
Job Displacement and Broader Societal Impacts
Concentration and distribution
AI market power has more than one layer
Market boundary · Separate input suppliers from model developers, application vendors and employers.
Map the layers first
Specialized chips, advanced memory, foundry capacity, electricity, cloud infrastructure, models and applications are inputs and products at different layers. A subscription bought by a clinic is not a unit of accelerator capacity, and the market share of a cloud provider is not the share of workers employed by companies on that cloud. Applying one concentration figure across the stack obscures who can exercise control, who can avoid it and what a remedy might change. [92] [93]
At each layer identify suppliers, paying customers, capacity constraints, how contracts allocate access and how quickly buyers can switch. Include bottlenecks in integration, such as identity, data access, application defaults and evaluation pipelines. Switching can be technically possible yet commercially unattractive.
Separate evidence from mechanism
Upstream concentration can coexist with strong downstream price competition. A handful of providers can supply several competing services; many front-end brands can also depend on the same vulnerable input. Neither count alone answers whether customers have effective alternatives. Diversification should be tested against the specific workload, compliance controls and service-level capacity needed, not the number of entries in a catalog.
Document the exposure in one chain: supply constraint or contract → procurement conditions → firm access or cost → task deployment → staffing and distribution. Each arrow is a hypothesis requiring evidence. Treat direct pricing, availability and switching data as evidence of procurement frictions, not of a specific loss of jobs. The remainder of this chapter follows the chain from access to distribution, without assuming all links move in the same direction.
Cloud partnerships combine capital and distribution
Not an adjudication · The FTC staff study analyzed arrangements and possible effects; it did not establish a violation.
Why the bundle matters
In January 2025 the FTC released a section 6(b) staff report examining partnerships involving [*REDACTED*] and [*REDACTED*] and [*REDACTED*] and [*REDACTED*] with [*REDACTED*]. The report discusses investment, revenue sharing, consultation and exclusivity provisions, cloud spending commitments and access to sensitive information; arrangements varied. Under some structures, a share of the investment can flow back to the provider through cloud purchasing. The headline investment is thus not a simple measure of cash available for unrelated costs or of a provider’s market share. [92]
Such relationships may finance capacity and reduce risk for both parties. They can also raise entry or switching barriers if competitors on either side cannot obtain comparable inputs or deploy on alternative infrastructure.
Contract terms are vintage-specific
The FTC snapshot largely reflects information available in 2024, with public updates through early 2025. It should not be recycled as a current list of exclusivity terms. In April 2026 [*REDACTED*] announced that [*REDACTED*] remained its primary cloud partner but that [*REDACTED*] could serve products across cloud providers and [*REDACTED*]’s intellectual property license became non-exclusive; the disclosed capped [*REDACTED*]-to-[*REDACTED*] revenue share was described as continuing through 2030. In October, [*REDACTED*] announced planned access through [*REDACTED*]’s [*REDACTED*], explicitly reiterating the primary-cloud relationship with [*REDACTED*]. These company announcements are not independently verified usage, exit capacity or a complete contract. [95] [96]
For procurement, separate first-release priority, hosting rights, spend commitments and actual workload distribution. A new route to buy the same model can improve choice without diversifying the underlying model provider.
Switching cost can preserve incumbent advantage
UK evidence · A UK public-cloud infrastructure finding is not a global AI model-market share.
The UK cloud investigation
The Competition and Markets Authority’s UK public-cloud infrastructure investigation concluded in July 2025 with adverse competition findings; it recommended that the CMA consider strategic market status investigations involving the largest providers, [*REDACTED*] and [*REDACTED*]. This concerns a defined UK supply market rather than all computing, all software, or all AI. The evidence on switching includes commercial terms, technical barriers and software licensing; consumer benefits and switching costs may coexist. [93]
In October 2026 the CMA made a provisional strategic market status designation proposal about [*REDACTED*]’s business software ecosystem, not a final decision or a finding of wrongdoing. Its separate cloud engagement covered both [*REDACTED*] and [*REDACTED*], including identity and access management and workload migration. [94]
Test whether customers can leave
Measure what a switch would require: export data and access policy; rebuild or port integrations; rerun benchmark and red-team suites; recertify controls; replace model-specific tuning; and operate an overlap period. Count lost volume discounts and transfer charges, including any restrictions associated with commitments. Switching complexity depends on workload; a disposable text drafting tool differs from a regulated workflow with embedded case routing and audit requirements.
A proof of portability should reroute one representative workflow and meet its quality, latency and audit constraints using independent resources. Identify shared failure points, rather than equating several contracts with redundancy. Procurement can reward genuine substitutability while still allowing legitimate bundled discounts. A policy evaluation should then look for changed prices, provider entry or switching time and adverse effects, not just a public commitment that more choice exists.
Large firms report more AI use
U.S. Census BTOS · May 3, 2026 period endpoint; reported use in any business function during the past two weeks.
Distinguish two denominators
Census’s BTOS analysis found that national firm use was 19.8% in the period ending May 3, 2026. In that same analysis, about 37% of firms with 250 or more employees reported use and 32% of firms with 100–249 employees did so; the corresponding percentage for firms of four employees or fewer was below 20%. The question covered any business function, not only production of goods and services. Before November 2025 the wording was narrower, making a direct long-run splice inappropriate. [97] [99]
A separate revised analysis of the November 2025–January 2026 supplement found 18% firm-weighted use and 32% employment-weighted use. Weighting firms by the number of people they employ puts more weight on larger firms. The latter measures employment at firms reporting use, not the percentage of people personally using AI or an adoption jump over time. [98]
Why the difference matters
Larger organizations may be able to amortize information security, legal review, integration and training across more workers. Smaller organizations may move faster on a narrow workflow using commercial software but lack staff to verify results or renegotiate a problematic contract. These are plausible mechanisms; size comparisons are descriptive and reflect industry mix and organization design too.
Compare adoption, intensity, duration and outcomes. A small company using AI daily across all billable work is not less reliant than a large company testing a drafting tool in one unit simply because the latter employs more people. Report who has access, where deployment is authorized and how much workflow volume is touched rather than treating organizational adoption as a sufficient statistic.
Exhibit 13 · Business adoption varied by sector and firm size
Census BTOS reports the share of businesses reporting use in the prior two weeks. Starred groups are firm-size categories across industries, not sectors. The categories overlap. Values are not the share of workers using AI, a productivity measure, or a measure of market power.
U.S. Census Bureau, Business Trends and Outlook Survey: primary source
Sector averages hide how workflows change
U.S. Census BTOS · Percentages refer to surveyed businesses; individual workers and tasks have different denominators.
Look inside the industry
For the period ending May 3, 2026, Census reported 39.7% AI use in Information and 33.9% in Finance and Insurance, both above the 19.8% national rate. The comparisons do not show the probability of AI-caused displacement in either industry. Each includes different occupations, and firms adopting AI may have used it for unrelated work, including marketing or internal document processing. [97]
In the separate winter supplement, 57% of adopting firms used AI in three or fewer business functions. The leading reported functions included sales and marketing (52%), strategy and business development (45%) and IT (41%). Respondents could select more than one function, so these shares are not slices of an exclusive pie. [98]
Combinations, not averages, drive impact
An insurer may use office AI to draft correspondence yet require a qualified person to decide claims. A manufacturer may use AI in logistics without changing the work of every machinist. A customer-service team may automate routine tickets but raise staffing needs for escalations. These are illustrative deployment pathways, not evidence of observed changes in those sectors.
To compare workplaces, retain the job family, baseline workload, review requirements, unit of output and actual deployment date. Distinguish depth from reach: a system used by hundreds of people for occasional drafts differs from one that does most first-pass application review. Follow both managerial and frontline work. Adoption surveys identify sectors for targeted follow-up, not a ranking of careers to abandon.
The value from time savings has many claimants
Distribution lens · Measure verified incremental gains, then trace who receives them.
Gross capacity is not distributable gain
An AI tool can shorten drafting but increase review, corrections, security oversight or customer disputes. Measure output that passes the applicable quality standard and count all time spent to produce it. Deduct recurring technology and operating costs, remediation and transition expense before inferring a financial surplus. Changes in case mix can make output look faster without improving the work; preserve the distribution of case difficulty and compare like-for-like. [107]
The downstream value can accrue to workers as wages, bonus, paid training or reduced work time; to customers as lower prices or faster service; to the firm as margin or reinvestment; or upstream as license and cloud payments. Added revenue does not by itself identify any party’s share of net surplus.
Track pathways explicitly
A fixed-fee professional service can retain a time saving as margin while an hourly billing model might see revenue fall at unchanged prices and demand. A budget-limited public agency might serve more cases without cutting headcount. A contractor paid per task could lose assignments even as directly employed staff gain time. These are illustrative allocations under different contracts.
Use a ledger with separately denominated rows for quality, output, labor hours, labor compensation, vendor costs, prices and working conditions. Avoid adding customer welfare, wages and revenue into a single monetary gain unless a formal accounting framework prevents double counting. Compare both the near-term pilot and steady-state operations. A product can improve average throughput yet transfer pressure, costs or risk to the least powerful part of the workflow.
Exhibit 14 · Who captures the productivity gains?
Higher output or lower effort does not establish who benefits. Competitive conditions, contracts, worker voice, and business strategy influence how gains reach workers, customers, firms, or suppliers. The diagram assigns no shares.
48PEGASUS Research conceptual exhibit. Not empirical data.
Labor share measures compensation, not individual pay
BLS nonfarm business · November 5, 2026 release: Q3 labor share 53.0%, preliminary and subject to revision.
What the national ratio says
The labor share is the portion of covered economic output that accrues as compensation to labor, rather than the median worker’s wage or the fraction of workers who gained a raise. The Bureau of Labor Statistics reported a 53.0% third-quarter 2026 share in its November preliminary release, describing it as the lowest in its revised series beginning in 1947. That release incorporated national-accounts revisions; the earlier September Q2 publication showed 52.8% at the time. Do not combine old and revised figures into a single unannotated line or infer a contradictory ranking. [100] [101]
In the November release, third-quarter productivity grew at an annualized 3.6% quarter-to-quarter while real hourly compensation grew at an annualized 1.1%. Neither rate is an AI-specific estimate. Different price measures and compensation concepts matter when comparing productivity and pay. [100]
What a distribution analysis needs
Sector composition, prices, labor costs, capital deepening and shifts in the measurement of output can change the aggregate ratio. To investigate a technology contribution, add firm and sector adoption dates and control for plausible contemporaneous shifts in demand, investment and labor composition. National productivity and labor-share series alone do not isolate those effects.
Track levels, growth and the distribution separately. A worker can receive higher nominal pay while a firm’s output grows faster and its labor share falls. Average compensation can rise even if a typical worker does not gain. Use household income, matched wage records or compensation distributions for household and worker claims, and use the nonfarm business accounts for claims about the covered aggregate.
Asset ownership is a separate distribution channel
Federal Reserve DFA · 2026 Q2 selected asset category: aggregate holdings by wealth group, not AI-only ownership.
Capital can transmit gains unequally
As of 2026 Q2, the Federal Reserve’s Distributional Financial Accounts reported corporate-equity and mutual-fund holdings of $16.15 trillion for the top 0.1% of households by wealth, $16.74 trillion for the next 0.9%, $24.01 trillion for the 90th–99th percentiles, $7.36 trillion for the 50th–90th percentiles and $0.37 trillion for the bottom half. The groups are not equal in population and other asset categories, including pension entitlements, are shown separately. The table is not a measure of AI-company ownership, total wealth or realized investment gains. [102]
This baseline illustrates why a distributional analysis can change when adding capital income to wages. Asset appreciation can accrue to existing ownership even when the people who lost work have little exposure to the same assets.
Do not turn a channel into a forecast
An IMF working paper models scenarios in which automation can compress wages at the top in some settings while increased capital returns widen wealth inequality. The result depends on assumptions about who owns capital, which tasks can be substituted, and the model’s counterfactual; it is not an observed effect of today’s systems. [104]
Report asset exposure and labor outcomes as separate series. Do not infer which shareholders benefit from individual deployments without firm cash flow, ownership and valuation evidence. Positive gains accruing through diversified retirement accounts can coexist with uneven direct equity wealth, missed employment opportunities and shifts in business ownership. Policies intended to share benefits should specify the mechanism, the eligible group and when returns would actually become available.
Worker bargaining determines how savings are allocated
U.S. wage and salary workers · 2025 union membership 10.0%; public/private and occupational differences matter.
Bargaining power is more than union density
BLS estimated a 10.0% union membership rate for U.S. wage and salary workers in 2025, with 14.7 million members. The 2025 CPS annual average excludes October because the survey was not conducted that month; comparisons with full-year averages require caution. Membership is one institutional measure, not a comprehensive index of bargaining power and not a prediction of an AI-related wage effect. [105]
Workers’ leverage also depends on meaningful outside options, local hiring competition, the portability of credentials, unemployment protection, information, and formal or informal channels for input. The same technology can distribute gains differently when deployment rules are negotiable or when pay and scheduling decisions remain opaque.
Make consultation consequential
Before rollout, identify which decisions can affect pay, schedules, performance ratings and discipline. Where appropriate, document notice, input from workers and representatives, a named decision owner, appeals and a process for correcting source data. Invite feedback from contingent workers and junior staff who may not be represented by aggregate employee surveys. [107]
Compare three governance choices: identical wage and work expectations, revised productivity targets alone, and shared gains through compensation, paid training, time or redesign of work. For each, test changes in job quality, output and turnover. Avoid reading the descriptive wage difference between union and nonunion workers as the causal effect of membership; occupations, employer types and industries vary. A mechanism is only meaningful if workers can use it without fear of reprisal and the organization acts on substantiated feedback.
Algorithmic management can change work without cutting jobs
Worker impact · Scheduling, monitoring and evaluation deserve separate controls.
Distinguish assistance from authority
The OECD’s 2025 employer study covered more than 6,000 firms across six countries. Managers reported potential benefits in decision processes as well as concerns about accountability, explainability and worker health. The study is employer-reported and covers algorithmic management tools, some of which need not use generative AI. It cannot establish that a specific form of monitoring improves welfare or causes job losses. [106]
A schedule suggestion with a human authorized to override it has a different impact from a system that reduces shifts or pay without meaningful review. Retaining headcount does not rule out greater work intensity, unreliable hours, slower advancement or less discretion. Conversely, automation of repetitive clerical coordination can make shifts easier to plan.
Control consequential changes
Measure disputes, erroneous evaluations, involuntary schedule changes and successful appeals in comparable units with denominators and time windows. Preserve the workflow rule and system version behind decisions; separate audit logging from expanded surveillance. Avoid metrics that reward supervisors for fewer appeals if workers may instead be deterred from reporting issues. [107]
Define a decision hierarchy: routine assistance, recommendations subject to documented review, and decisions that cannot be delegated to a system. Test whether reviewers can access and understand source evidence before accepting a recommendation. Include workers in post-launch review, particularly the groups most likely to be affected by sparse histories, atypical schedules or data quality issues. Report not only whether the tool works but which people gain discretion and which lose it.
Employer concentration is not platform concentration
Labor-market competition · Define the real set of employers competing for a worker’s skills and availability.
Look at practical alternatives
Employer market power concerns the ability to set worse pay, benefits or working conditions without losing enough workers to make the change unprofitable. Relevance depends on occupation and skills, commuting or remote options, scheduling constraints, licensing and hiring practices. A cloud market share cannot establish concentration among local buyers of a worker’s labor. An industry’s local employment share also does not tell us how many employers compete for its workers.
The U.S. merger agencies’ Guideline 10 describes how they examine mergers between competing buyers, including employers, and what reduced competition could mean for wages and conditions. These are agency analytical guidelines; applying them to a particular transaction requires facts and qualified assessment. [108]
Follow hiring behavior and terms
Examine where workers actually go when they leave, which firms recruit from the same pool, whether applications convert into credible offers, and the costs and duration of a transition. Look at pay bands, benefits, contract type and hours, rather than wages alone. The employer of record may differ from the firm controlling work if agencies or platform intermediaries are involved.
A model-assisted hiring platform can reduce search frictions if it helps compare a broader set of jobs, or reinforce dependencies when it steers candidates into narrow channels. Verify both mechanisms with outcomes and testing. Distinguish monopoly power among upstream vendors from employer buyer power in local labor markets so that competition and workforce remedies are aimed at the correct decision makers.
Places with similar exposure can adjust differently
Regional dashboard · Pair industry composition with hiring alternatives, wages and practical access.
Build an industry baseline
BLS Quarterly Census of Employment and Wages data provide establishment counts, employment and wages by industry and geography for covered workers. A location quotient compares an industry’s share of local employment with its share in a reference area. A share above the reference area flags specialization, not an AI adoption rate, the probability of displacement, or the concentration of individual employers. More detailed data may be suppressed to protect confidentiality. [109]
Start with the commuting labor market and nearby job alternatives, not a state average. Pair industry exposure with occupation mix, actual implementation evidence, and typical routes into alternative work. Compare working-age access to connectivity, training schedules and transport so program seats are not mistaken for useful options.
Distinguish construction from durable capacity
A proposed data center may create construction work and a smaller continuing operations workforce, but neither investment announcements nor electrical capacity indicate net employment in a community. Require separate local counts for contractors, permanent positions, project phase and place of residence before describing local gains. Compare public incentives, resource needs and alternative uses under transparent assumptions.
Regional development can broaden capabilities through employers, universities and workforce providers. EDA announced in July 2026 intent to award $169 million across six Tech Hubs; those are selection or implementation actions, not a demonstrated employment impact and not exclusively AI funding. Track disbursement, participation, durable jobs and earnings over time. A region that can host a building is not automatically positioned to retain high-value functions or workers. [110]
Public compute can widen access, not guarantee parity
Resource policy · Measure usable capacity, queue delays and institutions able to complete projects.
Match the shared resource to the barrier
National Science Foundation’s National Artificial Intelligence Research Resource connects research and education users with resources such as computing, datasets, software, expertise and models. Shared access can relax fixed costs for institutions that cannot buy dedicated systems. A grant allocation, however, is not the same as completed research, a competitive commercial platform, a job placement or proof of equal access. [111]
Open weights can widen the ability to run or adapt models without a hosted API, but execution, storage, governance and technical support still cost money. License terms and portability differ. An organization may also depend on one supplier for critical integrations even when it can download a model separately.
Design for actual use
Set goals for participation by resource-constrained institutions, completion, and usable results. Publish units and scope for capacity: training, inference, reserved allocations and actual runtime are not interchangeable. Count waiting time and projects that could not proceed, not only grants awarded. Combine allocations with onboarding support, data stewardship and verification to avoid excluding organizations without existing technical teams.
Evaluate additionality: which projects would otherwise have stalled, run at lower scale, or used an alternative? A recipient survey alone cannot fully answer it; compare similar unfunded applicants where fair and feasible and guard against collecting sensitive records unnecessarily. Seek enough provider diversity to avoid single-supplier dependence, while preserving clear controls for safety, data access, reproducibility and budget.
Competition remedies and transition support solve different problems
Policy design · Identify the mechanism before choosing the remedy.
Two complementary tracks
Where a bottleneck is weak cloud or software competition, access and switching remedies could include workable data export, interoperability, transparency on licensing, and proportional scrutiny of exclusionary conduct. Public procurement can require tested migration and access to audit logs while permitting lawful efficiency benefits. The UK CMA’s ongoing cloud engagement and provisional business software proceedings illustrate different policy channels rather than a one-size-fits-all response. [94]
Where the barrier is a shrinking set of credible local employers, training alone does not create an opening. Conversely, a more competitive infrastructure market does not itself give a displaced worker income replacement, career advice, transport or time to retrain. Labor competition needs its own evidence and response. [108]
Connect workers before the separation
The federal WIOA Adult and Dislocated Worker programs support eligible employment and training services, with Rapid Response working with employers and worker representatives to limit disruption as appropriate. These are service channels subject to eligibility and implementation constraints, not evidence that every affected person will receive aid or that AI-specific displacement will be covered automatically. [112] [113]
A practical portfolio pairs early notice and placement with occupation-specific training, earnings and retention measures, and accessible supports. Measure time to reemployment, whether earnings and benefits recover, and how results differ for workers who do not complete training. Name which organization owns each outcome and what funding would sustain it. Policy should address short-run costs without promising that assistance reverses every loss.
An executive scorecard should track outcomes by group
Suggested metrics · Define time, population, baseline and change before rollout.
Measure reach and concentration separately
At the vendor layer, report shares of spending and workload by provider after separating model, cloud and application. Record dependency on the same underlying host and the practical tested time to switch. At the business layer, report AI use by function and workflow volume, not the count of purchased seats. At the worker layer, measure task mix, hires, separations, hours, pay and reassignments. At the community layer, compare employer diversity and post-transition earnings. [98] [109]
For each measure record the denominator and source. A spend share uses invoices; workflow share uses reviewed units of work; employment outcomes need a stable cohort and explicit exclusions. Observe differing timelines instead of forcing all metrics into a month-end percentage.
Decisions need triggers and counterfactuals
Set governance triggers before deployment: an inability to switch the most critical workflow, quality erosion after escalation, transfers into lower-paid roles, or costs shifted to contractors. Predefine meaningful outcomes and fair comparisons; track changes against a credible baseline before attributing them to AI. A phased rollout can help separate technology effects from demand shifts if the chosen groups are comparable and policies are applied fairly.
Revisit thresholds under new circumstances but keep a log so the pilot is not declared successful retroactively. Report subgroup results only where sample size and confidentiality permit. Distinguish organization-specific evidence from national trends or provider claims. Assign owners for implementation, worker response, post-launch monitoring and the independent decision to pause or narrow the rollout.
Benefits can diffuse while returns concentrate
Conclusion · Access, work outcomes and returns to ownership must remain separate in public debate.
What can coexist
AI tools can become cheaper or more widely available while a small set of firms controls critical infrastructure. Productivity can improve without broadly higher wages. Local hiring can weaken even if national employment grows. Higher exposure in an occupation can produce either assistance or substitution, and scenario studies depend on assumptions about complementarity, ownership and demand. None of these potential combinations follows automatically from a survey response. [92] [103] [104]
The democratic question is therefore not merely whether the economy benefits in aggregate. It is which groups receive useful services, better work or investment returns, which groups bear transition costs, and which decision makers can alter those distributions.
Priority actions
First, record the specific dependency and test real alternatives before treating a multivendor catalog as open competition. Second, audit job transitions and compensation, not merely payroll snapshots or seat usage. Third, make distribution explicit in budgets and contracts: training time, worker notice, appeals, pricing and reinvestment are design choices. Fourth, support places with little institutional capacity, measuring actual participation and earnings rather than only announcements. [94] [112] [110]
Use two simultaneous tests for public claims. For a competition claim, identify the product and geographic market, time period and mechanism. For a distribution claim, identify the people and assets, numerator and denominator, and baseline. If those boundaries cannot be named, show the uncertainty rather than a headline that presents potential exposure as observed displacement or concentrated wealth as a measured gain from AI.
Browse: White Papers · Reports · Downloads
