AI and wages: the power to walk away

A worker can keep a job while losing one reason to expect a pay rise: the possibility of finding a better offer elsewhere. An employer can still need a person to do the work while finding that person easier to replace. Wages are partly shaped by these alternatives. Artificial intelligence could change them well before a redundancy programme makes the shift visible.

Speaking in Dublin on 13 October 2026, Philip Lane brought this issue into the European Central Bank’s assessment of AI. He described aggregate employment trends as largely unchanged, alongside subdued hiring plans that particularly affect new graduates. The prospect of substitution, he suggested, could weaken wage bargaining and reduce labour’s share of income. His speech identifies a possible mechanism; it does not estimate a causal effect on European pay. ^[1]

Testing the argument means looking inside the employment relationship: which knowledge is difficult to transfer, how readily firms can hire a replacement, and where workers can go next. The evidence is mixed. Productivity gains have been documented, some entry routes are narrowing, and several European studies find stable or higher earnings among adopters. The distributional outcome remains unsettled. ^{[4] [5] [7] [9]}

The value of another job offer

Consider an illustrative team handling complicated customer cases. An experienced employee knows how to spot inconsistencies, find the relevant rules and interpret a poorly expressed request. Replacing her would mean training time and mistakes. The firm has a reason to retain her that goes beyond the number of cases she completes each week.

A tool might make part of that knowledge available to the rest of the team. The service becomes faster and less dependent on one person. For the experienced employee, two things happen together: her output increases, while some of the obstacles to replacing her diminish. Her future pay will depend in part on the skills that remain scarce and on the offers she can obtain elsewhere.

Economists call the alternative available when a relationship ends, or an agreement fails, an outside option. A worker’s option might be another employer. The firm’s might be a different hire, a redesigned workflow or automation. To matter, an alternative must be workable. A persuasive demonstration leaves questions about reliability, integration and the skills needed to operate the service unanswered.

Henrique Basso and Omar Rachedi examine this logic for industrial robots. Their research combines US metropolitan data from 2008–2018 with a bargaining model in which the ability to automate can change the response of wages before the contemplated adjustment is made. It offers a mechanism to investigate, rather than a coefficient that can be carried over to generative AI in 2026. ^[2]

Method and scope

Conditional analytical diagram with no monetary or probability scale. Automation and standardisation are distinguished in the text. The possible effects depend on effective adoption, demand and labour-market institutions.

The worker’s side matters just as much. When several employers need the same skill, an employee may turn greater efficiency into a better-paid move. When suitable openings become scarce, excellent performance in the current job can coexist with weaker bargaining power. The surrounding labour market can therefore transmit the change to people who make little use of AI themselves.

The ability to replace and the ability to leave have to be considered together. Counting technically automatable tasks alone misses customer demand, workplace organisation and the practical alternatives available to both parties.

Knowledge becomes part of the workflow

A September 2026 working paper by Masao Fukui, Emi Nakamura and Jón Steinsson formalises a related possibility: standardisation can increase output while making workers more interchangeable. People still perform the tasks in their model. What changes is how much the outcome depends on their individual knowledge. This is a theory of income distribution, rather than an estimate of the current effect of chatbots. ^[3]

The design of work is crucial. Return to our hypothetical team. Suppose the assistant proposes a procedure and the employee learns to question it, correct it and understand its limitations. That can build a skill useful at another firm. Alternatively, the procedure might become a series of instructions whose underlying logic the employee never learns. Performance could improve while remaining heavily dependent on the employer’s system. Both arrangements raise the case count; the portable expertise acquired may differ considerably.

Evidence from an actual workplace shows how sharing knowledge can help newcomers. A study published in the Quarterly Journal of Economics in 2025 follows 5,172 customer-support agents during a staggered rollout of generative assistance. The sample includes workers who received the tool and workers who did not. On average, issues resolved per hour rise by 15%, with particularly strong benefits for less experienced staff. That is a workplace output measure, not a pay increase. ^[4]

Our earlier review of AI and productivity discussed earlier versions of several studies. This analysis uses the published or revised versions identified in the source notes.

The same technical improvement can open a role to more applicants and change the scarcity of a skill. The next step depends on demand, the skills still needed and competition between employers. A firm winning new customers may recruit more people even as each employee handles more cases.

There is an important counterweight within Fukui and his co-authors’ model itself. When firms are identical and equally standardised, easier replacement on the employer’s side is offset by better job-finding options for workers. Labour’s share is unchanged. Differences between firms are central to obtaining the alternative distributional outcome. ^[3]

That condition makes it unsafe to infer a general decline in wages simply because software becomes easier to use. The useful question is where accessible jobs are being created and whether new skills genuinely allow people to move between employers.

The first adjustment may be at the entrance

For someone starting a career, an outside option first requires an opening. A business can slow recruitment while retaining its existing team. Employees already inside continue working; applicants wait longer for a first opportunity. Headcount can initially suggest continuity even as the prospects of incoming cohorts change.

In the August 2026 revision of their study of US ADP payroll data, Stanford researchers document such a divergence among 22–25-year-olds. Between November 2022 and June 2026, employment falls by about 11% in the two most AI-exposed occupational groups and rises by about 10% in the three least-exposed groups. The adjustment is concentrated primarily in hiring. ^[5]

Method and scope

Common baseline of 100 in November 2022. Higher exposure: top two occupational quintiles, approximately −11%. Lower exposure: bottom three quintiles, approximately +10%. Calculation from these rounded values: 89 / 110 − 1 = −19.09%. Only start and end points are shown; no monthly path is invented. Revision announced on 12 August 2026; pre-trends and limitations discussed on pages 3–4.

The approximately 19% figure highlighted in that revision compares paths. It is the shortfall relative to where employment would stand had the more-exposed group kept pace with its peers. The authors flag pre-existing trends, sensitivity to controls and a stronger divergence in ADP than in national survey benchmarks. They explicitly distinguish the descriptive finding from an isolated causal effect of AI. ^[5]

France provides another observation at the entry point. An Insee analysis published on 24 March 2026 reports a 7.4% year-on-year fall in employment among 15–29-year-olds, excluding work-study trainees, in computer and information-service activities in the fourth quarter of 2025. Its age group and period differ from the US study. The descriptive result cannot attribute the entire decline to AI. ^[6]

These findings justify tracking flows alongside announcements of job cuts. To assess bargaining power, useful evidence would include successful applications, the time needed to find a comparable job and the pay offered to new hires. Fewer openings could reduce applicants’ negotiating room before altering the wages of established teams.

That transmission still needs testing. Recruitment may also reflect the business cycle, earlier over-hiring or changing skill requirements. Linking actual adoption to hiring decisions within individual firms would help separate these explanations. An occupation’s theoretical exposure tells us less about what a particular employer has changed than a record of its redesigned workflow.

A slower risk concerns learning. When straightforward tasks disappear from an entry role, firms need other ways for newcomers to acquire the judgement required for difficult tasks. An experienced team may temporarily operate with fewer juniors. Replenishing that experience becomes a later problem. This is a scenario about the organisation of training, whose eventual scale the available numbers do not establish.

European results challenge a uniform story

The Danish work of Anders Humlum and Emilie Vestergaard is an important counterweight. It links surveys of roughly 25,000 workers in eleven exposed occupations to administrative records through the end of 2024. The authors find no clear differential average effect on earnings or hours. Their confidence intervals rule out average earnings effects exceeding 2% in magnitude. That bound concerns their estimated average, rather than every individual worker’s possible experience. ^[7]

A separate paper published by the BIS in January 2026 uses firm observations from 2019–2024. Its authors estimate a productivity gain and find higher wages among adopters, without a detected short-run decline in employment. Their identification strategy relies on matching European firms with comparable US firms; its causal interpretation depends on the underlying assumptions. AI here covers a broader range of technologies than generative assistants alone. ^[9]

Expectations collected by the ECB are also more favourable among intensive users. In the late-2025 Survey on the Access to Finance of Enterprises (SAFE), these firms anticipate stronger wage and employment growth over the following twelve months than non-users. The chart reproduces the published unconditional weighted means. Those averages are different measures from the adjusted regression estimates. ^[8]

Method and scope

Source: footnote 4 of the SAFE box, ECB Economic Bulletin 2/2026. Wages: 3.1% with no use and 3.7% with significant use. Employment: 0.6% and 2.8%. Raw gaps are not mixed with adjusted coefficients of +0.6 percentage point for wages and +1.3 points for employment. AI includes predictive and generative tools and process automation. No confidence intervals for the means are reconstructed.

A business with expanding orders may both adopt AI and plan to hire because it needs more capacity. A simple comparison between users and non-users risks assigning to the tool a difference rooted in the firms’ prospects. Conversely, sorting businesses by the technical exposure of their occupations is an imperfect measure of actual adoption.

The purpose of adoption matters too. In its analysis published on 4 March, the ECB associates AI use for research and innovation with job creation, while firms aiming to reduce personnel costs show a different pattern. These are survey comparisons with controls, over a limited horizon. They do not establish the same outcome for every type of deployment. ^[14]

The studies cannot be pooled into a single meaningful average. They concern different years, technologies and populations. Together, however, they impose a useful discipline: the wage-weakening hypothesis must contend with evidence of business expansion and with average effects still close to zero.

A null result through late 2024 does not settle every transformation taking place in 2026. It does require a further explanation of what has changed: intensity of use, reliability, integration or demand. Simply declaring that the adverse effect has been delayed would otherwise protect the hypothesis from any contrary evidence.

Pay can rise while labour’s share falls

Productivity describes output per unit of labour. Value added is the value of production less the goods and services consumed in producing it. In our company example, the employee compensation share divides total employee compensation by that value added. Pay and productivity can both rise while that share falls. ^[15]

Take an entirely hypothetical example, with unchanged prices, headcount and working hours. Value added increases from 100 to 110 units. Total employee compensation, including employers’ social contributions, rises from 60 to 61.2. ^[16] Compensation has increased by 2%, against a 10% increase in value added. Labour’s share drops from 60% to approximately 55.6%.

Method and scope

Inputs chosen to explain arithmetic, not empirical estimates or forecasts. Before: value added 100, labour 60, remainder 40. After: 110, 61.2 and 48.8. Final share: 61.2 / 110 × 100 = 55.636%. Bars share a zero-based scale. Compensation includes employers’ social contributions; other value-added components are not net profit. Source 3 motivates the distributional question but supplies none of the hypothetical inputs.

The example predicts none of these rates. It shows why higher pay can coexist with a redistribution of the income generated by production. The remaining value added includes other components, such as gross operating surplus and, depending on the valuation convention, taxes less subsidies. Calling all of it net profit would omit depreciation and several further accounting steps. ^[17]

Average pay can also become misleading when the workforce changes. In a second hypothetical example, two employees paid 30 and 50 produce an average of 40. If only the job paid 50 remains in the calculation, the average rises to 50 without anyone receiving a raise. Slower hiring into junior roles could produce this kind of compositional change. Tracking distribution therefore requires comparable workers and jobs, hours and labour-market entry, as well as the headline average.

The issue becomes macroeconomic when aggregate income is distributed differently across households. The effect on consumption depends on how much recipients spend, their savings and their access to credit. Capital owners may also be employees, directly or through pension funds. The accounting distinction between labour and other income cannot by itself identify which households gain or lose.

Our reading of US productivity and real-compensation measures also distinguishes their scope and observation period. It does not estimate an AI effect.

The calendar of wage agreements

Previously negotiated pay rises create inertia. A firm may reorganise production this year while an existing agreement still governs increases over the coming months. Newly signed agreements and offers to recruits may therefore reveal a change in bargaining conditions sooner than the wages currently being paid.

The ECB wage tracker, published on 16 September, incorporates agreements signed through the end of August. Known clauses imply year-on-year growth of 2.7% in the first quarter of 2027 and 2.8% in the second. Statistical coverage, however, declines from 32.5% to 25.1% of employees across the nine participating countries. These forward-looking signals are conditional on the contracts already in the database. ^[11]

Method and scope

ECB release of 16 September 2026, Table 1. Agreements known by end-August. Q1 2027: annual growth 2.7%, coverage 32.5%. Q2: 2.8% and 25.1%. Coverage bars are proportional on a 100 base; the calendar is schematic, not time-proportional. Denominator: employees in the nine participating countries. Information is conditional and revisable; forward-looking clauses are not a forecast.

New agreements can enlarge or change that information. The grey portion of the chart represents employees outside the tracker’s available data, rather than employees without collective bargaining coverage. The release also identifies base effects from earlier one-off payments in the 2026 measure. An apparent slowdown has to be examined against those contractual calendars before it is assigned a technological explanation. ^[11]

For inflation, another step lies between productivity and selling prices. Return to the hypothetical business: producing more while total compensation grows more slowly reduces labour compensation per unit of output. The firm might lower prices, preserve margins, improve quality or spend more on other inputs. Competition and demand help determine the outcome. The wage-productivity gap alone cannot deliver a precise prediction of selling prices.

Saved time also has to become usable capacity. The ECB’s worker survey, published on 26 August, stresses that self-reported hours saved translate into additional output only if the organisation can put those hours to productive use. Completing a task faster leaves the next part of the workflow unresolved. ^[10]

A backlog of orders could absorb that capacity. Without new customers, a service might instead shorten waiting times, improve quality, reduce hours or revise recruitment. Those responses have different implications for growth, welfare and income. Compressing them prematurely into one measure of “AI productivity” would obscure the process we need to understand.

A possible two-speed expansion

Investment needed for adoption creates demand before all the production gains have arrived. The ECB’s financing analysis published on 6 October examines links between technological news, credit and investment, alongside differences across sectors and between Europe and the United States. Investment can consequently run on a different timetable from labour income. ^[12]

Financing decisions also connect to competition for credit between AI and other borrowers.

Method and scope

Analytical synthesis with no amounts, calibrated lags, probabilities or causal coefficients. Arrows represent possible transmission channels. Consumption depends on distribution across households, wealth, saving and expectations. No net effect on GDP, inflation or policy rates is estimated.

A favourable path combines useful tools, expanding activity and skills sought by several employers. Efficiency gains then generate more opportunities to work and to bargain. In a less balanced scenario, investment remains strong while some households face fewer professional opportunities. Consumption could become more dependent on other groups’ income and wealth. The direction and size of the aggregate effect remain empirical questions.

Perceptions may matter before losses occur. The August European survey associates negative views of AI with weaker expectations for employment, income and spending. Its authors also connect those views with the wider economic environment. Causality remains open: concern about AI may accompany a broader deterioration in confidence rather than produce it. ^[10]

For monetary policy, these are different configurations. Improving productive capacity, investment-led demand and weaker consumption among particular groups can occur simultaneously. Inferring an automatic interest-rate response from the label “AI” would ignore their relative scale and timing. The article maps possible transmission channels; it does not allocate a measured portion of current inflation or interest-rate movements to the technology.

Keeping alternatives open

Collective organisation is part of bargaining power as well. The OECD chapter on social dialogue and AI describes a role in technology choices, training and working conditions, while noting the constraints on representatives’ expertise and resources. Published in 2023, it identifies mechanisms worth examining rather than guaranteeing the outcome of every workplace negotiation. ^[13]

In our hypothetical case-handling team, a useful agreement could organise learning around exceptions, protected training time and access to the tasks that build expertise. It could specify how quality and quantity will be assessed before performance targets change. These are analytical proposals. Their purpose would be to make changes in the work observable, beyond a licence purchase or a count of people trained to use a particular tool.

Testing Lane’s hypothesis would then require following the same workers and firms over time: actual task changes, available offers, job moves, comparable pay and the content of new agreements. A sustained improvement in those opportunities among adopters would favour the complementary-technology scenario. A contraction concentrated where AI integration has genuinely changed production would strengthen the weaker-bargaining scenario. Sector-wide difficulties would still have to be separated from adoption itself.

As of 14 October 2026, the evidence assembled here does not establish a general decline in European wages caused by AI. It supports a more precise question than how many jobs have disappeared: are these tools giving workers more professional alternatives than they take away? The answer will be found as much in the jobs people can move into as in those they already hold.

Sources and method

Documentary research closed on 14 October 2026. Lane’s speech was checked against research papers, survey evidence and contractual wage indicators. The article distinguishes theoretical mechanisms, descriptive observations, estimates resting on identification assumptions and respondents’ expectations. Each result retains its own observation period; the studies are not combined into a common estimate.

Figures use explicitly published values or labelled calculations. The case-handling team and the distributional and workforce-composition examples are fictional. No original interviews, individual payroll inspection or replication of the ADP, Danish or SAFE datasets was carried out. For the QJE article, the published abstract, Table I and relevant sections of the paper were consulted, including the version hosted by Danielle Li; for Denmark, from the authors’ research summary. The source notes identify the material reviewed.

The editorial image is an AI-generated conceptual illustration. It depicts no actual company or research observation.

1. The AI opportunity for Europe. 13 October 2026. Speech delivered in Dublin. Labour-market paragraph distinguishes possible bargaining effects from observed aggregate trends. No causal European wage estimate.
2. Robot adoption and inflation dynamics. 29 September 2025. Working Paper 2536, abstract and introduction. US metropolitan evidence for 2008–2018 and a structural model of industrial robot adoption. No robot coefficient is transferred to generative AI.
3. The Commoditization of Labor. 19 September 2026. Author-hosted working paper, opening pages. A model of standardisation and outside options, including the identical-firms neutrality result. No current causal chatbot estimate.
4. Generative AI at Work. 4 February 2025; May 2025 issue. Quarterly Journal of Economics 140(2), pp. 889–942. Published abstract and Stanford author page checked. Final sample 5,172; 15% means issues resolved per hour. The published abstract, Table I and sections of the full paper were consulted; 5,172 is the total sample, with 1,636 agents observed after gaining access to the tool. full paper hosted by Danielle Li.
5. Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence. August 2026; revision announced 12 August. ADP payroll data through June 2026. Printed pp. 3–4 distinguish a descriptive shortfall from causal identification. Approximate employment changes concern ages 22–25 and November 2022–June 2026. Earlier 13%/16% adjusted estimates are not reused. Revision announcement.
6. Avec l’essor de l’intelligence artificielle générative, l’investissement numérique tire davantage la croissance aux États-Unis qu’en France, tandis que l’emploi recule dans les activités informatiques des deux côtés de l’Atlantique. 24 March 2026. Insee economic outlook, section on entry-level workers and Figure 11 reading note. Year-on-year change in Q4 2025; ages 15–29 excluding work-study trainees. Different sample and period from the US research; descriptive attribution only.
7. Still Waters, Rapid Currents. July 2026. Research Insight 28/26, authored by Humlum and Vestergaard; based on Discussion Paper 078/26. Roughly 25,000 Danish workers, eleven occupations, administrative outcomes through December 2024. Average null findings and uncertainty bounds cannot be applied to every individual.
8. Adopting and investing in AI: evidence from euro area firms in the SAFE. 2026, Economic Bulletin 2/2026. October–December 2025 SAFE round, footnote 4. Unconditional weighted expectations over the next twelve months. AI has a broad definition. Adjusted regression gaps are distinct. The box shows no separate publication day.
9. AI adoption, productivity and employment: evidence from European firms. 23 January 2026. BIS Working Paper 1325, abstract, data and identification. EIBIS–ORBIS pooled observations 2019–2024. US-peer instrument; interpretation depends on identification assumptions. Broad AI and analytics, not generative AI alone.
10. AI adoption and the productivity promise: what workers report. 26 August 2026. ECB Consumer Expectations Survey. Reported time savings and attitudes are survey evidence; additional output requires usable capacity. Associations between AI sentiment and economic expectations do not settle causal direction.
11. ECB wage tracker at 2.7% in H1 2027, pointing to a modest uptick in negotiated wage growth. 16 September 2026. Table 1 and methodological notes. Agreements signed through end-August; Q1/Q2 2027 contractual signals. Nine participating countries. Future agreement information is not a forecast; coverage is statistical, not bargaining coverage.
12. Funding the AI revolution: evidence from euro area sectors. 6 October 2026. Sectoral funding analysis and model notes. Series have different observation cut-offs; macro model estimated over 1999–2024. Used for transmission mechanisms, with no numerical GDP or interest-rate attribution.
13. Social dialogue and collective bargaining in the age of artificial intelligence. 11 July 2023. OECD Employment Outlook, Chapter 7, by Sandrine Cazes. Social dialogue, technology choices and training. Historical evidence and institutional mechanisms; no universal causal guarantee or current legal claim.
14. Artificial Intelligence: friend or foe for hiring in Europe today?. 4 March 2026. SAFE Q2 and Q4 2025. Cross-sectional hiring comparisons with controls. The stated survey count was corrected to 5,000 on publication day. Motive and intensity of adoption matter; these are not October 2026 employment observations.
15. Eurostat: value added. Output less intermediate consumption; an accounting definition, not a source for the scenario’s invented inputs.
16. OECD: employee compensation. Gross wages and salaries plus employers’ social contributions; a definition, with no new empirical observation.
17. Eurostat: gross operating surplus. A gross balance before capital consumption and further accounting steps; it is distinct from net profit.