AI is not eliminating work: it is moving the bottleneck

Abstract

In 2025, 18% of French companies reported using at least one artificial-intelligence technology. Yet more than half of AI-using firms still said a lack of expertise was holding them back. That gap helps explain a paradox that experimental research can now measure: AI can cut the time required for some writing tasks by 40%, raise customer-support productivity by 14%, or dramatically accelerate an isolated programming exercise. But it can also reduce the probability of a correct answer by 19 percentage points when a task falls outside the model’s capabilities, slow experienced developers working in complex mature repositories, and weaken learning when answers replace reasoning.

The conclusion is neither “AI increases productivity” nor “AI destroys quality”. It moves the work: from production toward framing, verification, integration and accountability. Economic value therefore depends less on how much output is generated than on whether an organisation can convert local speed into a reliable, maintainable result that people actually use. This article offers an evidence-based reading of that shift — and an operational framework for measuring it.

Thesis. AI reduces the cost of producing a first result. It does not automatically reduce the cost of producing a correct, integrated and durable result.

1. Measure productivity in the wrong place and you get the wrong conclusion

Counting documents, tickets or lines of code measures production throughput. An organisation pays for a validated outcome: a correct decision, a feature that is used, an incident avoided, a process shortened.

A more useful way to reason is:

Net AI value = production gain − verification − rework − coordination − risk − lost learning.

This is an analytical framework, not a universal metric. Its purpose is to stop a visible gain at task level from hiding a cost that has simply moved into review, security, maintenance or the downstream process.

The findings only look contradictory:

Context	Method	Main result	What the number does not prove
Professional writing, 453 graduates	Randomised experiment	Time −40%; quality +18%	Not long-running, collective or regulated work
Customer support, 5,179 agents	Staggered deployment, operational data	Productivity +14%; +34% for novices	Much smaller effect for experienced agents
Consulting, 758 professionals, including 385 on tasks “inside the frontier”	Pre-registered experiment	12.2% more tasks; 25.1% faster; quality increased	On one task outside the frontier, correct answers fell by 19 points
Isolated JavaScript server task, 95 developers	Randomised experiment	Time −55.8%	Standardised greenfield exercise, not an existing product
Three companies, 4,867 developers	Three field experiments	Completed tasks +26.08%	Preprint; volume metric, quality not included
Mature open-source repositories, 16 experts and 246 tasks	Randomised experiment	Time +19% with AI	Early-2025 tools; small developer sample

Sources: Noy & Zhang, Science, Brynjolfsson, Li & Raymond, NBER, Dell’Acqua et al., Organization Science, Peng et al., Microsoft Research, Cui et al., Microsoft Research, METR.

The determining factor is therefore not simply model capability. It is the fit between the model, the task, the context available to it and the cost of being wrong.

The METR result is especially revealing. Before the experiment, developers expected AI to save them 24% of their time. After completing the tasks, they still believed they had been around 20% faster. The measured result was a 19% slowdown. In February 2026, a new data collection using more recent tools produced a favourable point estimate, but METR considered the sample too biased to quantify the improvement reliably. The durable lesson is not “AI slows developers down”. It is that feeling faster is not a delivery metric. METR, 2026 update.

2. Cognitive load does not disappear — it changes shape

An experiment across 66 companies and 7,137 knowledge workers randomly provided access to a generative tool integrated into email, meetings and documents. Among users in the treatment group, time spent on email fell by roughly two hours per week and work outside normal hours decreased. But researchers found no detectable change in the overall quantity or composition of tasks. AI freed time locally without, by itself, redesigning the job. Dillon et al., NBER.

A CHI 2025 study offers a complementary mechanism. Across 936 work situations reported by 319 knowledge workers, higher confidence in AI was associated with less self-reported critical thinking. Cognitive effort moved:

• from finding information to verifying it;
• from solving the problem to integrating the response;
• from execution to supervision.

This is survey evidence, not proof of cognitive decline. But it shows where the effort can reappear after “generation time” disappears from the metric. Lee et al., CHI 2025.

Immediate performance and future capability are different variables

In a field experiment involving almost 1,000 high-school students, unrestricted access to GPT-4 improved performance during practice by 48%. Once the tool was removed, that group performed 17% worse than the no-AI group. A tutored version designed to provide hints rather than answers produced an even larger immediate gain (+127%) without a significant penalty on the later no-AI assessment. Bastani et al., PNAS.

The sample is educational, so directly transferring the effect size to work would be wrong. The mechanism, however, is testable in organisations. If AI performs the steps that previously developed junior judgement, who will build the expertise needed to supervise AI tomorrow? Documentation, explanatory review, task rotation and knowledge transfer then become risk controls rather than administrative extras.

3. AI can replace part of coordination — not collective accountability

A pre-registered experiment published in 2026 assigned 791 Procter & Gamble professionals to work individually or in pairs, with or without AI, on real product-innovation problems. Individuals assisted by AI reached the performance level of teams without AI. The proposals also became more balanced: R&D profiles incorporated more commercial considerations and commercial profiles incorporated more technical ones. Participants reported more positive emotions.

The most useful management insight sits elsewhere: AI mainly improved idea generation, while human judgement retained its value in selection. Dell’Acqua et al., Organization Science.

That suggests a new division of labour:

AI is well placed to	The team remains responsible for
widen the option space	define the real problem and constraints
generate variants quickly	select according to strategy and risk
translate between functional vocabularies	arbitrate interests and own the decision
produce a first formal version	guarantee coherence, security and maintainability

The same technology can therefore reduce a functional silo and weaken a team if it replaces the conversations in which disagreement, project memory and accountability are built. The P&G experiment measures bounded problem solving; it does not prove that an AI agent can replace the social functions of a durable product team.

4. Well-being: operational relief and insecurity can coexist

The psychological evidence is mixed, and that ambiguity is useful.

In OECD surveys of workers in finance and manufacturing, more than 80% of AI users reported better performance; 63% reported greater enjoyment of work; 54% in finance and 55% in manufacturing reported an improvement in mental health. But the OECD also identifies work intensification, more stress in some studies and less favourable outcomes for people subject to algorithmic management. Around one worker in five said they had lost autonomy. These are self-reported data collected before generative AI fully diffused, and they are exposed to selection effects because they mostly describe workers who remained in their organisations after adoption. OECD, Employment Outlook 2023.

The apparent contradiction is rational: a worker can enjoy the tool day to day while fearing its economic consequences. In the same OECD survey, three in five workers said they were worried about losing their job to AI in the next ten years; 41% of manufacturing respondents and 42% of finance respondents expected wages in their sector to fall.

How management uses the technology matters as much as the technology itself. EU-OSHA data on algorithmic task allocation and monitoring associate those systems with more time pressure, overload, isolation and loss of autonomy. That should not be conflated with voluntary generative assistance; it shows what can happen when AI becomes a control mechanism rather than a resource. EU-OSHA.

Operational implication: saving time improves well-being only if some of that gain remains available to reduce overload, support learning or improve judgement. If it is immediately converted into a higher volume norm, AI can make each task easier while making the day more intense.

5. Economics: minutes saved do not automatically become payroll reductions

The strongest Danish study so far links adoption surveys to administrative data covering 25,000 workers and 7,000 workplaces. Despite rapid adoption, reported gains and new AI-related tasks, the authors find no detectable effect on earnings or recorded hours two years after ChatGPT and rule out effects larger than 2%. What changes first is the structure of work: content production, supervision, integration and occupational transitions. Humlum & Vestergaard, NBER, revised 2026.

That result invalidates a common shortcut:

20% less time on a task ≠ 20% removable payroll capacity.

Time saved can be absorbed by complementary tasks, higher volumes, additional controls or organisational friction. Conversely, a study of French firms adopting AI between 2017 and 2020 finds higher employment and sales after adoption, consistent with an expansion effect stronger than substitution. The effect varies by use case: employment growth for some applications, including cybersecurity, and declines for some administrative processes. This study covers pre-generative AI and should not be extrapolated mechanically to LLMs. Aghion et al., AEA Papers and Proceedings.

Globally, the ILO estimates that one in four jobs has some exposure to generative AI, but only 3.3% of jobs fall into the highest exposure category. Its central diagnosis is job transformation rather than wholesale replacement. ILO–NASK, 2025.

The economically useful questions are therefore not “how many people can AI replace?” but:

1. what currently constrains value creation;
2. which part of that constraint AI actually reduces;
3. which new bottleneck appears next;
4. who captures the gain: customer, company, team or worker;
5. what delayed costs appear in quality, incidents, skills or turnover.

6. In software, local speed has to survive the delivery system

DORA data illustrate the move from individual workstation to delivery system. In analysis based on 2024 data, a 25% increase in AI adoption was associated with a 1.5% drop in throughput and a 7.2% drop in delivery stability despite individual-level benefits. In the 2025 report, based on nearly 5,000 professionals, the relationship with throughput and product performance had turned positive, while the relationship with stability remained negative. DORA describes AI as an amplifier: it accelerates strong teams and exposes weak systems. These are statistical associations from surveys, not causal effects. DORA, GenAI analysis ; DORA 2025.

Why? Faster code production increases the size and frequency of changes. Without small batches, automated tests, disciplined versioning, decoupled architecture and fast feedback loops, review and operations become the next bottleneck.

Metrics worth following

Level	Useful metrics	False signal to avoid
Individual	time to validated result, rework rate, verification time	lines of code or prompts
Team	cycle time, review delay, escaped defects, out-of-hours load	unweighted tickets closed
Delivery	lead time, deployment frequency, change failure rate, recovery time	velocity alone
Product	adoption, time to first value, journey success rate, cost per outcome	features shipped
Skills	autonomy without AI, review quality, documentation coverage, effective transfer	training volume

7. Where a senior taskforce now creates value

In France, AI use tripled between 2023 and 2025, yet 54% of firms not using it cited a lack of expertise; more than half of firms already using AI reported the same barrier. Insee, July 2026.

The expertise required is not simply the ability to call a model. It includes the ability to:

• choose tasks where AI creates measurable advantage;
• build the context the model does not possess;
• set autonomy limits according to the cost of error;
• integrate testing, security, observability and rollback from the start;
• measure product value and delivery quality rather than generated volume;
• preserve learning and leave the client team autonomous.

That is the role of a senior taskforce: turn generation capability into a reliable production system.

At Averiane, this logic translates into MVPs generally delivered in six weeks, with scope, timeline and expected outcomes defined before development. Fixed-price delivery aligns the engagement with results rather than accumulated days. Tests, metrics and dashboards make quality and risk visible during the engagement. Documentation, training and knowledge transfer prevent initial speed from becoming future dependency.

AI can produce faster. The value of expertise is knowing what to produce, how to verify it and how to make it durable.

Averiane — robust MVPs in six weeks. Measurable quality. Complete knowledge transfer.

Methodology note

This article prioritises randomised trials, field experiments, peer-reviewed research and data from public institutions. Comparisons across studies are used to show heterogeneity; they are not a meta-analysis. Tasks, tools, dates, populations and metrics differ. Self-reported findings are identified as such. Long-term effects on health, learning and employment are still less established than immediate effects on bounded tasks.

Main references

1. Noy, S. & Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science.
2. Brynjolfsson, E., Li, D. & Raymond, L. (2023, revised). Generative AI at Work. NBER.
3. Dell’Acqua, F. et al. (2026). Navigating the Jagged Technological Frontier. Organization Science.
4. Dell’Acqua, F. et al. (2026). The Cybernetic Teammate. Organization Science.
5. Cui, Z. et al. (2025). The Effects of Generative AI on High-Skilled Work. Microsoft Research, preprint.
6. Becker, J. et al. (2025). Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity. METR.
7. Dillon, E. et al. (2025, revised). Shifting Work Patterns with Generative AI. NBER.
8. Lee, H.-P. et al. (2025). The Impact of Generative AI on Critical Thinking. CHI.
9. Bastani, H. et al. (2025). Generative AI without guardrails can harm learning. PNAS.
10. Humlum, A. & Vestergaard, E. (2025, revised 2026). Still Waters, Rapid Currents. NBER.
11. Aghion, P. et al. (2025). How Different Uses of AI Shape Labor Demand: Evidence from France. AEA Papers and Proceedings.
12. OECD (2023). Artificial intelligence, job quality and inclusiveness.
13. ILO & NASK (2025). Generative AI and Jobs: A Refined Global Index of Occupational Exposure.
14. DORA (2025). State of AI-assisted Software Development.
15. Insee (2026). Information and communication technologies in firms in 2025.