EOL evaluates whether AI, robotics, and automation are moving from technical possibility to economically meaningful adoption, task substitution, and measurable labor-market effects. The goal is not to produce an automation score. It is to build an occupation-level forecast that can change as the evidence changes, with the supporting evidence and intermediate judgments open to inspection.
EOL begins with observed occupational structure and the official BLS projection, then asks whether emerging technology is altering the relationships that historically connected output, tasks, staffing, and employment.
EOL starts with the BLS structural projection, evaluates six stages of technological transition, then asks whether the combined evidence supports a different employment path. The stages inform judgment; they are not converted into a mechanical score.
Panel scope, why exposure is not displacement, how BLS is incorporated, and the research logic underlying the framework.
EOL studies 20 deliberately selected occupations that represent a broad cross-section of work relevant to AI and robotics, spanning knowledge-intensive roles and physically oriented jobs. This focused design allows detailed analysis of different routes from technology to changes in labor demand without attempting to cover every occupation.
The sentinel panel is a group followed consistently over time for early evidence of meaningful technological and labor-market transitions. Its diversity of tasks, training and credential requirements, labor-supply conditions, and adoption environments supports comparison across distinct mechanisms and helps EOL interpret implications that may extend beyond the occupations themselves. Because the panel is purposefully selected, its findings inform wider labor-market understanding but are not mechanically extrapolated into a national employment total.
EOL considers the current 20 occupations a sufficiently broad and useful cross-section for its analytical objectives, and substantial expansion is not presently planned. This does not preclude adding, replacing, or reconsidering occupations when future technological developments or research needs warrant it. Any changes to panel composition would be documented to preserve transparency and longitudinal comparisons.
Measures of technological exposure identify where AI or robotics may matter. They do not establish that a technology is economical, adopted, capable of absorbing enough of the work, or reducing employment.
EOL therefore treats capability and exposure evidence as the beginning of the analysis. An occupation can be highly exposed while remaining strongly dependent on human labor because of regulation, physical presence, accountability, customer preference, workflow complexity, or rapidly growing underlying demand.
Can more output increasingly be produced without proportional additions of human labor, and is that change beginning to affect hiring, staffing, hours, or employment?
Most long-horizon economic forecasts are necessarily anchored in historical relationships. When the future develops broadly like the past, that is a major strength. The same approach becomes more vulnerable when the economy experiences a genuine structural break, a change large enough that previously stable relationships no longer provide a reliable guide to what comes next.
This is a well-established forecasting problem. Major discontinuities can cause historically grounded models to lag a changing underlying regime, a problem examined directly in the literature on structural breaks, including Clements and Hendry (2006). AI and robotics create precisely this concern. If capability, cost, adoption, or the amount of work machines can perform changes much faster than in previous decades, the relationship between output and human labor can diverge sharply from its historical reference points.
BLS itself explicitly recognizes this limitation. Its current employment-projections methodology assumes that productivity and technological progress will remain broadly consistent with historical experience, while noting that unusually rapid technological progress can make historical relationships less reliable. Recent advances in AI are one reason BLS identifies that possibility.
BLS is not blind to technological change. It researches emerging technologies, productivity, occupational staffing patterns, industry demand, and other structural developments and makes forward-looking adjustments when the evidence is sufficiently convincing. The issue for EOL is therefore not whether BLS considers technology, but whether a particular occupation may be entering a technological regime that departs materially from the relationships embedded in the current projection.
What employment outcomes remain supportable if AI and robotics produce a meaningful break from the technological and labor relationships embedded in the historical record?
BLS serves as EOL's structural reference point, not as a no-automation baseline. Before EOL moves away from the BLS path, it reviews what the current projection already appears to capture and defines only the additional departure that current evidence can support.
Review BLS methodology, occupation and industry trends, and published factors affecting occupational utilization. Historical technology, productivity, demand, and staffing effects are presumed to be in the structural reference unless evidence shows otherwise.
State exactly what EOL believes may differ from the current BLS path, such as faster diffusion, broader task coverage, greater staffing leverage, a different demand response, or a new constraint on human labor requirements.
When BLS publishes a new projection or materially changes its methodology, EOL reassesses how much of the prior residual departure has been absorbed into the new baseline. The EOL-minus-BLS gap is never carried forward mechanically.
Where historical continuity remains a reasonable assumption, the EOL range should remain relatively close to the structural reference. Where the six-stage evidence indicates that capability, economics, diffusion, task subsumption, or human labor requirements are moving beyond the regime represented in BLS, EOL can support a larger residual departure.
EOL does not subtract a separate automation penalty from BLS. Effects already reflected in historical productivity, occupational staffing patterns, demand, or explicit BLS utilization adjustments are not counted again. EOL departs only for the incremental occupation-specific mechanism supported by current evidence.
The review is deliberately qualitative where BLS does not publish a numerical decomposition. EOL will not invent a precise percentage for the technology effect already embedded in BLS. As later BLS releases absorb new evidence, EOL reassesses the residual departure rather than preserving an old gap.
EOL combines established ideas from economic research and institutional forecasting while remaining explicit about where its own approach is less formal.
Partial-identification research asks what conclusions remain supportable under stated evidence and assumptions when a single parameter cannot credibly be point-identified. EOL applies that logic prospectively: current evidence constrains a set of economically coherent assumptions, which in turn constrains the employment outcomes EOL is willing to publish. See Tamer (2010).
Economic institutions routinely distinguish a baseline from outcomes under alternative assumptions. EOL uses the BLS structural projection as its baseline and makes departures conditional on explicit occupation-level mechanisms and evidence rather than treating the baseline as a technology-free world.
EOL tests what would be required to move materially outside the published range and preserves forecast snapshots for later evaluation. This follows the broader institutional discipline of sensitivity analysis and retrospective forecast evaluation used by organizations such as the Congressional Budget Office.
EOL ranges are not formal econometric identified sets, statistical confidence intervals, or confidence sets. Some restrictions are prospective judgments informed by evidence rather than statistically estimated population parameters. The partial-identification reference describes the logic of bounding claims under uncertainty, not a claim that EOL has implemented a formal set-identification estimator.
The six-stage assessment and the interpretive lenses used to characterize the accumulated evidence.
Every occupation is assessed through the same sequence from technological capability to observable labor-market effects. The stages are judgments about the current state and direction of the transition, not six mechanically combined scores.
What can the technology do today in relevant occupational environments?
⌄Does substitution or augmentation make economic sense at current or near-current capability?
⌄Are capable and economic systems actually being deployed?
⌄How much of the occupational task bundle is moving to technology?
⌄How strongly does additional output still require additional human labor?
⌄Are upstream changes beginning to appear in hiring, hours, staffing ratios, wages, or employment?
⌄After the six stages are assessed, EOL summarizes what the evidence implies for the occupation’s structural trajectory without turning the framework into a black-box score.
Evidence Tilt summarizes the direction and weight of the current evidence relative to the occupation’s structural trajectory. It is categorical, not a probability, percentile, automation share, or mathematical adjustment.
The test asks whether the technological and labor relationships underlying conventional historical projections remain reliable enough to serve as the central forecast. Its categories are Reasonable, Increasingly questionable, and Unlikely reliable as the central assumption.
These assessments are analyst-directed judgments rather than outputs produced mechanically by a model or consensus panel. EOL uses a standardized framework together with AI-assisted research, source verification, computational checking, and multiple rounds of QA. Final responsibility for interpreting the evidence and publishing the forecast ranges remains with EOL, while the underlying evidence, assumptions, Forecast Bridge, and revision history are published so the judgments can be inspected, challenged, and revised.
How EOL moves from the structural reference to bounded working outlooks and separately interprets labor supply.
The forecast is built as an evidence-constrained departure from the structural reference, not as an automation score and not as a mechanical average of outside forecasts.
Start with the official 2025–35 employment path and the demand, productivity, staffing, and technology effects already embedded in it.
Use the BLS Incorporation Review, six stages, Historical Continuity, demand offsets, and occupation-specific mechanisms to define what may differ from the current BLS path.
Test the upper and lower endpoints against economically coherent assumptions supported by the evidence, then test materially outside the range.
External probabilistic forecasts such as Metaculus provide independent calibration where EOL can establish a defensible occupation match. They can challenge the range, but they are not mechanically averaged into it.
A constrained near-term outlook. Because BLS does not publish a 2030 occupation point in the 2025–35 release, EOL uses a neutral bridge benchmark derived by carrying the BLS ten-year growth path halfway through the decade at a constant compound rate. This is an EOL analytical reference, not a BLS forecast. The constant-compound path is itself a modeling assumption rather than a claim about when BLS-projected growth will occur; EOL's 2030 range is separately constrained by the current employment position, adoption, economics, and realistic diffusion time.
A medium-term outlook anchored directly to the published BLS 2025–35 occupation projection. The longer horizon allows materially more time for capability improvement, deployment, workflow redesign, supply adjustment, and substitution to compound.
EOL identifies the economically important quantities linking technological change to employment, uses current evidence to constrain those quantities, and publishes the range of employment outcomes that remains economically coherent. The six stages are not converted into numerical scores. Instead, they constrain the assumptions that can support each endpoint.
Each endpoint is an approximate scenario boundary that must be reconcilable with the structural reference, the residual departure, demand offsets, and the occupation-specific mechanism. EOL then asks what would be required to move materially beyond the range. The bounds are not statistical confidence intervals, do not assign equal or specified probabilities inside the range, and do not imply that outcomes outside the range are impossible. Small differences immediately around an endpoint are not economically meaningful.
Every occupation page places the Forecast Bridge after the six-stage assessment, evidence synthesis, and labor-supply analysis. The bridge then summarizes, in simple language, what BLS already captures, why EOL differs, what supports the lower and upper edges of both horizons, and which binding factors keep the range from moving materially higher or lower. It is a synthesis of the analysis above, not a second analysis layer.
Range width is not a statistical confidence interval or a generic confidence score. It represents the distance between materially different employment paths that remain supportable after the six-stage assessment. Ranges widen when unresolved forks in capability, economics, diffusion, task subsumption, demand response, or the translation of productivity into staffing could still produce very different outcomes. They narrow when durable constraints such as high Human Labor Dependency, regulation, physical bottlenecks, strong underlying demand, or already-mature adoption rule out much of that scenario space.
The location of the range reflects EOL's assessment of the direction and magnitude of likely employment effects; its width reflects how far apart the remaining evidence-supported pathways are. Two occupations can therefore have similar Evidence Tilts but materially different range widths.
SOC 37-2012 · BLS 2025–35 structural projection +0.6%
BLS projects +0.6% employment change from 2025 to 2035. BLS already reflects historical cleaning productivity, staffing patterns, and demand changes such as reduced housekeeping frequency in some hotels. EOL does not treat existing cleaning mechanization as a new effect.
The residual departure is specialized and eventually compositional cleaning automation beyond utilization changes and cleaning productivity already embedded in the structural baseline.
Key mechanism Compositional cleaning automation plus consumer-side substitution for purchased cleaning services.
The 2030 range remains close to flat. By 2035, specialized cleaning automation can materially reduce labor intensity, but the core range no longer assumes generalized room or household automation reaches stress-case scale.
Counterpressure Cleaning-service demand remains durable, but both provider automation and household robot adoption reduce human hours purchased.
The downside gives greater weight to the displacement mechanism and faster diffusion or task subsumption. Developments that would move the view further in this direction include: The prior -10% 2035 lower edge remains a serious adverse tail, not a core bound.
The upper edge gives greater weight to countervailing demand and persistent human bottlenecks. Developments that would move the view further in this direction include: A materially higher outcome would require the employment-supportive conditions to dominate more strongly than current evidence supports: Multi-task robotics remains unreliable or uneconomic; consumers use robots only as supplements while maintaining professional cleaning frequency; human finishing and exception work remains dominant.
Given the current evidence, EOL’s working range is -1% to +2% by 2030 and -8% to +1% by 2035. Each endpoint has been reconciled to the structural reference and the occupation-specific mechanism; materially outside outcomes require assumptions current evidence does not yet support.
SOC 15-1252 · BLS 2025–35 structural projection +10.2%
BLS projects +10.2% employment change from 2025 to 2035. BLS already reflects historical software-demand growth, productivity change, occupational staffing patterns, and analyst judgment about emerging AI. EOL does not subtract today's coding automation a second time.
The residual departure is faster agentic and recursive development plus direct non-developer software creation, reducing professional developer-hours per unit of software faster than the BLS structural path implies.
Key mechanism Agentic workflow substitution; output per developer.
The 2030 range still permits positive developer growth as demand expands and production bottlenecks persist. By 2035, the downside reflects smaller teams and direct non-developer software creation, while the +4% upper edge preserves the possibility that software demand and persistent human bottlenecks outpace automation.
Counterpressure AI-induced software demand and new applications can offset productivity, but only if they require proportional developer labor.
The downside gives greater weight to the displacement mechanism and faster diffusion or task subsumption. Developments that would move the view further in this direction include: A materially lower outcome would require the greater-displacement conditions to be established at broader scale than current evidence supports: Sustained software output growth with smaller teams; reliable production-grade project autonomy; larger shares of end-to-end workflows completed by agents; rising output per developer without proportional hiring.
The upper edge gives greater weight to countervailing demand and persistent human bottlenecks. Developments that would move the view further in this direction include: A materially higher outcome would require the employment-supportive conditions to dominate more strongly than current evidence supports: Strong developer hiring and team expansion alongside AI adoption; persistent architecture, integration, security and customer bottlenecks; AI-induced software demand consistently creating more human work than automation removes.
Given the current evidence, EOL’s working range is -2% to +5% by 2030 and -16% to +4% by 2035. Each endpoint has been reconciled to the structural reference and the occupation-specific mechanism; materially outside outcomes require assumptions current evidence does not yet support.
SOC 53-3032 · BLS 2025–35 structural projection +3.8%
BLS projects +3.8% employment change from 2025 to 2035. BLS already recognizes autonomous haulage in mining and incorporates historical trucking productivity and freight demand. EOL does not count those effects again; the question is whether driverless long-haul deployment expands materially beyond the automation embedded in the current projection.
The residual departure is commercial driverless line-haul substitution on autonomous-eligible routes, translated through eligible employment share, penetration, human driver-hours removed and retained residual work.
Key mechanism Route-level job substitution in autonomous-eligible long-haul.
The 2030 range remains near flat because driverless deployment starts from a small fleet base. By 2035, the core range spans limited national diffusion through an aggressive autonomous-line-haul path; the prior -14% case moves outside the core range.
Counterpressure Freight growth is the principal offset; it weakens as marginal freight is carried by autonomous trucks.
The downside gives greater weight to the displacement mechanism and faster diffusion or task subsumption. Developments that would move the view further in this direction include: The prior -14% 2035 lower edge remains an aggressive stress case requiring very high autonomous-eligible penetration and driver-hour removal.
The upper edge gives greater weight to countervailing demand and persistent human bottlenecks. Developments that would move the view further in this direction include: A materially higher outcome would require the employment-supportive conditions to dominate more strongly than current evidence supports: Regulatory, weather and edge-case constraints keep autonomy on limited routes; human labor remains necessary across enough of the journey; freight growth materially outpaces autonomous deployment.
Given the current evidence, EOL’s working range is -2% to +2% by 2030 and -10% to +2% by 2035. Each endpoint has been reconciled to the structural reference and the occupation-specific mechanism; materially outside outcomes require assumptions current evidence does not yet support.
Registered Nurses provide a useful counterexample. BLS projects +5.6% employment growth through 2035, while EOL's current range is +4% to +9%. Strong underlying demand, labor scarcity, and continued high Human Labor Dependency mean that technological productivity gains can expand care capacity rather than translate into employment decline over this horizon.
These examples show the reasoning discipline, not a universal coefficient system. Different occupations can reach similar numerical ranges through different mechanisms, adoption constraints, demand offsets, and human labor requirements.
Employment outcomes depend on both the demand for workers and the supply of people available to perform the work. EOL therefore assesses supply separately from the core labor-demand forecast, then uses it to interpret how technological transformation is likely to appear in the labor market.
EOL asks whether entrant pipelines, replacement flows, occupational transfers, and current shortages are aligned with the demand trajectory implied by the technology assessment. An expanding pipeline entering an occupation whose labor demand is weakening can create pressure first in entry-level hiring, placement, wages, or underemployment, even before broad employment declines appear. Conversely, a constrained pipeline can allow automation to absorb unfilled openings or reduce replacement hiring before it displaces incumbent workers.
Supply works differently across occupations. Physicians move through medical school, residency, and licensure. Accountants and engineers typically enter through degree programs. Truck drivers move through commercial licensing and training. Many service and manual occupations have no discrete education pipeline and instead depend on young-worker entry, occupational transfers, turnover, and employer training.
Current shortages are treated separately from pipeline evidence. A shortage shows that available labor is not matching current demand; it does not by itself show whether the future entrant pipeline is expanding or contracting. Supply does not receive a mechanical numerical weight in the forecast. Instead, it helps EOL test adoption incentives, identify emerging demand-supply mismatches, and interpret whether a changing demand trajectory is likely to appear through fewer replacement hires, weaker entry-level hiring, underemployment, wage pressure, or eventually broader displacement.
How uncertainty, falsifiability, revision, source translation, and the published record are handled.
EOL separates confidence in the direction of change from confidence in its magnitude and makes the conditions that could change each forecast explicit.
How confident EOL is that the occupation will deviate from its structural trajectory in the stated direction.
How confident EOL is in the size of the 2030 or 2035 range. Longer horizons generally carry greater uncertainty.
Concrete upstream developments that would move the outlook toward greater displacement or greater employment.
Occupation-level technological transitions are uncertain in timing, diffusion, and magnitude. EOL therefore uses ranges to represent a defensible set of outcomes rather than imply precision the evidence cannot support.
Working outlooks are current assessments, not fixed terminal views. Material new evidence can change a stage assessment, the interpretation of historical continuity, the expected inflection point, or the range itself. Prior assessments are retained so revisions remain visible.
EOL gives particular attention to upstream indicators that can falsify the mechanism behind a forecast. Examples include reliable end-to-end task performance, changing economics, sustained adoption, staffing leverage, output per worker, replacement hiring, and occupation-specific demand offsets.
EOL separates source facts from analytical translation and final judgment so readers can see where the evidence ends and the forecast begins.
Government data, academic research, company deployment evidence, industry surveys, forecasting platforms, and other material sources.
The occupation match, transition stage, mechanism, relevance, limitations, and whether the evidence is supporting or countervailing.
The stage assessment, evidence synthesis, confidence, forecast rationale, or working range when warranted.
The evidence item, stage movement, forecast hold or change, and resulting revision record remain inspectable over time.
EOL does not assign artificial numerical source weights. Evidence strength, independence, occupation match, and limitations are tracked explicitly. Company and vendor evidence can be important for capability, deployment, adoption, and economics, but it is not treated as neutral evidence of labor-market effects.
The core EOL assessment asks what technology is doing to work now, and how that evidence informs employment forecasts. The Scenario Lab asks an earlier question: what could make technological progress suddenly move faster? It follows five emerging developments, including ways they might reinforce each other, using the same six questions about technology, work and jobs.
For each pathway, the Lab distinguishes current binding constraints from secondary constraints, identifies observable tests for their release, and records counterevidence. These are qualitative scenario judgments, not numerical stage scores or predictions that a breakthrough will occur. A scenario becomes eligible to affect EOL's forecast only after new evidence passes the normal six-stage assessment and independent review.
Explore the Scenario Lab and its source limitations →Signal Tracker shows the evidence record, stage movements, forecast holds and changes, and the historical/live record over time.