A task-based labor model (Acemoglu–Restrepo family): jobs are bundles of tasks; AI absorbs task-share as capability × adoption; three counterforces fight back — demand growth from cheaper output, new-occupation creation, and the physical bottleneck of robot fleets. Annual steps, 2026–2036, six regions × six task classes.
1 · Baseline: who does what today
Global employment ≈ 3.55 billion (ILO-modeled). Regions: US 168M · China 735M · EU+UK 250M · India 600M · Other advanced 155M · Rest of world 1,640M. Each region's work-hours are split across six task classes, calibrated to ILO sector shares (agriculture 26% of world employment) and IMF exposure findings (≈60% of advanced-economy jobs AI-exposed vs ≈40% in emerging markets):
- Routine cognitive — clerical, support, bookkeeping
- Non-routine cognitive — coding, analysis, law, finance, design
- Creative & strategic — research, senior judgment, invention
- Interpersonal & care — health, teaching, hospitality, relationship sales
- Routine physical — assembly, warehousing, driving, field agriculture
- Dexterous physical — trades, repair, construction, complex manual work
2 · Capability: stacked S-curves, not a straight exponential
Cognitive capability follows a logistic (S-curve) per task class whose speed comes from the doubling-time dial (METR: ~7-month doubling in autonomous task length 2019–25, ~4 months in 2024–25; frontier agents at ~half-day tasks by early 2026). A 0.35 translation factor discounts benchmark progress into real-task coverage — a doubling of task-length horizon does not double the share of jobs' tasks covered. Each class has a ceiling (routine cognitive 95% … interpersonal 50%, a human-preference floor). Capability is not one curve: Wave 2 (the arrival dial) models foundation models transferring to robots, re-accelerating physical and interpersonal capability as a second S-curve stacked on the first — the "saturate, then a new round" pattern of every prior technology. And the ceilings themselves are generation-indexed: preference floors are measured against today's alternative (a Gen-1 bot doing Gen-3 work), so as capability generations advance the ceilings drift upward (interpersonal 50%→72%, creative 65%→85%), holding only an irreducible core for genuinely relational work.
3 · Adoption: tipping points, not constant crawl
Adoption is a logistic per region×class — but its speed is coupled to capability: below a reliability threshold (≈35–55% of tasks, by class), adoption runs at pilot speed (~0.15× base — the slow phase we are in now); once capability crosses the threshold, speed jumps ~1.85× (the launch). For physical work, speed is additionally scaled by wage-vs-machine cost parity: robot labor at ~$10/hr all-in (2026, falling ~16%/yr) tips against a $36/hr US wage years before a $2.50/hr Indian wage — cheap human labor is its own protection. 2026 starting points anchor to US Census BTOS (18% of firms, 32% employment-weighted, 50–70% in large information/professional/finance firms) and Stanford AI Index 2026 (88% of surveyed orgs). Regional base speeds: China ≈ US > other advanced > EU > India > rest of world; India's IT-export sector adopts at client-country speed.
4 · Robotics: the physical bottleneck
Physical task classes cannot automate beyond the installed fleet, in worker-equivalents: industrial robots 4.66M units operating in 2024, 542k installed/yr, ~2 worker-equivalents each (multi-shift), growing 8%/yr; autonomous vehicles from ~100k (Waymo alone: 450k+ paid rides/week in 2026), growing ~60%/yr; humanoids from ~150k units in 2026 (Morgan Stanley: 50k China shipments in 2026 → 446k/yr by 2030, 13M in use by 2035; Tesla targeting 10k→50k→1M/yr capacity), ramping on a logistic manufacturing curve — near-exponential early at the fleet-growth dial's rate, saturating toward a 250M-unit carrying capacity (factories are S-curves too; for scale, world vehicle production is ~90M units/yr). Worker-equivalence is explicit, not 1:1 — utilization asymmetry means one machine ≠ one job: industrial robots 2.0 equivalents (multi-shift); autonomous vehicles 2.5 (a robotaxi runs ~20 hr/day vs an 8-hr driver shift and truckers are capped at 11 driving hours — today's math: Waymo's 500k weekly rides across ~3,000 vehicles ≈ 165 rides/vehicle/week vs ~100 for a full-time rideshare driver, ≈1.6 driver-workloads per vehicle and rising with utilization); warehouse AMRs 1.0 (Amazon alone crossed 1M robots against ~1.5M employees, with its Vulcan system handling 75% of item types); humanoids 0.8→1.6 as uptime improves. The AV fleet starts from ~10k truly driverless commercial units in 2026 (Waymo ~3,000, Apollo Go ~1,000, Pony.ai 1,400+, WeRide ~1,000) growing ~90%/yr on a logistic capped at 25M; AMRs from ~2.5M at 30%/yr capped at 40M. Fleet is allocated China 45% · US 18% · EU 12% · other advanced 10% · India 5% · RoW 10%.
5 · The three counterforces
(a) Demand elasticity: cheaper output → more demand → workers re-absorbed in-industry (why ATMs coincided with more bank tellers for 20 years). (b) New-task creation: new occupations = elasticity × displaced work, with a 2-year lag, allocated 40% non-routine cognitive · 25% interpersonal · 15% creative · 12% dexterous physical · 8% other. Crucially, new tasks join the automatable pool — they are made of tasks too, and AI eats them at their class's automation rate in later years. The optimists' accounting error is granting new work permanent immunity; this model does not. Two further mechanisms: the tool→worker transition — below full substitution, AI compresses headcount on remaining work (augmentation: 0.55 productivity factor on cognitive classes, 0.25–0.35 on physical/interpersonal), and substituted work retains only a supervision residual of 1 human per span-of-control dial digital workers, a span that widens as reliability grows. And a chain-completion kicker: when make + move + coordinate (routine physical, routine cognitive, non-routine cognitive) jointly cross ~45% automation in a region — the fully autonomous build→warehouse→ship→sell→deliver chain — coordination labor collapses by up to a further 35%, capturing the synergy of automation vectors compounding each other. (c) Background growth: GDP-driven demand growth per region (India 5.5%/yr … EU 1.2%/yr) and labor-force drift (China −0.3%/yr, India +1.2%/yr, RoW +1.6%/yr).
6b · Obviation: work that stops existing
Distinct from automation: when your agent talks to their agent, the interaction is not performed by AI — it ceases to exist. No chat queue, no escalation, no scheduling email. Switchboard operators weren't replaced by robot operators; direct dialing deleted the task. The model tags each class's intermediation share — work existing only because humans can't interface directly (routine cognitive 45% · non-routine cognitive 20% · interpersonal 15% · creative 5% · physical 2–5%) — and eliminates it as consumer personal-agent penetration (a logistic gated on high agent reliability, from ~3% in 2026) meets enterprise-side agent adoption, scaled by the agent↔agent obviation dial. Eliminated work leaves the pool entirely — it consumes no AI compute, no robots, and creates displacement with nothing to supervise. It appears as the third band in the top chart. The displaced value recycles through the same new-task engine as all displacement.
6 · What the model does NOT claim
It does not predict. It makes the debate quantitative: the "past repeats" case is literally the new-task-elasticity dial near 0.85; the pessimist case is the same dial near 0.3 plus fast capability. It also omits wages (displacement shows up partly as wage compression before job loss), policy response, and AI-driven scientific acceleration beyond what the new-task dial captures — all of which would matter in reality. The model's reliability gates cut both ways, and the evidence demands both: tool penetration is deeper than surveys suggest (Google reports ~75% of new code AI-generated; young software developers −20% since 2022), yet durable unsupervised substitution keeps failing below the threshold — Klarna publicly walked back its showcase AI customer-service replacement and rehired humans for judgment-heavy work after satisfaction dropped, and surveys find a majority of firms regretting AI-first staff cuts. That is exactly the tool→worker distinction: augmentation compresses now; substitution waits for reliability. Read the Klarna reversal through the model's gates: it is what mid-gate premature adoption looks like — a Gen-1 product doing Gen-3 work — and the gates are indexed to capability, not calendar, so when a Gen-5 product does Gen-3 work the same deployment succeeds and adoption rips. Deeper still, the honest epistemic limit: obviation of enumerable intermediation work is modelable; wholly new interaction patterns that delete work categories we cannot yet name are not — and symmetrically, neither are the new categories they create. The obviation and new-task dials bracket that uncertainty rather than pretending to resolve it. Early real-world signal worth carrying: entry-level (22–25) employment in the most AI-exposed US occupations fell ~13% relative since ChatGPT while older cohorts in the same jobs grew — displacement is arriving rung-by-rung, youngest first, exactly as a task model predicts.