AI can reduce payroll. But who will buy what companies produce?

Abstract

Reducing payroll through AI has become an explicit objective for part of the corporate world. In 2025, 41% of employers surveyed by the World Economic Forum said they expected workforce reductions where AI could automate tasks. Yet observed data remain much more nuanced: in 2026, McKinsey reported that 14% of respondents had actually reduced headcount because of AI over the previous year, compared with 32% who had expected to do so one year earlier. In France, research on earlier waves of AI adoption even found higher employment and sales among adopting firms.

The central question is therefore not whether “AI will destroy jobs”. It is more subtle: what happens when productivity grows faster than the economy’s ability to recreate labour income, skills and demand?

At firm level, reducing labour cost can improve margins. At economy-wide level, a worker’s wage is also a consumer’s income. If AI gains concentrate in margins and capital returns without sufficiently translating into lower prices, higher wages or new activity, demand can become the next bottleneck.

Thesis. AI is not only a productivity technology. It is a technology for redistributing value. Its impact will depend as much on where the gains go — wages, prices, profits, investment, new tasks — as on how much work is automated.

---

Executive summary

If you only retain six ideas:

• A broad-based fall in payroll is not yet established. Reductions exist, but remain selective by function, industry and experience level, with particularly visible signals among some junior workers.
• Payroll is not the same thing as headcount. A company can reduce headcount while maintaining payroll if the remaining scarce skills become more expensive.
• Purchasing power depends on wages AND prices. If AI lowers prices substantially, wage pressure can be offset. If gains accrue mainly to margins and capital, it cannot.
• The most interesting macroeconomic risk is a demand bottleneck. The Bank for International Settlements explicitly models a scenario in which automation eventually slows growth because displaced workers are also lost consumers.
• Social risk is also an operating risk. AI-related insecurity can weaken psychological safety and knowledge sharing — exactly when firms need employees to explain and formalise their processes.
• The junior career ladder may break before aggregate employment collapses. Early US data show a specific deterioration in employment for young workers in highly AI-exposed occupations.

---

1. Payroll reduction is not “inevitable” — but it has become an explicit strategy

A seductive but overly simple idea needs correcting first:

More automation ≠ automatically lower payroll.

Payroll depends on at least two variables:

Payroll = number of employees × average compensation.

A company can therefore:

• reduce headcount;
• retain a larger share of senior profiles;
• pay more for scarce AI, data, architecture or governance skills;
• and still end up with stable or even higher payroll.

The available signals already show this ambiguity.

The World Economic Forum reported in 2025 that 41% of surveyed employers expected to reduce their workforce where AI could automate tasks. But 77% also planned to invest in upskilling, and 47% expected to move employees into other roles rather than simply eliminate them.

One year later, McKinsey found a significant gap between intention and reality: 32% of 2025 respondents had expected AI-related headcount reductions; in 2026, only 14% said those reductions had actually happened in the preceding year. That did not prevent 39% from still expecting AI-related headcount reductions in the year ahead.

The most accurate reading is therefore neither “nothing is happening” nor “mass layoffs are certain”. It is:

companies are actively trying to convert AI productivity into lower costs, but the outcome depends on occupation, experience level and whether work can be reallocated.

The ILO reaches a similar conclusion. Its 2025 global index estimates that roughly one job in four has some exposure to generative AI, but sees job transformation as more likely than complete disappearance.

The most worrying signal may not be total employment

In August 2026, the Stanford Digital Economy Lab published an update based on ADP payroll data covering millions of US workers. The researchers do not find broad displacement across the entire economy.

They do find something else: among 22–25-year-olds in highly AI-exposed occupations, employment stands 19% below the level it would have reached had it followed the path of comparable workers in less exposed occupations.

That is much more specific than a “future without work” story.

AI may first disrupt entry points into careers before it materially reduces aggregate employment.

Sources: World Economic Forum, Future of Jobs 2025, McKinsey, AI job losses fall short of forecasts, 2026, Stanford Digital Economy Lab, 2026, ILO–NASK, 2025.

---

2. Purchasing power depends on more than wages

Public debate often treats lower wages and lower purchasing power as the same thing.

That is incomplete.

Real purchasing power depends on nominal income and the price level.

As a first approximation:

Change in real purchasing power ≈ change in income − inflation.

If AI enables a company to produce 30% more cheaply and competition passes those gains on to consumers, prices can fall. A worker whose nominal income stagnates may still be able to buy more.

Conversely, if productivity gains:

• mostly increase margins;
• accrue to capital owners;
• are captured by a small number of model, cloud or chip providers;
• and reduce end prices only slightly,

then consumers do not automatically recover the gain.

A simple example

Scenario A

• wages: −3%
• prices: −7%

Real purchasing power can increase.

Scenario B

• wages: −3%
• prices: −1%

Purchasing power falls.

The decisive question is therefore not only:

How many jobs can AI eliminate?

It is:

Who captures the difference between the old production cost and the new one?

The IMF highlights precisely this distributional risk: if AI disproportionately complements high-income workers and increases returns to capital, it can increase both labour-income inequality and wealth inequality.

---

3. The economic paradox: payroll saved by one company is potential revenue elsewhere

This may be the most important idea in the article.

For an individual firm:

lower labour cost → potentially higher margin.

For the economy as a whole:

lower labour income → potentially lower consumption.

A worker is simultaneously a production cost for one employer and a customer for other companies.

The Bank for International Settlements formalised this point in its 2026 Annual Economic Report. One of the scenarios studied is a “demand bottleneck”: automation initially raises output, but growth eventually slows if labour’s income share falls enough to weaken demand.

The mechanism is straightforward:

1. AI reduces labour required for some tasks;
2. unit costs fall;
3. companies improve margins or lower prices;
4. if labour income falls too far, consumption slows;
5. companies face less of a production constraint and more of a demand constraint.

The BIS states the mechanism bluntly: a displaced worker is also a lost consumer.

This is not a certain forecast. The loop can be broken if productivity gains finance:

• new tasks;
• new industries;
• genuinely lower prices;
• higher wages in complementary activities;
• productive investment;
• or redistribution mechanisms.

But it reveals a limitation in business cases that look only at the firm.

An economy cannot optimise every company indefinitely as if demand for everyone else’s output were independent of the income paid by the system as a whole.

Source: BIS Annual Economic Report 2026.

---

4. Automation can increase revenue — without making the advantage permanent

Early evidence does not show only cost savings.

It also shows expansion effects.

A 2025 study of French companies adopting AI between 2017 and 2020 finds higher employment and sales after adoption, consistent with a classic mechanism: productivity gains make a firm more competitive, it wins market share and then grows.

But the effect is not uniform. AI applications in some administrative processes are associated with more negative employment effects.

This is exactly what the economics of automation predicts:

• displacement effect: machines perform tasks previously done by labour;
• productivity effect: lower costs make production more profitable;
• reinstatement effect: new human tasks appear.

The final outcome depends on the relative strength of those three forces.

A company can gain while its sector employs fewer people

The history of industrial robots in France provides an instructive precedent.

Robot-adopting firms increased productivity, value added and even their own employment. But they also took market share from competitors. At industry level, the employment effect became negative.

That distinction matters for executives:

“Our company is creating jobs because of AI” does not mean AI is creating jobs for the sector.

A highly productive company may simply be shifting economic activity toward itself.

Sources: Aghion et al., 2025, Acemoglu, LeLarge & Restrepo, France.

---

5. Some business models are threatened by their own productivity

AI does not only change costs.

It can attack the metric on which a company charges customers.

IT services and consulting

If revenue is largely based on:

number of consultants × number of days × day rate

then a technology that reduces the number of days required can create a paradox.

It improves operational productivity while potentially reducing billable volume.

The economic response is a shift toward:

• fixed-price delivery;
• outcomes;
• delivered capability;
• intellectual property;
• or higher-value judgement and arbitration services.

SaaS

A different paradox appears in per-seat pricing.

If AI agents perform work that previously required several users, the client organisation may need fewer software seats.

McKinsey already notes that AI-native software businesses are more likely to use consumption- or outcome-linked pricing, while traditional models remain more dependent on flat subscriptions or seats.

The number of software companies using consumption-based pricing more than doubled between 2015 and 2024.

The question changes

Before:

How many users do we have?

Tomorrow:

How much value does our system produce?

That is a profound shift.

AI can push providers toward charging more often:

• per transaction;
• per resolution;
• per task;
• per unit processed;
• per economic outcome.

Sources: McKinsey, AI software pricing, 2025, McKinsey, 2026.

---

6. Psychological risk becomes knowledge risk

An AI-related layoff does not only have a social cost.

The expectation of layoffs can already change the behaviour of those who remain.

That is particularly dangerous because AI programmes often ask employees to:

• document their work;
• make rules explicit;
• reveal exceptions;
• correct AI outputs;
• train new workflows.

In other words:

the organisation asks employees to transfer exactly the knowledge that may appear to reduce their own future value.

That is a classic incentive problem.

Recent work on AI-induced job insecurity suggests that insecurity can reduce psychological safety, then reduce knowledge sharing and increase knowledge-hiding behaviour.

A longitudinal 2026 study involving 407 employee–HR dyads finds that AI-related insecurity affects innovation through an indirect pathway:

insecurity → lower psychological safety → less knowledge sharing → lower innovation.

The most useful result is that the effect can be mitigated when employees perceive more:

• transparency;
• control;
• human agency in the system.

The management implication is clear:

an AI strategy built primarily around the threat of headcount reduction can damage exactly the cooperation automation needs in order to work.

Sources: Kim & Lee, 2026, Jeong, Kim & Lee, 2023.

---

7. Work intensification is another plausible outcome

Automating part of a job does not automatically mean working less.

A company can convert saved time into:

• more volume;
• more control;
• more reporting;
• higher availability expectations.

OECD surveys of AI at work show an ambivalent picture.

Many workers report better performance and sometimes higher job satisfaction.

But where AI becomes a tool for algorithmic management, outcomes are less favourable:

• higher work intensity;
• sometimes lower autonomy;
• privacy concerns;
• stress linked to monitoring.

In finance, the OECD reported that 85% of AI users subject to some form of algorithmic management said their work pace had increased, compared with 74% of other surveyed AI users.

The most credible psychological scenario may therefore not be:

“AI eliminates work”.

It may be:

fewer employees, each assisted by AI, under a higher production norm.

Source: OECD, Employment Outlook 2023.

---

8. The least visible risk: remove juniors today, run out of seniors tomorrow

A large share of junior work has always served two functions.

It produced something.

It also trained the future expert.

A junior analyst builds summaries before learning to arbitrate.

A junior developer handles simple tasks before understanding architecture.

A junior lawyer prepares files before carrying strategy.

A junior consultant builds analyses before challenging an executive.

If AI takes over those tasks first, the company saves time today.

But a question appears:

where will tomorrow’s experts learn the elementary steps that build judgement?

The Stanford data showing deterioration among 22–25-year-olds in AI-exposed occupations make this less theoretical.

PwC also reports that in 2026, junior vacancies in highly AI-exposed occupations were much more likely to require skills traditionally associated with senior profiles — particularly judgement and leadership.

That creates a risk of career-ladder compression:

junior roles ↓
while demand for senior judgement stays stable or ↑

In the short term, this looks efficient.

In the long term, it can create:

• shortages of people able to verify AI;
• wage inflation for experienced specialists;
• weaker succession pipelines;
• greater dependence on a small group of experts.

Removing a training task is not only a saving. It can also create skills debt.

Sources: Stanford Digital Economy Lab, 2026, PwC Global AI Jobs Barometer 2026.

---

9. AI may move a significant share of value from labour to capital

The global question goes beyond individual companies.

AI is extremely capital intensive.

Building frontier systems requires:

• chips;
• data centres;
• cloud infrastructure;
• energy;
• data;
• models;
• scarce talent.

These inputs have strong economies of scale.

UNCTAD estimates that the AI market could grow from $189 billion in 2023 to $4.8 trillion by 2033.

But infrastructure and investment are already highly concentrated. UNCTAD notes, among other things, that 100 companies account for around 40% of global corporate R&D spending.

More broadly, between 2017 and 2025, the share of global sales held by the five largest digital multinational companies rose from 21% to 48%, according to UNCTAD.

This creates a second redistribution loop:

client company
→ spends less on human labour
→ spends more on compute, cloud, models and infrastructure
→ part of the value leaves local payroll
→ and flows toward highly concentrated global suppliers.

Productivity gains can therefore be real without remaining fully inside the local economy.

Sources: UNCTAD, Technology and Innovation Report 2025, UNCTAD, digital-market concentration.

---

10. Four economic futures are possible

It is dangerous to present one inevitable trajectory.

The economic outcome will depend mainly on two forces:

1. how many new tasks and new markets AI creates;
2. how much of the productivity gain flows to labour rather than capital.

Scenario 1 — Complementarity

AI raises employee productivity.

Companies earn more.

Complementary skills become more valuable.

Wages rise.

Scenario 2 — Expansion through new tasks

AI automates heavily, but new activities appear fast enough to absorb displaced workers.

That is one of the mechanisms through which previous general-purpose technologies expanded output without eliminating work altogether.

Scenario 3 — Lean-firm economy

A small number of highly productive, highly skilled employees generate large amounts of value.

Employment polarises.

Capital income grows faster than labour income.

Economic concentration increases.

Scenario 4 — Demand bottleneck

Automation progresses faster than the creation of new labour income.

Potential output rises.

Demand does not keep pace.

Growth becomes constrained not by the ability to produce, but by the ability of customers to buy.

None of these scenarios is guaranteed.

The key point is that technology alone will not choose which one materialises.

---

11. Specific dangers for the company

Executives should monitor at least seven risks.

Risk	Mechanism
Demand erosion	Customers themselves may face income pressure
Business-model cannibalisation	Fewer seats, billable days or human tasks to monetise
Skills debt	Junior tasks that trained future experts disappear
Knowledge hiding	Insecurity reduces knowledge sharing needed for automation
Vendor dependency	Payroll savings become recurring cloud, model and token spend
Work intensification	Productivity gains become the new volume norm
Temporary advantage	Once every competitor adopts AI, the gain often flows into lower prices

The last point is particularly important.

A technology can create a huge advantage for the first adopter and almost no durable advantage once the whole industry uses it.

At that point, productivity simply becomes the new baseline.

---

12. What an executive committee should actually measure

“Jobs eliminated” is too weak a KPI.

A serious dashboard should cover several layers.

Productivity

• revenue per employee;
• cost per transaction;
• time to validated outcome;
• full AI cost per process.

Labour

• payroll / revenue;
• headcount by experience level;
• junior hiring;
• wage premium for scarce skills;
• voluntary turnover.

Skills

• number of learning tasks removed;
• time required to make a junior autonomous;
• dependency on a small number of experts;
• ability to operate without the external provider.

Demand

• average basket / contract value;
• price sensitivity;
• churn;
• customer affordability;
• share of productivity gains passed through into prices.

Organisational psychology

• perceived job security;
• trust in the AI strategy;
• autonomy;
• knowledge sharing;
• voluntary versus imposed use.

Economic dependency

• AI spend / revenue;
• marginal cost per request or workflow;
• number of critical suppliers;
• model substitutability;
• share of margin transferred to compute providers.

---

Conclusion

The question “how many people can we replace with AI?” is probably too narrow for executives.

A better question is:

What happens to every euro of productivity gained through AI?

Does it become:

• margin;
• lower prices;
• new investment;
• new jobs;
• higher wages;
• dividends;
• cloud and model consumption;
• or simply less disposable income somewhere else in the economy?

Payroll reduction is possible, and in some occupations it is already visible.

But it is neither automatic nor economically neutral.

A company can improve its own economics by reducing the labour required for a given output.

An entire economy can prosper sustainably only if it simultaneously finds new ways to distribute enough income, skills and purchasing power to absorb what it produces.

The real AI issue may therefore not be the disappearance of work.

It may be the movement of income, economic power and the ability to consume.

And for companies, that question eventually comes back to one place:

revenue.

---

Methodology note

This article deliberately separates observed data, employer expectations and macroeconomic scenarios. World Economic Forum and McKinsey figures reflect employer or executive responses; they are not certain forecasts. Stanford data describe recent US labour-market developments and should not be mechanically extrapolated to France. Aghion et al.’s French study focuses mainly on AI adoption between 2017 and 2020, before mass diffusion of LLMs. BIS scenarios and Acemoglu’s frameworks are economic models used to analyse possible mechanisms, not precise predictions.

---

Main references

1. World Economic Forum (2025). The Future of Jobs Report 2025.
2. Brynjolfsson, E., Chandar, B. & Chen, R. (2026). Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence. Stanford Digital Economy Lab.
3. Gmyrek, P. et al. (2025). Generative AI and Jobs: A Refined Global Index of Occupational Exposure. ILO Working Paper 140.
4. Aghion, P. et al. (2025). How Different Uses of AI Shape Labor Demand: Evidence from France. AEA Papers and Proceedings.
5. Acemoglu, D. (2024). The Simple Macroeconomics of AI. NBER Working Paper 32487.
6. Acemoglu, D. & Restrepo, P. (2019). Automation and New Tasks: How Technology Displaces and Reinstates Labor. Journal of Economic Perspectives.
7. Acemoglu, D., LeLarge, C. & Restrepo, P. (2020). Competing with Robots: Firm-Level Evidence from France. NBER.
8. Bank for International Settlements (2026). Annual Economic Report 2026.
9. Cazzaniga, M. et al. (2024). Gen-AI: Artificial Intelligence and the Future of Work. IMF Staff Discussion Note.
10. OECD (2023). Artificial intelligence, job quality and inclusiveness.
11. Kim, B.-J. & Lee, J. (2026). How Human-Centered AI Buffers the Negative Effects of AI-Induced Job Insecurity on Sustainable Innovation. Sustainable Development.
12. Jeong, J., Kim, B.-J. & Lee, J. (2023). The effect of job insecurity on knowledge hiding behavior. Frontiers in Public Health.
13. UNCTAD (2025). Technology and Innovation Report 2025: Inclusive Artificial Intelligence for Development.
14. McKinsey & Company (2026). AI job losses fall short of forecasts.
15. McKinsey & Company (2025). Upgrading software business models to thrive in the AI era.
16. PwC (2026). Global AI Jobs Barometer 2026.