DEFINITIONS — HOW WE MEASURE EACH METRIC
FULL REPLACEMENT (ACTUAL)
The role has been eliminated. No human is employed in it. AI performs the function end-to-end without human oversight. Measured by hiring deceleration, employment decline in exposed occupations, and firm-level workforce data.
PARTIAL ADOPTION (ACTUAL, NOW)
A human holds the role, but AI currently handles tasks representing 70% or more of that role's economic value. The human provides oversight, handles exceptions, and bears accountability. AI does the substantive work.
FULL REPLACEMENT POTENTIAL
Roles that could be fully eliminated today at economically competitive cost if companies chose to act. Grounded in MIT Iceberg Index (Nov 2025) — the threshold where AI cost ≤ human labor cost for the full role.
PARTIAL ADOPTION POTENTIAL
Roles where AI could currently handle tasks representing 70%+ of economic value with human oversight, if deployed. Synthesized from McKinsey task-hour data, Penn Wharton T2/T3/T4 categories, and economic value weighting adjustments.
NOTE ON METR BENCHMARK DATA
The METR time horizon figures cited throughout this site (e.g. Claude Opus 4.6: 14h 30m) measure the
human expert task completion time for tasks the AI can complete with 50% reliability — not the AI's own operating duration, and not a direct measure of labor substitution. METR researcher Sydney Von Arx:
"there are a bunch of ways that people are reading too much into the graph." We use these figures as a proxy for task complexity reach, adjusted for conversion lag (V3) and organizational adoption friction (V9) before applying to displacement estimates.
(MIT Technology Review, Feb 5, 2026)
THE FOUR METRICS — CURRENT STATE
METRIC 01 · ACTUAL
Full Replacement
Happening Now
HIGH CONFIDENCE
~1%
RANGE: 0.5% – 1.5% of US desk jobs
Roughly 325,000–975,000 desk roles have been fully eliminated. Concentrated in entry-level data processing, customer service tier-1, image/content tagging, document review, and basic translation workflows. These roles no longer appear in job postings and their former occupants have not been rehired into equivalent positions.
PRIMARY BASIS: Penn Wharton (2025) T4 employment data (-0.75% vs 2021) · MIT Iceberg Index · HBS 25-039 skill demand analysis
METRIC 02 · ACTUAL (NOW)
Partial Adoption
70%+ Economic Value, With Oversight
MEDIUM CONFIDENCE
~8%
RANGE: 6% – 12% of US desk jobs
Approximately 5–8 million desk workers are currently in roles where AI handles the majority of economically significant tasks, with the human serving primarily as reviewer, approver, or exception-handler. Highest concentration in software development, financial analysis, legal research, marketing content, and data analysis. This is the most contested estimate on this page — no study directly measures the 70% economic value threshold at current adoption rates. We synthesize it from Anthropic Economic Index usage data, HBS skill-demand shifts, and Penn Wharton T3 employment trends.
BASIS: AI Labs synthesis · Anthropic Economic Index (Feb 2025): 36% of occupations using AI for 25%+ of tasks · HBS 25-039: 24% decline in automatable skill demand · Penn Wharton T3 employment deceleration
METRIC 03 · POTENTIAL
Full Replacement
Technically & Economically Viable Today
HIGH CONFIDENCE
~11.7%
RANGE: 7% – 14% (friction-adjusted floor: ~7%)
If companies chose to act today, AI could fully replace roles representing approximately 11.7% of US employment at cost-competitive rates — roughly 19 million worker-equivalents, $1.2 trillion in annual wages. This is the MIT Iceberg ceiling: where AI capability meets economic viability. The gap between this figure and the ~1% actual displacement is the adoption lag — organizational inertia, integration cost, risk tolerance, and implementation time. This gap is what closes over the horizon windows below. Goldman Sachs' lower estimate of 6–7% reflects a friction-adjusted view that accounts for deployment costs and transition time.
PRIMARY BASIS: MIT Iceberg Index, Nov 2025 (MIT/Oak Ridge) · Goldman Sachs Global Economics Research (2025) · Acemoglu & Restrepo task-substitution framework
METRIC 04 · POTENTIAL
Partial Adoption
70%+ Economic Value Viable Today, If Deployed
LOWER CONFIDENCE — SEE REASONING
~40%
RANGE: 35% – 50% of US desk jobs
This is our most consequential and most uncertain estimate. If deployed today with human oversight, AI could handle tasks representing 70%+ of the economic value of approximately 40% of desk jobs. This is larger than the full-replacement ceiling (11.7%) because partial adoption — where the human remains but AI does the heavy lifting — requires a lower capability bar. The human absorbs edge cases, maintains relationships, and holds accountability. The 40% figure is derived by applying economic value weighting to McKinsey's 60–70% automatable task-hour data, discounting for the fact that high-value judgment tasks are disproportionately concentrated in the remaining 30–40% of work hours. This is a AI Labs synthesis — not a direct research finding.
BASIS: AI Labs synthesis from: McKinsey (2023) 60–70% automatable task-hours · Penn Wharton T2+T3+T4 categories · Eloundou et al. (2024) task exposure tiers · Economic value weighting adjustment (high-value tasks ≠ high-hour tasks)
THE ADOPTION GAP — WHAT'S VIABLE VS. WHAT'S HAPPENING
FULL REPLACEMENT
ACTUAL NOW
~1%
PARTIAL ADOPTION
ACTUAL NOW
~8%
FULL REPLACEMENT
VIABLE TODAY
~11.7%
PARTIAL ADOPTION
VIABLE TODAY
~40%
The space between each pair of bars is the adoption gap — organizational inertia, integration cost, and implementation time. The horizon windows below model how fast that gap closes.
AI LABS ANALYTICAL NOTE — WHY METRIC 04 IS LARGER THAN METRIC 03
The counterintuitive result — that partial adoption potential (~40%) is so much larger than full replacement potential (~11.7%) — deserves explanation. Full replacement requires AI to perform an entire role end-to-end at competitive cost. Partial adoption (70%+ economic value, human oversight) requires only that AI handle the bulk of substantive work, while a human covers the remainder.
The 11.7% MIT Iceberg figure is a strict threshold: the role must be fully substitutable at competitive cost. The 40% partial adoption estimate applies a looser test: can AI handle tasks representing 70% of the role's economic value, with a human overseeing? This is achievable for a much wider set of roles — including those where the remaining 30% involves judgment, accountability, or relationship management that currently requires a human.
The key methodological caveat: economic value is not evenly distributed across tasks. A lawyer's billable hour isn't uniform — the first-draft research is high-hour but lower-margin; the strategic judgment and client relationship are lower-hour but higher-margin. We adjust for this by discounting McKinsey's task-hour figures downward when applying them to economic value. This produces our 35–50% range, with 40% as the central estimate. This range carries the most uncertainty of any figure on this page.
HORIZON WINDOWS — WHEN DOES FULL REPLACEMENT HIT 50% OF DESK JOBS?
These windows model full replacement only (Metric 01 trajectory). The current ~1% actual / ~11.7% viable gap is the starting condition. Each scenario differs in how fast that gap closes and how fast the viable ceiling rises.
| SCENARIO |
KEY ASSUMPTION |
2027 |
2029 |
2031 |
50% WINDOW |
| Conservative LOW |
Friction persists; recursive loop doesn't materialize; capability plateau hits |
3–5% |
8–12% |
18–25% |
2040–2045 |
| Base Case CENTRAL |
Market-driven adoption; minimal corporate friction; moderate recursive acceleration |
8–12% |
22–28% |
40–50% |
2031–2033 |
| Recursive Model AGGRESSIVE |
AI self-improvement compresses doubling period; capability ceiling not hit before 2030 |
12–18% |
45–55% |
75–85% |
2029–2030 |
| Singularity SPECULATIVE |
Hyperbolic curve fully materializes; no compute or energy constraint before 2030 |
20–30% |
65–80% |
90%+ |
2028 |
WHAT MOVES THE WINDOW — KEY VARIABLES
ACCELERATES TIMELINE
API cost declining ~75%/year · Agentic AI handling multi-step workflows · First-mover earnings beats forcing board-level adoption mandates · AI doubling period shortening · Energy infrastructure catching up
DECELERATES TIMELINE
Energy grid constraints · Capability ceiling / alignment problems · Liability and regulatory exposure for AI-made errors · Integration complexity at scale · 70–85% of AI initiatives fail to meet expected outcomes (MIT/RAND)
GENUINE UNKNOWNS
Whether new task categories emerge fast enough to absorb displaced workers · Degree to which productivity gains raise wages and sustain demand · Policy responses (UBI, hiring floors, retraining mandates) · Whether recursive acceleration hits a ceiling