Who Gets Augmented, Who Gets Replaced: The 2026 Job-Displacement Map

Section 01Stop Asking 'Will AI Take My Job.' Start Asking Two Better Questions.

In every public conversation about AI and work, the same sentence appears: "AI will take X% of jobs by year Y." It is the wrong sentence. It hides more than it reveals — and it has produced a labor policy debate stuck on the question of whether displacement is coming, when the more useful question is which workers are being augmented, which are being replaced, and which of them have the bargaining power to do anything about it.

This piece offers a 2×2. It is not the only way to think about AI's labor effects. It is the way we think has been most useful in our conversations with workers, employers, and policymakers over the last two years.

The 2×2 has two axes:

• Axis 1: Augmented vs. Replaced. Does AI make this worker more productive in their existing role (augmentation), or does it eliminate the role and require the worker to move to a different one (replacement)? This is the technological question.
• Axis 2: High vs. Low Bargaining Power. Can this worker capture some of the surplus that AI produces — in wages, hours, benefits, or autonomy — or does the surplus accrue entirely to the firm or to consumers? This is the political-economy question.

The point of putting both axes on the same chart is that the worker's experience of AI is determined as much by the second axis as the first. A radiologist whose throughput doubles because of an AI image-reading assistant and whose specialty has restricted labor supply will see most of that productivity flow back to them as compensation. A customer-service agent whose throughput doubles because of a generative-AI assistant and whose role has no licensing barrier, no union, and a global remote labor pool will see most of that productivity flow back to the firm. Same technology. Different outcome.

Section 02The Four Quadrants

The four quadrants the 2×2 produces are not theoretical. They map onto observable groups of workers right now.

One observation about this table that is easy to miss: the bottom row (Replaced) is not necessarily worse for the worker than the top row (Augmented). A worker in Quadrant C, with strong bargaining power, may end the transition with a better job than a worker in Quadrant B who is augmented but losing real wages relative to a rising performance bar. The technology axis is real; the bargaining axis is just as real; and it is the combination that determines outcomes.

Section 03What the 2025–2026 Data Actually Says (Quadrant by Quadrant)

Reasonable people can disagree about how to populate the quadrants. The data points below are the ones we have found most reliable. We have cited each one.

Quadrant A: The Compounders

The clearest 2024–2025 evidence of AI augmentation producing worker capture comes from software engineering and from radiology. A controlled field experiment of GitHub Copilot at Microsoft, Accenture, and a Fortune 100 employer found a 26% increase in completed tasks per developer among less-experienced engineers, with no degradation in code review pass rates [1]. Critically, in the U.S. software labor market, that productivity gain did not translate into proportional headcount reductions in 2025 — total software employment was roughly flat, while compensation for engineers with demonstrated AI-tool proficiency rose 8–12% in the major metros. The augmentation surplus partially accrued to workers.

Radiology is the often-cited Quadrant A example. Despite Geoffrey Hinton's 2016 prediction that "we should stop training radiologists now," the AAMC's 2025 physician workforce data shows demand for radiologists at an all-time high. AI image-reading assistants have shifted radiologists' work toward complex case adjudication and patient-facing consultation; volume per radiologist has risen, and licensed supply has not. The result has been higher compensation, not lower.

Quadrant B: The Treadmillers

This is the quadrant that gets the least public sympathy and arguably has the most workers in it. The defining 2025 pattern is "same number of jobs, rising performance bar."

The clearest case is entry-level analyst and junior knowledge work. A study published in NBER in late 2024 estimated that generative AI raised the output of less-experienced consultants by 43% in tasks within the technological frontier — but for tasks outside the frontier, AI use degraded performance [2]. Firms in 2025 responded not by reducing entry-level hiring but by raising what they expect at the same compensation. The hiring threshold at McKinsey, BCG, and the Big Four for first-year analysts in 2025 included demonstrated proficiency in at least two AI-augmentation workflows. The work is not gone; the work is now what used to be the second-year work, done by first-year hires.

The same pattern has played out in marketing copywriting, basic graphic design, and contract paralegal work — the surplus is real, but is being absorbed by firms or by consumers (faster turnaround, more deliverables per dollar), not by the workers doing the work.

Quadrant C: The Negotiators

2023 was the year Quadrant C made itself visible. The Writers Guild of America's strike and the SAG-AFTRA strike both centered on AI-related provisions: minimum guarantees against AI-written scripts being used as a basis for "rewrites" at lower rates, consent requirements for digital replicas of performers, and revenue-share frameworks for AI-trained data. The settlements that emerged are imperfect, but they established the principle that AI's effect on a role is a bargainable item.

In 2024–2025, dockworkers in the U.S. and Europe achieved comparable wins on terminal automation; nursing unions in California negotiated guardrails on AI-driven staffing software; and the EU's AI Act required works-council consultation for high-risk workplace AI deployments. The pattern is consistent: where bargaining power exists, the transition is being shaped, not just absorbed.

Quadrant D: The Casualties

This is the quadrant the public conversation has the hardest time naming. The clearest 2025 evidence is in call-center work, particularly in low-protection labor markets. NWLB's 2025 survey of the Indian and Philippine business-process-outsourcing workforces (n=3,400 workers across 14 employers) found that 31% of respondents in agent roles reported their team headcount had been reduced in the prior 12 months, attributed to "voice AI tools" or "automated resolution"; 22% of respondents in their twenties reported leaving the industry entirely in that window. Wages for the remaining roles have not risen.

The casualties of Quadrant D are not abstract. They are workers who entered a sector during a hiring boom in 2018–2022 and are exiting it in 2024–2026 with comparatively few transferable credentials, comparatively little severance, and a labor market that does not yet have a clear next destination for them.

Section 04Three Frames This Replaces

The 2×2 above is most useful if we are explicit about what it replaces.

It replaces "AI will take X% of jobs." The number — whether it is the Oxford 47% from 2013, the OECD 14% from 2018, the Goldman Sachs 300 million from 2023, or the McKinsey 30% of hours from 2026 — is computed by counting tasks within current job descriptions that current AI could perform. It is a measure of technological exposure, not of labor outcomes. A high-exposure role in Quadrant A produces compounders; the same role in Quadrant D produces casualties. The exposure number alone tells you neither.

It replaces "AI augments rather than replaces." This sentence — most associated with the productivity-optimist case from researchers like Brynjolfsson and McAfee — is true on average and false in distribution. AI augments average workers and replaces marginal workers within the same occupation, and the difference between the two often correlates with race, geography, or seniority [3]. "Augment vs. replace" is the right framing; it is just under-applied when used as a blanket statement about the technology.

It replaces "we need to reskill workers." Reskilling is necessary, but a 2025 reskilling policy without an accompanying bargaining-power story is asking workers to absorb the entire cost of the transition while the firm captures the entire gain. The quadrants make this concrete: Quadrant D workers cannot reskill their way out of Quadrant D unless either (a) their bargaining power changes, or (b) policy moves the transition costs off of them. Reskilling alone moves them from Quadrant D to a different but still-Quadrant-D occupation a few years later.

The history of automation in the United States is not the history of jobs lost. It is the history of bargaining position eroded. David Autor, MIT, in testimony before the U.S. Senate HELP Committee, 2024

Section 05What Each Actor Should Actually Do

For workers

Identify which quadrant you are in. The diagnostic is not your job title — it is the combination of (a) whether AI is multiplying your output and (b) whether you can capture the multiple. If you are in Quadrant B, the move is not "use AI more"; it is to make the captured surplus visible (in performance reviews, in market-rate compensation comparisons, in negotiated terms). If you are in Quadrant D, the move is a deliberate pivot to a Quadrant A or C adjacent role now, while you still have current employment to fund the transition.

For employers

Most enterprise AI deployments in 2025 produced internal narratives of augmentation and external observed behavior of replacement. The credibility cost of that gap is rising. If your company is realizing AI productivity gains and is not, at the same time, sharing some of them with the workers whose work was augmented — in pay, in hours, in role design, in retraining — your retention numbers in 2026 and 2027 will tell you what your workers concluded.

For policymakers

The single most important labor-market intervention you can make in 2026–2028 is not retraining money. It is moving bargaining power. Concretely, this means:

• Portable benefits — health, retirement, paid leave — that travel with the worker rather than the role, so transitions become less catastrophic
• Sectoral standards for AI-augmented roles (the EU's AI Act consultation requirements are a starting model, not an endpoint)
• Public reskilling infrastructure priced as a public good, not as consumer-paid online courses
• Antitrust enforcement against labor monopsony, which is the single biggest depressant of worker bargaining power in the modern U.S. labor market

For unions and worker organizations

The 2023 WGA settlement is the playbook for the 2026 organizing campaign. AI-related provisions — minimums, consent rights, transparency requirements, revenue share — are now bargainable items, and the workers who get them earliest will set the terms for the rest of the labor market.

Section 06The 2026 Job-Displacement Map — Now Live

The 2×2 above is a way to think. The populated map is the artifact that makes it usable.

What the first edition includes:

• Quadrant A (the Compounders, 24 occupations) — radiologists, senior software engineers, anesthesiologists, tax partners, AI/ML research engineers, venture-capital partners, licensed PEs, veterinarians, and others where AI multiplies output and bargaining power converts the multiple into compensation.
• Quadrant B (the Treadmillers, 11 occupations) — junior consultants, first-year associates, paralegals, copywriters, junior designers, SDRs, junior accountants — where AI multiplies output but the firm captures the multiple and the hiring threshold rises.
• Quadrant C (the Negotiators, 15 occupations) — screen actors, screenwriters, dockworkers, UAW assembly workers, unionized nurses, airline pilots — where the role is being restructured but collective bargaining is making the transition managed.
• Quadrant D (the Casualties, 22 occupations) — offshore call-center agents, data-entry clerks, basic content moderators, telemarketers, retail cashiers — where the role is going and workers lack leverage to shape the terms.

The methodology block on the map page explains the indices in detail. Augmentation index synthesizes Eloundou et al. (2023) GPT-exposure scores, McKinsey 2026 task-level estimates, and BLS occupational projections. Bargaining-power index is a composite of unionization rate, licensing barriers, labor-market concentration (HHI by occupation × MSA from BLS OEWS), and the NWLB HAPI Network Capital dimension averaged within the occupation.

The map is meant to be argued with, cited, and used. If you disagree with a placement, the open data file (/ai-job-map/data.csv) is the place to start — every score is published, every placement is defended in the per-occupation evidence panel, and the 2027 edition will incorporate substantive challenges. Write to [email protected].

"AI will take jobs" is the wrong sentence. The right one is: which workers can capture the surplus AI is creating, and which ones can we still help to capture it.

Section 07A Note on Confidence

The history of public predictions about AI and labor is the history of wrong predictions. The 2013 Oxford study estimated 47% of U.S. jobs at "high risk" of computerization by 2033; we are now eight years from that deadline and total U.S. employment is at an all-time high. The 2016 confident prediction that radiologists would be obsolete by 2021 has aged in the way that confident predictions about labor tend to age.

This is not an argument that the current generation of predictions is also wrong. Generative AI is a different technology, the deployment surface is broader, and the labor market is differently structured than it was in 2013. It is an argument for humility about the magnitude and the timing — and for routing the public debate away from the question of "how many jobs" and toward the question of "which workers, in which positions, with what bargaining power."

That is the question this piece is trying to make easier to ask.