AI ResearchJobsAIEconomicsFuture of Work

The pie gets bigger: a factor map of AI job disruption (US & UK)

Jon Twigge · 10 June 2026

Why this report exists. I published a blog post — The pie gets bigger. The pieces don't redistribute on their own. — arguing that what decides AI's effect on any given job is a small set of economic forces, not the technology's raw capability. This is the research that sits underneath that post: the verified findings it draws on, the directional sector bands, and — just as useful — the plausible-sounding claims that failed verification and should not be cited. The thread started with the Workonomics article "The Pie Gets Bigger. The Pieces Don't Redistribute Automatically.", whose central point the verification supports.

Methodology

Compiled 2026-06-07 via an adversarially-verified deep-research pipeline: 29 sources fetched → 117 candidate claims → the top 25 put to a 3-vote refute panel → 20 confirmed, 5 killed. Confidence and vote counts are shown per finding.

TL;DR

  • There is no single, knowable AI job-loss number. Credible forecasts disagree by an order of magnitude, and the famous headline figures are task exposure, not predicted headcount loss. Goldman Sachs: ~25% of US work hours are automatable; base case 6–7% of workers displaced over a ~10-year transition, with only +0.6pp unemployment if spread over a decade. The Tony Blair Institute: 1–3 million UK jobs displaced long-run, peaking at 60k–275k/year — small against the UK's ~450k normal annual job losses. Present these as a spread, never as one consensus number.
  • The deeper story is structural, not numerical. Automation reliably automates tasks, not whole jobs, and whether an occupation grows or shrinks is set by economics — demand elasticity, the productivity-vs-displacement balance, and whether a worker's tasks are complemented or substituted — not by the technology's raw capability.
  • This grounds the Workonomics thesis directly. In the canonical Acemoglu–Restrepo task model, automation "always" shifts the task content of production against labour: it can grow the pie while shrinking labour's slice. The offset — new tasks where labour has comparative advantage — is a separate, non-automatic force that has weakened over recent decades. Redistribution happens around displaced workers, not for them, and it carries a real, concentrated, poorly-cushioned adjustment cost.
  • The ATM/teller case is more double-edged than the popular telling. Teller headcount held roughly flat (even rose) for thirty years because demand was elastic — but tellers fell as a share of employment and declined absolutely after ~2010. The same Bessen "inverted-U" research shows that once demand satiates (steel, autos), the identical automation cuts jobs. Coexistence is real; so is the sting.

Key Findings (verified)

1. No single number — and the headline figures are task-exposure, not job losses. Goldman Sachs estimates AI "can potentially automate tasks that account for 25% of all work hours" in the US; in its base case wide-scale adoption takes ~10 years and "6–7% of workers will be displaced during that transition period," producing only a +0.6pp rise in unemployment if spread over a decade. The Tony Blair Institute models "1 to 3 million jobs could ultimately be displaced," peaking "between 60,000 and 275,000 jobs a year" against ~450,000/year of normal UK churn. Confidence: high. Votes: all 3-0. Sources: Goldman Sachs, Tony Blair Institute.

2. Speed of adoption is itself a major force. The same total displacement is benign spread over a decade but damaging if frontloaded: GS expects +0.6pp unemployment over a decade-long transition, "but if it's more frontloaded, the impacts on the economy are much larger." This is the displacement-to-recovery lag factor. Confidence: high. Vote: 3-0. Source: Goldman Sachs.

3. The core mechanism behind "pie grows, slice shrinks." Automation that replaces labour in tasks always shifts the task content of production against labour (the displacement effect), reducing labour's share of value added and potentially reducing labour demand even as it raises productivity — "automation therefore increases the size of the pie, but labour gets a smaller slice." The counterbalance is the reinstatement effect — new tasks where labour has comparative advantage — which "always raises the labour share and labour demand," but is a separate force that has weakened in recent decades. So "new job categories created" is a real but non-guaranteed offset. Confidence: high. Votes: displacement mechanism 2-1 (dissent was a ceteris-paribus scope quibble, not a refutation); reinstatement 3-0. Source: Acemoglu & Restrepo 2019, NBER w25684 (author is a 2024 Nobel laureate).

4. It's not "brilliant" automation that threatens jobs — it's "so-so" automation. Technologies that displace labour while delivering only small productivity gains do the most damage, because the modest productivity dividend is too thin to offset the displacement. A technology can be widely adopted yet net-negative for labour. Confidence: high. Vote: 3-0. Sources: Acemoglu & Restrepo; MIT Sloan ("The lure of so-so technology," e.g. self-checkout, automated phone systems).

5. Demand elasticity is the master switch (the engine under Modes B and C). Whether automating a task raises or lowers occupational employment is set by the Hicks–Marshall laws of derived demand — the elasticity of substitution, the elasticity of product demand, and the occupation's share of the wage bill. When demand is elastic, automation can grow employment (textiles, ATMs/tellers); once demand becomes inelastic/satiated, the same automation makes employment fall, producing an "inverted-U" path (steel, autos). Estimated price elasticities fall over an industry's life: cotton 2.13 (1810) → 0.02 (1995); steel 3.49 → 0.16; autos 6.77 → 0.15. Even where a sector's own demand is inelastic, freed income gets spent elsewhere, so income-elastic sectors can absorb the labour (Mode B). Confidence: high. Votes: all 3-0. Sources: Bessen, Brookings 2020; Autor 2015, "Why Are There Still So Many Jobs?".

6. Automation automates tasks, not whole jobs — and complement-vs-substitute is decided worker-by-worker. Across 271 detailed US occupations over ~60 years, only one (elevator operators) saw its decline chiefly attributed to automation. Workers who supply tasks complemented by automation gain; those supplying substituted tasks face falling wages: "a construction worker who is expert with a shovel but cannot drive an excavator will generally experience falling wages as automation advances." Confidence: high. Votes: all 3-0. Sources: Bessen, Brookings 2020; Autor 2015.

7. The ATM/teller case — coexistence, but with a sting. US ATMs roughly quadrupled (~100k → 400k, 1995–2010); bank-teller employment rose modestly (~500k → 550k, 1980–2010) — but tellers fell as a share of total US employment. The headcount held because ATMs cut the cost per branch (tellers per branch fell >⅓ between 1988–2004 while urban branches rose 40%+, aided by deregulation), and surviving teller roles shifted from cash-handling to relationship/sales work. Brookings separately cites 472,000 tellers in 2018 (a 10%+ rise since 2000) alongside 500,000+ ATMs. The "redistribution around, not for" part: the share decline and the post-2010 absolute fall (driven by online/mobile banking, not ATMs) are the cost borne by workers even as the industry grew. Confidence: high. Votes: all 3-0. Sources: Autor 2015; Brookings — "Not all robots take your job".

8. Historical evidence does not support mass unemployment — but it does show distributional harm. Occupations that used computers more grew faster (0.9%/year at the sample mean), with the net total-employment effect negligible (−0.07%/year) after inter-occupation substitution. Automating firms tend to grow faster than non-automating firms; only 1 of 12 reviewed studies found a substantial negative employment effect across all industries (manufacturing the most negative). Caveat: pre-generative-AI data — historical grounding, not an AI forecast. Confidence: high. Votes: all 3-0. Sources: Bessen 2015; Brookings 2020.

9. The effect is primarily distributional, and the adjustment burden is concentrated. 50–70% of the change in the US wage structure over 1980–2016 is accounted for by relative wage declines of workers specialised in routine tasks within rapidly-automating industries — and large inequality shifts occurred alongside only modest aggregate productivity gains (~3.4% cumulative TFP over 36 years), i.e. redistributive rather than growth-generating. In the five years after a firm automates, incumbent workers lose 11% of one year's earnings (€3,800), driven by spells of non-employment rather than wage cuts, with only ~13% recouped via the safety net. Caveats: the 11% is Dutch data (illustrative, revised to ~8–9% in a 2025 version); the wage-inequality attribution has rival explanations (skill-biased tech change, trade, deunionisation). Confidence: high. Votes: all 3-0. Sources: Acemoglu & Restrepo, Econometrica 2022; Brookings 2020.

Details

The full factor taxonomy (your three modes, expanded)

The user's seed model — Mode A (replaced wholesale), Mode B (expandable markets absorb labour), Mode C (inelastic markets need fewer humans), plus the market-share twist — maps cleanly onto canonical labour economics. The research expands it into a fuller set of forces, all evidenced above unless flagged:

  1. Demand elasticity — the master switch. Elastic → Mode B (automation can grow employment); inelastic/satiated → Mode C (the same automation cuts it). The "inverted-U" over an industry's life. (Finding 5)
  2. Task vs job — automation hits tasks, not whole occupations; whole-job destruction is historically rare. (Finding 6)
  3. Complement vs substitute — decided per worker, even within one job; complemented workers gain, substituted workers lose. (Finding 6)
  4. Productivity-vs-displacement balance ("so-so" vs "brilliant") — thin-productivity automation is the real labour threat; adoption alone doesn't predict harm. (Finding 4)
  5. Speed / timing of adoption — slow diffusion is absorbable, frontloaded shock is not. (Finding 2)
  6. New-task creation (reinstatement) — a real offset, but separate and not automatic, and weakening. (Finding 3)
  7. Share-of-employment vs absolute headcount — a sector can grow headcount while a role shrinks as a share (and later in absolute terms). The ATM/teller pattern. (Finding 7)
  8. Concentrated adjustment burden — even without mass unemployment, the cost lands hard on specific incumbents and is poorly cushioned. (Finding 9)
  9. The market-share twist (user's own, consistent with the above) — firms that adopt AI well take greater market share, but potentially a larger slice of a shrinking human-served market. Plausible and consistent with the task model, not independently verified here — present as reasoning, not citation.

Factors the user listed that this batch did NOT verify (source separately before asserting): capital intensity; regulation & liability; trust / human-preference premiums; reshoring; geographic concentration; firm-size / adoption concentration (the "~1 in 5 firms use AI / ~2–3% automate labour tasks" statistic — needs a primary source such as US Census BTOS or ONS BICS); and AV / contact-centre-specific forecasts. Driving/transport was covered only lightly and has no verified figure here.

Directional displacement-pressure bands (0–5)

⚠️ Important provenance note. These bands are reasoned editorial judgement built on the verified mechanisms above — they are NOT themselves verified figures. The tempting per-occupation numbers (e.g. telemarketers ~60%, admin 46%) were refuted in verification (see below), so precise sector percentages were deliberately not used. Read these as "relative pressure," with the reminder that pressure ≠ net headcount — demand elasticity (Finding 5) can convert task-automation into job growth or job loss.

Scale: 0 = negligible, 5 = severe. Timeframes: Near = 0–2y, Medium = 3–5y, Long = 6–10y.

Broad areaNearMediumLongNotes (US vs UK, mechanism)
Call centres / customer service344–5Already moving (cost-driven, codifiable, remotely delivered). Capped below 5 by a trust/quality premium — Klarna's reversal is the canonical limit of full automation. UK & US similar; offshored English-language voice work is most exposed.
Admin & back-office (data entry, bookkeeping, basic finance/legal ops)344Routine cognitive work — high substitution. Mode C risk: demand for "bookkeeping hours" is fairly inelastic, so automation tends to reduce headcount rather than expand output.
Software / knowledge work (coding, copywriting, analysis)22–33The augmentation-heavy case. Substitution concentrated at the junior/routine end (boilerplate code, commodity copy); complementarity at the senior/judgement end. Copywriting pressure > senior engineering. Mode B candidate: elastic demand for software may absorb freed capacity.
Driving / transport (light coverage)0–11–22–3Slowed by capital intensity, regulation & liability, and contested AV timelines — the brakes the other sectors lack. Long-run pressure real but back-loaded.

The spread of headline estimates (show this, don't pick one)

SourceFigureWhat it actually measures
Goldman Sachs~25% of US work hours automatable; 6–7% displaced over ~10y; +0.6pp unemploymentTask-hour exposure + transition displacement, not jobs lost
Tony Blair Institute1–3m UK jobs long-run; peak 60k–275k/yr (vs ~450k normal churn)Scenario-modelled UK long-run displacement
IPPR (UK)"up to 8 million UK jobs at risk" (worst-case "second wave")Worst-case scenario, widely reported as a ceiling — treat with care
Acemoglu & Restrepo50–70% of 1980–2016 US wage-structure change from routine-task automationDistributional effect, historical, not a headcount forecast

⚠️ Killed claims — do NOT cite these (refuted in verification)

These are tempting for the sector bands but failed the refute panel — listed here so they don't creep back in:

  • "23.8% of UK private-sector workforce time saved" by full AI adoption — vote 0-3. (Source: Tony Blair Institute.)
  • Per-occupation AI time-savings (telemarketers ~60%, admin 46%, sales 33.4%, construction 16%, etc.) — vote 1-2. (Source: Tony Blair Institute.)
  • "ATMs and tellers function as complements rather than substitutes" (the clean "robot as co-worker" framing) — vote 1-2. The fuller, verified picture (Finding 7) is more double-edged. (Source: Brookings.)
  • "Bank-teller FTE jobs grew ~2.0%/year since 2000" — vote 0-3. (Source: Bessen/BU.)
  • "Automation, not market power/markups/deunionisation, is the dominant driver of US wage inequality" — vote 1-2 (rival explanations remain live). (Source: Acemoglu & Restrepo, Econometrica.)

Open questions (not answered in this batch)

  1. Verified current adoption-concentration figure — the "~1 in 5 firms use AI / ~2–3% automate labour tasks" stat needs a primary source (US Census BTOS / ONS BICS).
  2. Sector-specific verified forecasts for call centres, admin/back-office, and software — needed to convert the editorial 0–5 bands into cited figures, since the TBI per-occupation numbers were refuted.
  3. Does generative AI behave like past task automation, or is it different? Autor (2024) raises the possibility that AI may reverse rather than extend past polarisation patterns — outside this batch's verified set.
  4. Trust/human-preference-premium and regulation/liability dynamics — listed as taxonomy factors but unverified here (key for AVs and regulated professional services).

Caveats on the whole report

  • Headline forecasts are not directly comparable (task-hours vs displacement vs long-run job counts) — always present as a spread.
  • Most of the structural economics (Acemoglu–Restrepo, Autor, Bessen) is grounded in past automation episodes (ATMs, computers, robots, 1980–2016). Whether generative AI behaves the same is an open debate — frame these as principles/precedent, not settled AI predictions.
  • The per-worker adjustment figure (~11% of a year's earnings) is Dutch data, revised to ~8–9% in 2025 — use it to illustrate the principle of concentrated, poorly-cushioned adjustment cost, not as a US/UK number.

Sources (primary unless noted)

Related reports:

More research →