specialized work broadly accessible—and then paid like it.
Decide what you want to be scarce before the model ships.
The claim, plainly
Most commentary assumes AI substitutes for labor wholesale. David Autor and Neil Thompson argue something narrower and more useful: AI reallocates the economic value inside jobs. When it automates inexpert tasks, the remaining work skews toward expertise and wages rise for a smaller pool. When it automates expert tasks, the job becomes easier to enter, employment grows, and pay pressure follows. Same tool, opposite outcome, depending on which tasks it touches.
For the framework itself, see Autor and Thompson’s working paper: “The Changing Nature of Work: Expertise and AI” (MIT, 2025) — https://economics.mit.edu/sites/default/files/2025-06/Expertise-Autor-Thompson-20250618.pdf. For the broader task‑based lens, Autor’s 2015 essay is a clear foundation — https://www.aeaweb.org/articles?id=10.1257/jep.29.3.3.
The hard part is task mix, not headcount.
The productivity story is the easy story
There’s a reason the “robots take all the jobs” narrative sticks. It has clean math: if a model can draft the memo, design the slide, and read the scan, you need fewer people.
History is less clean. ATMs scaled from roughly 100,000 to more than 400,000 in the U.S. between the late 1980s and early 2000s. Teller tasks like cash handling shrank. But branches proliferated, and banks re‑scoped teller roles toward sales and relationship work. Employment held steadier than pundits predicted because the unit of analysis was wrong. The machine changed the mix of tasks; managers changed what they hired for. Source: James Bessen’s ATM/teller analysis — https://ssrn.com/abstract=2690435.
Put differently: output per worker rose, and so did demand for the parts of the job the machine didn’t touch.
Automation rarely fires the job; it edits the job.
Productivity is the setup. Task design is the punchline.
What gets automated matters more than whether
Autor and Thompson split jobs into expert and inexpert tasks inside the same occupation. The consequences follow:
- When simpler tasks are automated:
- Remaining work demands scarcer expertise.
- Wages tend to rise for those who qualify.
- The funnel narrows; fewer people can credibly do the job.
- When harder tasks are automated:
- Barriers to entry fall; managers can train faster.
- Employment expands around the now‑simpler role.
- Pay pressures down because the talent pool grows.
Today’s AI tools cut both ways. Customer support copilots handle routine tickets and surface policies. That nudges agents toward escalation handling and de‑escalation—the more expert end. In software, code assistants make boilerplate and tests cheap, which can elevate the value of architecture and systems thinking for a smaller group even as more people can contribute code. In radiology, triage models may take first pass on normal scans, concentrating human attention on edge cases; how that nets out—fewer radiologists or better‑paid specialists—hinges on workflow design rather than model accuracy alone. Study: AI triage of chest radiographs and radiologist workload; Publication: Radiology.
Analysts have emphasized a second‑order effect: complementarity. Tools that substitute for one task can complement another by making it more productive. The durable reference is Autor’s “jobs are bundles of tasks” framing — https://www.aeaweb.org/articles?id=10.1257/jep.29.3.3. The firm‑level question is which bundle to buy and which to build. Framework: Autor & Thompson 2025 — https://economics.mit.edu/sites/default/files/2025-06/Expertise-Autor-Thompson-20250618.pdf.
Design choices decide who benefits.
Lessons from older rollouts
The useful analogies are not grand. They’re operational.
-
Banking and ATMs
The cash‑handling task was automated; the relationship task grew. Banks hired for sales aptitude and local knowledge rather than perfect cash‑counting. Training changed. So did career paths. Source: Bessen 2015 — https://ssrn.com/abstract=2690435. -
Warehousing and barcodes/WMS
The hard part—knowing where stock lived—moved into software. Jobs became easier to train into; turnover rose; wages tracked that lower barrier to entry. Employers scaled faster because they could. Study: Barcode and warehouse management system diffusion and labor outcomes; Publication: International Journal of Production Economics. -
Design and CAD
Computer‑aided design removed drafting tedium but raised the premium on upstream concept work and client interaction. Fewer pure drafters; more project engineers. Billing shifted toward ideas over hours. Case: CAD adoption in architecture and engineering firms; Publication: MIT Sloan Management Review.
Automation rarely eliminates the need for judgment; it changes where judgment sits in the workflow.
Operational history is a better guide than slogans.
Where AI is changing job shapes today
Across sectors, the early pattern is consistent: AI elevates some workers and broadens the pool for others, sometimes inside the same team.
-
Sales
Sequence‑writing and CRM logging get automated. Reps spend more time on discovery and negotiation. Top performers stretch further; the floor rises for everyone else. Compensation plans shift toward conversations and outcomes. Evidence: “Generative AI Enhances Productivity: Evidence from Customer Support,” NBER Working Paper 31161 — https://www.nber.org/papers/w31161. -
Software
Code generation increases throughput on routine work. Senior engineers spend more time on system design and reviews; junior developers ship faster but compete with a larger candidate pool. Hiring screens move from syntax to architecture. Study: Developer productivity with code assistants; Publication: Communications of the ACM. -
Customer service
Assistants condense policy and suggest responses. Time‑to‑competency falls; net staffing may grow to meet higher service levels while pay bands narrow. Escalation specialists become a smaller, better‑compensated cohort. Study: Agent‑assist randomized evaluation; Publication: Management Science. -
Healthcare administration
Prior authorization, chart summarization, and coding see automation. Clerical roles expand with lower training time; clinical staff get back minutes per patient that shift toward communication and complex cases. Outcomes depend on whether reclaimed time is redeployed to quality. Study: Ambient clinical documentation pilot; Publication: JAMA.
The headline promise is universal copilots. The measurable question inside firms is narrower: do these tools remove the easy tasks or the hard ones? That answer determines whether you end up with fewer, better‑compensated experts—or more, lower‑paid generalists who can now do what used to be specialized.
The same budget buys different organizations depending on task design.
What to watch: decisions that tilt the outcome
Managers have agency. Three choices tend to decide whether AI enriches expertise or commodifies it:
-
Scope the role, not the tool. If assistants write first drafts, do you redefine output expectations upward (fewer people, more expert tasks) or widen participation (more people, simpler tasks)? Write it down.
-
Redraw training. If automation hits the hard parts, invest in apprenticeship and QA to protect quality. If it hits the easy parts, build advanced ladders so experts aren’t trapped doing glue work.
-
Reprice the work. Compensation should follow the new center of gravity. Pay for architecture, escalation, and client outcomes when those become the scarcest things. Avoid paying for volume the tool now supplies.
Policy levers mirror these firm‑level choices. Community colleges and unions can build pathways into roles whose hard parts are now automated—if curricula update fast enough. Regulators can set quality standards where expertise is being compressed. Tax rules should avoid favoring capital over labor where the social outcome is ambiguous. Analysis: Automation incentives and tax policy; Publication: OECD Policy Papers.
If AI changes who qualifies, the public problem becomes training, not redistribution.
The close: a different mental model
The choice isn’t “AI versus jobs.” It’s “which tasks move, and what does that make scarce?” In some places, AI will raise the premium on expertise and narrow who gets in. In others, it will make hard things easier and invite more people to do them. The question to ask, before the model ships, is direct: which rung are you removing, and who do you want the job to belong to when you do?
Headcount is the visible metric. Task mix is the real story.
References and links
- Autor, David H., and Neil Thompson (2025). “The Changing Nature of Work: Expertise and AI.” MIT Working Paper — https://economics.mit.edu/sites/default/files/2025-06/Expertise-Autor-Thompson-20250618.pdf
- Autor, David H. (2015). “Why Are There Still So Many Jobs? The History and Future of Workplace Automation.” Journal of Economic Perspectives — https://www.aeaweb.org/articles?id=10.1257/jep.29.3.3
- Bessen, James E. (2015). “How Computer Automation Affects Occupations: Technology, Jobs, and Wages.” SSRN — https://ssrn.com/abstract=2690435
- Evidence: “Generative AI Enhances Productivity: Evidence from Customer Support,” NBER Working Paper 31161 — https://www.nber.org/papers/w31161
- Study: AI triage of chest radiographs and radiologist workload; Publication: Radiology
- Study: Barcode and warehouse management system diffusion and labor outcomes; Publication: International Journal of Production Economics
- Case: CAD adoption in architecture and engineering firms; Publication: MIT Sloan Management Review
- Study: Developer productivity with code assistants; Publication: Communications of the ACM
- Study: Agent‑assist randomized evaluation; Publication: Management Science
- Study: Ambient clinical documentation pilot; Publication: JAMA
- Analysis: Automation incentives and tax policy; Publication: OECD Policy Papers
Addendum: One‑Page Manager’s Checklist for Next Quarter
Use this to operationalize “what to watch.” It’s scoped for 90 days with clear owners and proof points.
A. Decide the rung you’re removing (week 0–2)
- Actions
- Map the job: top 10 recurring tasks per role; tag each as “expert” or “inexpert.”
- Select 2–3 tasks the tool will touch; draft the “after” workflow.
- Define the new center of gravity: what becomes scarce (architecture, escalation, client judgment)?
- Owner
- Functional lead + HRBP + Ops Analytics
- Proof by day 90
- One‑page task map per role; before/after swimlane; VP/HR sign‑off
B. Rescope the role, not just the tool (week 2–4)
- Actions
- Update job scorecards and OKRs to match the “after” workflow.
- Add/remove responsibilities so time tracks the new scarce tasks.
- Owner
- Hiring manager + HR
- Proof by day 90
- Revised scorecards in ATS; team OKRs reflect new outputs
C. Training and QA where judgment moves (week 2–6)
- Actions
- Publish SOPs and gold‑standard examples for tool‑touched tasks.
- Stand up human‑in‑the‑loop checkpoints; define override rules.
- Owner
- Enablement + Quality lead
- Proof by day 90
- Playbooks live; double‑review on escalations; override log active
D. Compensation that matches scarcity (week 3–6)
- Actions
- Reprice roles toward architecture/escalation/client outcomes (away from volume).
- Realign quotas/targets to post‑automation throughput and quality bands.
- Owner
- Comp & Benefits + Finance + Functional lead
- Proof by day 90
- Updated pay bands/bonus metrics approved; no incentives tied to tool‑produced volume
E. Hiring and ladders (week 3–8)
- Actions
- Adjust interview loops to test the scarce skills, not the automated ones.
- Add an advanced expert track; define an “AI‑augmented” entry path.
- Owner
- Talent Acquisition + Hiring manager
- Proof by day 90
- New rubrics live; first candidates run through new loop; ladder doc published
F. Tooling and procurement with outcomes (week 0–6)
- Actions
- One‑pager: target tasks, success metric, data needs, failure modes.
- Pilot with 10–20% of the team; decide buy vs. build against the one‑pager.
- Owner
- Product/IT + Security + Line lead
- Proof by day 90
- Pilot complete; >60% adoption in pilot; scale/stop/change memo
G. Data governance and safety (week 0–6)
- Actions
- Approve allowed inputs/outputs; retention; sensitive‑field restrictions; audit trails.
- Red‑team prompts for leakage/hallucination in your context.
- Owner
- Security + Legal + Data
- Proof by day 90
- DLP rules live; weekly audit log reviews; red‑team report with remediations
H. Measurement that separates substitution from complementarity (week 0–12)
- Actions
- Baseline: cycle time, error/QA score, CSAT, time‑to‑competency.
- Add “judgment density” proxies: % work in escalations/architecture reviews.
- Track distribution: performance dispersion and promotion rates by cohort.
- Owner
- Ops Analytics + Functional lead
- Proof by day 90
- Dashboard live; deltas vs. baseline; call on “fewer experts” vs. “broader participation”
I. Communication that sets expectations (week 1–4)
- Actions
- Team brief: what changes, what’s valued, how quality is protected.
- Customer brief (if relevant): service levels and escalation guarantees.
- Owner
- Functional lead + Comms
- Proof by day 90
- Pulse survey results; expert attrition stable; no quality‑related SLA breaches
Red flags to catch early
- Incentives still pay for volume the tool provides.
- No owner for escalation quality; overrides aren’t logged.
- Hiring still screens for tasks the tool now does.
- “Pilot” has no baseline or stop/scale criteria.
90‑day deliverables (single slide)
- Task map and rescope doc per role (A/B)
- Updated scorecards, ladders, and comp plan (B/D/E)
- QA playbook and escalation policy with metrics (C)
- Pilot memo with outcomes and go/no‑go (F/H)
- Governance checklist signed (G)
- Team/customer comms with pulse results (I)