Claude was wrong about job impacts

Anthropic published a research analysis arguing that Claude and similar AI systems will change the composition of jobs rather than eliminate them outright, using an illustrative 40-hour production manager week in which assistance and automation shares are applied task by task. In the production manager scenario, the same baseline workload falls to 30.63592 actual human hours after adding 1.2728 hours of support work and 3.1820 hours of assumed new work, with automation shares ranging from 2.0% for checking safety and quality to 70.0% for tracking costs and progress. The analysis states these are changes in task time, not counts of jobs lost or gained, and draws its task taxonomy from O*NET while noting the allocations are not measured time diaries.

Research & perspectives https://foundationproject.net/research Work, technology, and the roof over our heads The work above us. AI is changing the office. Robotics may change the roof. What happens when both transitions arrive together? Start with the work. The economics follows. The production manager Before the first ladder goes up. A roof must be inspected. Materials estimated. A crew scheduled. A job title contains a whole collection of tasks. A week is a bundle of work. Some tasks take an hour. Others take a day. Let the boxes fill, and a familiar workweek comes into view. An illustrative 40-hour week, not a measured time diary. First, a better pair of hands. An assistant drafts the estimate, checks an order, or proposes tomorrow’s schedule. A person still does the work, with help. Then, someone else does part of it. Some tasks can be delegated to an agent. The work still exists. What changes is who carries it out. One conditional scenario, not a forecast of adoption. And the job changes shape. New responsibilities can appear: supervising digital work, improving workflows, and teaching people to use them. Those human roles must be designed and supported. Their creation is not automatic. Read the exact production manager task figures Illustrative 40-hour baseline, Moderate terminal assumptions. The task taxonomy draws on O NET; these allocations are not measured time diaries. Assistance and automation shares are fractions of each original task. | Original task allocation | | | | |---|---|---|---| | Task | Baseline hours | Assisted | Automated | |---|---|---|---| | Inspect the roof | 4.8 | 17.0% | 15.0% | | Build the estimate | 6.4 | 10.0% | 50.0% | | Order materials | 4.8 | 10.0% | 50.0% | | Schedule the crew | 6.0 | 12.0% | 40.0% | | Track costs & progress | 4.8 | 6.0% | 70.0% | | Check safety & quality | 6.0 | 19.6% | 2.0% | | Coordinate people | 4.8 | 18.0% | 10.0% | | Resolve exceptions | 2.4 | 19.6% | 2.0% | Added support: 1.2728 human hours. Assumed new work: 3.1820 hours. Actual human hours for the same baseline workload, including these additions: 30.63592. These are changes in task time, not counts of jobs lost or gained. The opening explains work changing; scroll position is not a date. The scenario explorer below varies these assumptions. But a roof is more than a plan. Someone still has to make it hold. The roofer Now step onto the roof. Rig safe access. Repair the deck. Fit the flashing around an awkward corner. Physical work is a bundle of tasks, too. A different kind of workweek. Every roof brings its own materials, weather, access and surprises. The work takes judgment as well as strength. The same illustrative 40-hour baseline and scale. Help can reach the jobsite. Better information and tools may help people work more effectively. Assistance changes a task before it replaces one. And machines may do some of it. Consider a future in which specialized machines take on portions of material handling or roof installation. A scenario assumption. Today’s demonstrations do not establish whole-roof autonomy. Human work has to evolve, too. Setting up machines. Supervising unfamiliar equipment. Trialing a safer method, then teaching it to the next crew. A responsible transition makes room for people to learn. Read the exact roofer task figures Illustrative 40-hour baseline, Moderate terminal assumptions. The task taxonomy draws on O NET; these allocations are not measured time diaries. Assistance and automation shares are fractions of each original task. | Original task allocation | | | | |---|---|---|---| | Task | Baseline hours | Assisted | Automated | |---|---|---|---| | Rig safe access | 4.0 | 20.0% | 0.0% | | Move materials | 4.8 | 14.0% | 30.0% | | Remove old roofing | 5.6 | 18.0% | 10.0% | | Repair the deck | 4.8 | 19.6% | 2.0% | | Lay underlayment | 4.8 | 17.0% | 15.0% | | Install roof covering | 8.8 | 14.4% | 28.0% | | Fit the flashing | 4.8 | 19.8% | 1.0% | | Inspect & hand over | 2.4 | 19.4% | 3.0% | Added support: 0.8100 human hours. Assumed new work: 1.3500 hours. Actual human hours for the same baseline workload, including these additions: 35.37600. These are changes in task time, not counts of jobs lost or gained. The opening explains work changing; scroll position is not a date. The scenario explorer below varies these assumptions. The question that connects the two What if the next job is changing, too? A move into the trades takes time. The work at the other end may change while a person is learning it. More qualified workers and fewer human hours per roof can arrive together. Stronger demand and new responsibilities can offset that pressure. The balance matters. That possibility does not close the door on a better future. It asks us to build the training, responsibility, and shared gains that make one possible. Explore three paths through that transition. The scenario engine More output is only part of the story. A growing economy can still leave particular workers under pressure. Explore three conditional paths in which digital automation, physical robotics, demand, and human adaptation move at different speeds. Illustrative scenarios, not forecasts. Model years have no assigned calendar dates. No scenario is assigned a probability. Read the assumptions. methods Digital displacement and bounded physical automation overlap. Adjustment works, but roofing wages still face pressure. Moderate · model year 6. GDP 109.8; unemployment 4.1%; roofing wage index 94.5. Read annual scenario values | Moderate scenario. GDP and wage indices baseline 100; unemployment and labor share in percent. | | | | | | | |---|---|---|---|---|---|---| | Model year | GDP | Unemployment | Labor share | Knowledge wages | Non-knowledge wages | Roofing wages | |---|---|---|---|---|---|---| | 0 | 100.00 | 3.80 | 60.00 | 100.00 | 100.00 | 100.00 | | 1 | 101.38 | 3.72 | 59.14 | 99.88 | 100.26 | 100.04 | | 2 | 102.77 | 3.67 | 58.28 | 99.74 | 100.55 | 99.62 | | 3 | 104.50 | 3.71 | 57.26 | 99.70 | 100.61 | 98.62 | | 4 | 106.25 | 3.79 | 56.23 | 99.68 | 100.59 | 97.34 | | 5 | 108.00 | 3.92 | 55.19 | 99.65 | 100.51 | 95.94 | | 6 | 109.76 | 4.08 | 54.14 | 99.62 | 100.39 | 94.50 | Two ways to see the distribution Where the pressure sits. A national workforce and the value it produces are different quantities. Keep their units separate, and a difficult possibility becomes clear: economic growth and broadly shared gains need not arrive together. Moderate · model year 6 · fixed national workforce of 100 Back to the roof A different way to do the same work. Trace the original workload into human-led, assisted, and automated execution. Then add the human responsibility that comes with the new tools. Moderate · Roofer · model year 6. This is the original 40-hour workload, reallocated by execution. It is not a new 40-hour human workweek. Assisted work takes 5.5 actual human hours. Digitally or robotically executed hours measure the original human workload replaced; they do not measure machine runtime. - Added human support outside the original workload - +0.8 h - New human work outside the original workload - +1.4 h After time savings and added duties, this fixed workload calls for 35.4 actual human hours . This is labor required for fixed output, not a count of jobs lost. Employment also depends on demand. Read the task-hour accounting | Moderate / Roofer, model year 6. First four numeric columns are baseline hours; the last four are actual human hours. Display rounding can affect totals. | | | | | | | | | |---|---|---|---|---|---|---|---|---| | Task | Original | Human-led | Assisted baseline | Automated baseline | Assisted human | Support | New work | Total human | |---|---|---|---|---|---|---|---|---| | Rig safe access | 4.00 | 3.20 | 0.80 | 0.00 | 0.64 | 0.00 | 0.00 | 3.84 | | Move materials | 4.80 | 2.69 | 0.67 | 1.44 | 0.54 | 0.22 | 0.36 | 3.80 | | Remove old roofing | 5.60 | 4.03 | 1.01 | 0.56 | 0.81 | 0.08 | 0.14 | 5.06 | | Repair the deck | 4.80 | 3.76 | 0.94 | 0.10 | 0.75 | 0.01 | 0.02 | 4.55 | | Lay underlayment | 4.80 | 3.26 | 0.82 | 0.72 | 0.65 | 0.11 | 0.18 | 4.20 | | Install roof covering | 8.80 | 5.07 | 1.27 | 2.46 | 1.01 | 0.37 | 0.62 | 7.07 | | Fit the flashing | 4.80 | 3.80 | 0.95 | 0.05 | 0.76 | 0.01 | 0.01 | 4.58 | | Inspect & hand over | 2.40 | 1.86 | 0.47 | 0.07 | 0.37 | 0.01 | 0.02 | 2.26 | | Total | 40.00 | 27.68 | 6.92 | 5.40 | 5.54 | 0.81 | 1.35 | 35.38 | Pressure-test the argument The assumptions matter. Remove an overlap, change demand, or alter the speed of adjustment. The purpose is to see what the conclusion depends on. Moderate scenario · terminal comparison at model year 6, independent of the year control above. | Full scenario. All entries are model year 6. Differences use unrounded values. | | | | |---|---|---|---| | Measure | Full scenario | Selected case | Difference | |---|---|---|---| | Roofing wage index | 94.5 | 94.5 | 0.0 index points | | GDP index | 109.8 | 109.8 | 0.0 index points | | National unemployment | 4.1 | 4.1 | 0.0 percentage points | | Labor share of output | 54.1 | 54.1 | 0.0 percentage points | These are sensitivity tests, not alternative probabilities. “No direct robot substitution” retains augmentation. Changes to output demand can move national unemployment and roofing wages in different directions. The argument beneath the illustration A transition is something people have to live through. The important question is what happens between one arrangement of work and the next. A company can become more productive while an employee loses a reliable income. A customer can receive a better roof while a new worker finds fewer ways to learn the trade. Those outcomes can coexist. Understanding them requires looking past the job title to the tasks, the economics of completing them, and the people carrying the transition. Our concern is an overlap: digital systems may change office employment while physical systems change the work into which some displaced people might move. That overlap is possible, not inevitable. We do not have an established date for widespread autonomous roofing. We do have reason to examine what would happen if the period of protection offered by physical work proved shorter than the time people needed to retrain. This paper combines occupational evidence, bounded robotics reports, and an original conditional model. Its purpose is to make the assumptions visible enough to challenge. Optimism has a useful role here: it can motivate investment in a better outcome. It cannot substitute for the evidence that the outcome is being achieved. 01 · The unit of change Less time on a task does not settle the fate of a job. A production manager might delegate the first draft of a takeoff while retaining the inspection, the unusual substitution, and the conversation with the crew. A roofer might use a machine for a repeated installation task while continuing to diagnose damage and resolve awkward transitions. In either case, some work changes hands. The amount of paid employment depends on what happens to the rest of the job and how much work customers purchase. The opening uses task descriptions from O NET’s roofing occupation https://www.onetonline.org/link/summary/47-2181.00 and a composite of its construction-management https://www.onetonline.org/link/summary/11-9021.00 and construction-supervision https://www.onetonline.org/link/summary/47-1011.00 roles. We assigned the time weights. Each illustrated week begins at 40 hours; it is not an observed time diary. A commercial membrane crew would need a different allocation from the residential reroofing and repair mix used here. Colored portions preserve the original workload. They show baseline-equivalent work assigned to different execution modes, not machine operating hours. Assistance can reduce human effort on the work that remains. Supervision and new responsibilities then add human hours of their own. Their presence does not guarantee enough new work to replace every hour saved. Accountability for an accepted, durable roof also remains a separate question from which system executed a task. 02 · The receiving occupation The trades offer opportunity. Their capacity is finite. Anthropic’s original scenarios https://www.anthropic.com/institute/econ-scenarios explicitly exclude rapid robotics and recognize that workers may struggle to move between occupations. Our extension examines that stated boundary. A destination occupation may experience growing labor supply at the same time that technology changes its demand for human effort. Treating physical work as a permanent refuge would require an additional assumption about how long that work remains scarce. Roofing is a substantial trade, but a small receiving market relative to the national workforce. BLS reports https://www.bls.gov/ooh/construction-and-extraction/roofers.htm 166,900 roofer jobs in 2025 and projects approximately 12,000 openings annually over 2025–35, many replacing people who leave. Its May 2025 median annual wage is $55,440. Employment includes self-employed workers; that wage estimate excludes them. These figures establish scale, not a forecast of how many displaced office workers could enter. An occupational move involves learning, local vacancies, physical suitability, and a period of earning under different conditions. BLS describes training that progresses through practical tasks and experience. We should therefore ask a second question: if repeated beginner tasks become easier to automate, how will the next generation acquire expertise? A narrower entry route is a hypothesis to investigate. Experienced judgment retaining value would not, by itself, solve that problem. 03 · Capability, reliability, deployment The useful question is what the machine can finish. Physical automation in roofing need not arrive in a human-shaped body. In October 2024, Saint-Gobain announced a partnership with Renovate Robotics https://www.saint-gobain-northamerica.com/newsroom/pressrelease/saint-gobain-partners-renovate-robotics-innovate-robotic-roofing-technology targeting asphalt-shingle installation and prospective contractor pilots. Renovate’s public program https://renovaterobotics.com/ makes specialized roofing systems part of the relevant evidence. These sources do not establish whole-roof autonomy, independently verified reliability across varied sites, or a complete installed-square cost. Broader robotics reports explain why this boundary deserves attention. Figure reports https://www.figure.ai/news/production-at-bmw more than 90,000 sheet-metal parts loaded and 1,250 runtime hours in a BMW deployment. In a separate household experiment, Figure reports https://www.figure.ai/news/helix-2-5-zero-shot-30-home-generalization complete-task success increasing from 9% to 56% across 30 unseen homes with a different pretraining approach. These are vendor-reported results in specific workflows; neither measures roofing productivity. Likewise, a hand capable of gripping a triggered tool https://bostondynamics.com/blog/robot-hands-for-modern-ai-and-real-work/ is a component capability, not an accepted roof installation. The economic threshold is cost per accepted task. Include transport, setup, supervision, downtime, maintenance, financing, rework, and warranty exposure. Compare the same roof system, access conditions, and finish quality. A low purchase price or many powered-on hours cannot answer that question. Public evidence can justify a conditional scenario before it justifies an adoption forecast; keeping those standards separate makes the research more useful. 04 · The overlap Two changes can put pressure on the same paid hours. Consider a deliberately simple example with unchanged order volume. If qualified labor availability rises 10% while required human hours fall 20%, available labor per required position rises by 37.5%: 1.10 divided by 0.80. This arithmetic does not predict a 37.5% wage cut. It identifies a source of pressure that hiring, pay, working time, and occupational exit might absorb in different ways. Demand can offset the pressure. Better execution might make more projects affordable, shorten waiting times, or allow a company to undertake work it previously could not serve. Additional maintenance, repair, supervision, and training may also create paid work. Each channel needs a mechanism and a customer or institution willing to pay for it. A larger theoretical market is not the same as enough completed orders at sustainable margins. Our scenarios make both sides explicit. They vary task substitution, assistance, new responsibilities, qualified entrants, and the speed of matching people to positions. The roofing module then combines fewer hours per project with an assumed expansion in demand. Its wage response is a chosen rule, not an estimate from observed roofing markets. The exercise asks which assumptions produce pressure and what would have to change to relieve it. It cannot tell us which case is most likely. The timing matters within a company as well. A saved hour can become an additional project, a shorter working day, time to teach an apprentice, or a reduction in paid work. Those are different uses of the same technical improvement. The model represents selected demand and wage responses; it does not choose a contractor’s staffing policy. Employers and crews still have decisions to make before a productivity gain becomes an employment outcome. 05 · Reading the result A larger economy can contain a harder transition. In the Aggressive case, the year-6 national output proxy reaches 138.0 from a baseline of 100, while aggregate labor income falls to 85.8. The separate roofing wage index reaches 85.0. This is an internally consistent possibility under the selected assumptions, not evidence that these changes will occur. The proxy omits many constraints on production and investment that a national forecast would need. The counterexamples matter as much as the headline. Aggressive national-model unemployment is 12.2% with direct robot substitution and 12.6% without it. The modeled demand response and changing labor availability can offset displacement elsewhere. Robotics does not worsen every aggregate measure in this experiment. The comparison changes both required hours and demand, so it does not isolate a mechanical replacement effect. Stronger demand changes the roofing result as well. Raising roofing volume 25% above each default path lifts the terminal wage indices to 108.2, 103.8, 93.4 in Mild, Moderate, and Aggressive, respectively. That assumption is not a program we can promise to deliver. It makes the size of the offset visible. A constructive response should examine how useful demand could grow and where the economics would still leave people exposed. 06 · Who receives the gains Productivity creates a possibility. Distribution is a decision. Output, wages, and household security measure different things. A wage index describes the jobs in an occupational group; it does not track the lifetime earnings of the people who started there. Someone moving into another occupation may earn less even if wages in that destination rise. Employment and paid hours also matter. Looking only at the wages of people who retain work would miss part of the transition. Labor share is compensation divided by output. A worker can receive capital income without increasing that share. Ownership and profit participation may therefore matter to household outcomes, but their value depends on purchase costs, risk, dilution, governance, and actual distributions. Our model does not estimate those arrangements. Its gross nonwage residual includes capital costs and depreciation; it is not cash available to distribute. We think workers should have a meaningful part in the gains created with their experience and effort. That is a Foundation position https://foundationproject.net/?key=manifesto . Whether a particular arrangement delivers it has to be tested against individual outcomes. Founder ownership, a training announcement, or an expanding company is insufficient evidence on its own. The relevant questions are what workers receive, what risks they bear, and whether people affected by the transition can participate. 07 · A positive path Build the transition with the people doing the work. We support automation that makes good roofs easier and safer to deliver. A responsible approach begins with crews helping identify the work worth changing and the failure cases a demonstration can miss. Decisions about quality acceptance and stopping unsafe work need clear human authority. Removing exposure to difficult tasks is valuable; the result still needs evidence from the full operating workflow. Training and income continuity belong in the deployment plan before paid work disappears. New roles in diagnosis, maintenance, quality, customer care, and machine supervision need real demand, adequate pay, and a route for people to qualify. Employers should examine how newcomers will learn when machines take on some of the repeated tasks that previously supplied practice. These are design obligations, not guaranteed effects of buying equipment. Responsibility also extends beyond any one employer. Education providers, customers, insurers, and public institutions influence which changes can work in practice. Income support or public training could alter the transition, but those choices have costs and eligibility questions that this model does not calculate. Company-level success should not be presented as proof that the economy-wide adjustment has been solved. Measure customer and worker outcomes together: accepted work, price, rework, injury exposure, paid hours, earnings, retention, and participation in gains. Our concern would weaken if economical automation remained narrow, few qualified entrants arrived, or demand and new responsibilities absorbed the saved hours. It would strengthen if deployment outran those adjustments. A better future remains possible. Taking that possibility seriously means publishing adverse results and changing course when the evidence calls for it. The research record Assumptions you can inspect. This is an author-constructed stress test, not a calibrated forecast or a replication of Anthropic’s equilibrium model. No probabilities are assigned. Model years 0–6 describe a hypothetical sequence. Index 100 is a constant no-additional-automation counterfactual; baseline growth and inflation are excluded. What each quantity means - Task footprint - An illustrative 40-hour week. Colored shares divide baseline human-duration-weighted task instances. They are neither current adoption estimates nor machine runtime. Gold shows additional actual human hours, separately from the original footprint. - National output proxy and workforce - Two synthetic sectors and a labor force of 100. The five workforce stocks conserve that total. Training participants remain active jobseekers and are included in national-model unemployment. - Roofing wages and capacity gap - A separate receiving-market calculation. The capacity gap compares modeled positions with fixed participating supply; it is not an observed or forecast roofing unemployment rate. Wage indices are synthetic real indices, not projections of the BLS median. - Labor share and income - Labor compensation divided by modeled output, and aggregate compensation indexed to its baseline. The gross remainder is nonwage value added, not net profit or a guaranteed distribution. How the task and workforce calculations work Human hours per unchanged unit of output H = Σ wᵢ 1 − aᵢ − zᵢ + zᵢ/gᵢ + o aᵢ + r aᵢ Task weights w sum to one. Automation a, augmentation z, and unassisted work are disjoint; a + z cannot exceed one. Productivity g applies only to augmented instances. Support o and new work r add human hours around automation. Support is assumed to be 0.10 per automated baseline hour for digital work and 0.15 for physical work. The initial knowledge/nonknowledge split is 62.4/37.6, adapting the source’s employment share to a labor force. Each group starts with 3.8% unemployment, an assumed wage ratio of 1.5, and wages normalized to a 60% aggregate labor share. Initial output weights are proportional to that compensation. These are simplifying choices, not an official sector decomposition. Digital progress rises linearly over six periods. Direct robotics begins after period 2 and rises over four periods. The macro physical-exposure factor is 0.5 times the task-weighted roofing substitution share. A separate 0.5 augmentation-access assumption applies to remaining nonknowledge task instances; these factors do not establish that half of physical work resembles roofing. Desired jobs = baseline employment × target output × H Target output combines chosen digital demand uplifts of 1.6%, 8.3%, and 32.4% with a physical uplift equal to 0.20 times the direct nonknowledge robot share. We reuse the published GDP magnitudes as demand assumptions over a different horizon. Excess incumbents enter jobseeker pools; hiring is limited by vacancies and scenario matching fractions of 70%, 50%, or 35% annually. A one-period training queue has capacity of 0.5%, 1%, or 2% of initial knowledge labor force per year and assumes full completion. Realized sector output equals employment divided by the product of baseline employment and H. Fixed baseline weights combine those outputs into the national proxy. Capital is assumed available. There is no production network, investment constraint, relative-price equilibrium, or modeled public budget. Pay and the separate roofing market Target wages are the square root of jobs per qualified participant relative to baseline tightness. Wages move halfway toward that target in log terms each year. Trainees remain in the knowledge pool until completion. This is an assumed pressure response; wages do not feed back into hiring. Roof volume = national output index / 100 ^0.5 × H roof^−0.2 Roofing jobs equal volume times human hours per unit. Supply begins at 1/0.962 per initially employed person, then adds qualified entrants of 3%, 10%, or 20% of baseline employment by year 6. Its 3.8% initial jobseeker share is an assumption, not a measured roofing rate. Entrants are not an empirically identified flow from the national model. The demand exponents, cost pass-through, matching rates, and pay response are unestimated assumptions. The year-6 results, with the units kept separate | Foundation’s conditional experiment · model year 6 | | | | | |---|---|---|---|---| | Measure | Unit | Mild | Moderate | Aggressive | |---|---|---|---|---| | National output proxy | Index | 101.8 | 109.8 | 138.0 | | Knowledge-work wage | Index | 100.0 | 99.6 | 95.3 | | Composite nonknowledge wage | Index | 100.1 | 100.4 | 99.1 | | Separate roofing wage | Index | 98.5 | 94.5 | 85.0 | | National-model unemployment | % | 3.8 | 4.1 | 12.2 | | Labor share | % | 58.9 | 54.1 | 37.3 | | Aggregate labor income | Index | 99.9 | 99.0 | 85.8 | All indices begin at 100. Baseline labor share is 60%; baseline unemployment is 3.8%. Stored precision supports reproducibility, not forecast confidence. The download contains all 21 annual rows. Anthropic’s published benchmark is a different model The technical paper’s Table 3, printed page 31 https://www-cdn.anthropic.com/files/4zrzovbb/website/cf58f84d46a4a76bf5a5b039ac695fba6b80041c.pdf , reports the following January 2030 results. They are published model outputs, not observations. The scenario names and horizon differ from Foundation’s experiment. The nonknowledge wage measure is not a roofer wage. | Published benchmark · changes relative to its no-AI baseline where stated | | | | |---|---|---|---| | Measure | Modest | Substantial | Extreme | |---|---|---|---| | GDP change | +1.6% | +8.3% | +32.4% | | Knowledge wage change | +0.4% | −0.3% | −11.5% | | Other wage change | +1.1% | +5.9% | +33.6% | | Labor share | 59.4% | 56.1% | 45.2% | | Aggregate unemployment | 3.9% | 4.6% | 11.9% | Where the argument could fail The 33 sensitivity endpoints vary direct substitution, entrants, robotics timing, demand, wage response, and physical eligibility. “No direct robot substitution” retains assistance. These are year-6 comparisons; the provided sensitivity file does not contain annual trajectories. Comparisons can change multiple linked outcomes and are not statistically identified causal effects. Important omissions include capital financing and scarcity, robot service capacity, material costs, seasonal utilization, geography, training failure, retirement, immigration, labor-force exit, and demand feedback from lost earnings. The model does not estimate contractor margins, fiscal transfers, ownership returns, or the effect of a particular policy. Fewer human hours alone do not establish lower total roof costs. Better evidence would include paid task-time diaries across roof systems, accepted output and intervention logs from varied sites, complete operating costs, and longitudinal worker earnings. Those measurements could change the allocation, adoption, and demand assumptions substantially. The aggressive case tests broad deployment; it does not assert that present machines can execute those tasks. Inspect or reproduce the calculation Save the Python file in a writable folder and run python3 calculate scenarios.py . It uses the standard library, verifies the accounting identities, and writes the four data files beside the script. Running it replaces files with those names in that folder. Evidence and provenance Read the sources. Keep their boundaries. Official estimates and projections, vendor reports, scenario assumptions, and Foundation positions have different evidentiary roles. A vendor’s measurement is not independent replication. An assumption becomes a model input, not an observed fact. The record below identifies both what a source supports and what it cannot establish. 1. Published economic model Scenarios for our Economic Future https://www.anthropic.com/institute/econ-scenarios Anthropic. The task-based explanation and the scope of the original economic scenarios.Rapid physical robotics is explicitly outside its scope. Its scenarios are conditional. 2. Published economic model The technical paper accompanying the economic scenarios https://www-cdn.anthropic.com/files/4zrzovbb/website/cf58f84d46a4a76bf5a5b039ac695fba6b80041c.pdf Anthropic. The robotics boundary on printed page 5 and January 2030 results in Table 3, page 31.Its broad nonknowledge category is not roofing; Foundation does not replicate its model. 3. Official occupational evidence 47-2181.00 — Roofers https://www.onetonline.org/link/summary/47-2181.00 O NET OnLine. Roofing tasks, work activities, and occupational context.The descriptions do not establish our workweek weights or automation percentages. 4. Official occupational evidence 11-9021.00 — Construction Managers https://www.onetonline.org/link/summary/11-9021.00 O NET OnLine. Planning, cost, coordination, and construction-management responsibilities.This broad occupation is one input to our composite roofing production-manager role. 5. Official occupational evidence 47-1011.00 — First-Line Supervisors of Construction Trades and Extraction Workers https://www.onetonline.org/link/summary/47-1011.00 O NET OnLine. Crew supervision, scheduling, inspection, and site coordination.Combining task descriptions does not create an official roofing-manager population. 6. Official occupational evidence Occupational Outlook Handbook: Roofers https://www.bls.gov/ooh/construction-and-extraction/roofers.htm U.S. Bureau of Labor Statistics. 2025 employment and median pay; training context; projected 2025–35 openings.Employment includes self-employment; wage estimates exclude it. Openings include replacement hiring. 7. Partner announcement Saint-Gobain partners with Renovate Robotics https://www.saint-gobain-northamerica.com/newsroom/pressrelease/saint-gobain-partners-renovate-robotics-innovate-robotic-roofing-technology Saint-Gobain North America. The October 2024 partnership and planned contractor pilots for robotic shingle installation.The announcement does not establish completed pilots, whole-roof autonomy, or total installed cost. 8. Vendor-reported evidence Renovate Robotics https://renovaterobotics.com/ Renovate Robotics. The vendor’s specialized roofing-automation program and commercial preview.The reviewed site does not supply independent whole-job reliability or accepted-square economics. 9. Vendor-reported evidence Figure’s production deployment at BMW https://www.figure.ai/news/production-at-bmw Figure. Reported sheet-metal loading volumes and runtime in a bounded factory workflow.Factory loading is not roof installation. The report does not establish a public robot-hour price. 10. Vendor-reported evidence Helix 2.5: zero-shot generalization across 30 homes https://www.figure.ai/news/helix-2-5-zero-shot-30-home-generalization Figure. Reported complete-task success for three household behaviors in previously unseen homes.Generalization across tested homes does not establish roofing ability or independent replication. 11. Vendor-reported evidence Robot hands for modern AI and real work https://bostondynamics.com/blog/robot-hands-for-modern-ai-and-real-work/ Boston Dynamics. Component capabilities, including grasps for triggered tools.Holding a tool does not establish safe movement, accepted fastening, or a weatherproof roof. 12. Foundation position Foundation Projects manifesto https://foundationproject.net/?key=manifesto Foundation Projects. The values behind this paper: human dignity, education, responsibility, and shared progress.An organizational position, not scientific evidence or a demonstrated effect of an intervention. Authorship and disclosure Prepared for Foundation Projects. Research record reviewed October 3, 2026. Foundation works with roofing businesses and has a commercial interest in this transition. The article and original model have not undergone external academic peer review. Arithmetic checks establish consistency, not empirical calibration. Citing Anthropic or a vendor implies no affiliation or endorsement. Our stance is that safer, more productive roofing should also create a durable future for the people doing the work. The evidence and model above make that responsibility open to scrutiny; they do not promise a particular investment return or worker outcome.