Will agentic AI reduce employment?

1. Will productivity gains be as visible as people think?

Probably not, and not straight away. In 1987, the economist Robert Solow wrote: “You can see the computer age everywhere but in the productivity statistics”¹. The gains from computing then remained invisible in measured productivity for a decade.

Brynjolfsson, Rock and Syverson described this mechanism in 2021 under the name of the J-curve³. The gains from a general-purpose technology arrive after complementary investments (organisation, skills, processes), and measured productivity falls before it rises. An AI agent produces nothing useful as long as the process has not been written down, its access has not been decided and nobody has learned to review its outputs.

The time freed up can also be absorbed by higher standards. In 1983, Ruth Schwartz Cowan showed that household appliances had not reduced the time spent on housework: people washed more often, and the standard of cleanliness rose with the machine². The same effect may occur with AI agents (a hypothesis that no survey measures yet): more documents prepared, more replies sent, and therefore more decisions to take. The student who accepted waiting three days for an answer will expect it the same day.

In the French civil service, the announced gains exist. The purchasing directorate expects more than 60% of time saved on finding and onboarding suppliers, and an experiment with 10,000 public servants in six ministries measured up to 16% of time saved on document summaries⁴. The report by the general inspectorates in July 2026 nevertheless finds that these gains remain “diffuse” and rarely measured, and that the vast majority of experiments do not reach the industrialisation stage⁵.

No one knows yet when these gains will become visible, or how large they will be. So do not base a budget or a staffing plan on an announced gain. Measure the processing time of a process before the agent goes into service, then again after a few months of operation.

2. Which tasks will move to agents, and what becomes of human work?

Most likely those that meet three conditions: 1) a high volume; 2) written rules; 3) a verifiable result. In an institution, these are internship agreements, certificates, and first replies to applicants and students. The job of the administrator who handles these files changes in content.

Take an internship agreement. A first agent extracts the information and checks its consistency with the programme's reference framework; a second checks the mandatory clauses (stipend, duration, supervisor); a third prepares the reply to the student and the company; a fourth flags to a person anything that falls outside the rule. The administrator no longer handles every agreement. They handle those that raise a problem (for example 10%, an order of magnitude to calibrate on your own figures) and check the others by sampling. They must keep the difficult cases: it is by handling them that they remain able to judge whether the simple cases are right.

The manager of an automated assembly line no longer tightens bolts. They are accountable for what the line produces and for its pace, and decide when to stop it. Human work shifts in the same way, towards five activities:

1. 1)designing: writing what the agent must do and must never do, based on the academic regulations and procedures;
2. 2)steering: choosing the tasks given to agents and the level of autonomy of each one;
3. 3)testing: checking on real cases, before the agent goes into service and then at each change, that the agent handles files correctly;
4. 4)deciding: settling the cases that fall outside the rule, the decision remaining human;
5. 5)controlling: reading the log of runs, checking a sample, stopping the agent if needed.

These activities require knowledge of the job. The administrator who knows the academic regulations by heart is best placed to write what the agent must do, and to spot what it misses.

Measured usage already follows this shift. The Work Trend Index 2026 analysed more than 100,000 conversations with an AI assistant at work: 49% of exchanges are for analysing, reasoning and deciding, only 17% for producing content⁶. In the same report's survey (20,000 people who use AI at work, ten countries), checking the quality of AI outputs comes first among the human skills growing in importance, cited by 50% of respondents (43% in France)⁷. These figures do not concern higher education: the survey covers all sectors, and the conversation analysis excludes education and covers North America only.

3. What will you do with the time saved?

Write it down before the agent goes into service, and have leadership sign the sentence. Whether agentic AI reduces employment in your institution depends to a large extent on this answer.

Your teams are waiting for this decision. In the Work Trend Index 2026, 65% of people who use AI at work fear falling behind if they do not adapt quickly, and yet 45% consider it safer to stick to their current objectives than to rethink their work with AI (53% and 43% in France)⁷. Only 26% of users consider their leadership clear and aligned on AI (20% in France)⁷.

Time saved that has not been allocated gets absorbed by meetings, reviews and existing tasks. Four allocations can be defended in an institution:

1. 1)analysis: understanding why a rule produces exceptions, and correcting it;
2. 2)relationships: meeting the student before their internship, meeting the host company;
3. 3)complex cases: the student in difficulty, the exemption that calls for judgement, the complaint that is escalating;
4. 4)growth: absorbing an increase in student numbers without additional recruitment.

The fourth is a legitimate choice, provided it is written down as clearly as the other three. One sentence is enough, on this model: “the time saved on internship agreements is used to meet each student before they leave”. It commits leadership and answers the question your teams are asking; a charter that lists prohibitions does not answer it.

The EU AI Act already requires staff to be informed in certain cases. Before putting into service or using a high-risk system in the workplace, the institution as employer must inform workers' representatives and the staff concerned (Article 26, paragraph 7)⁸. The high-risk uses in education are listed in Annex III, point 3 (access and admission, the evaluation of learning outcomes, the assessment of the appropriate level of education, exam proctoring); their obligations apply from 2 December 2027. A chain that checks internship agreements is not covered, as long as it is used neither to admit nor to evaluate a student. Inform staff anyway, for each agent, in writing and before it goes into service: the meeting costs little, and the habit will be in place when the obligation applies to admission.

4. What should you expect in fourteen months?

Agents roughly four times more autonomous. METR measures the length of a human task that a model completes on its own one time in two⁹. This duration was 4 seconds in 2019; it reaches 16 hours in 2026, or two working days for an expert. It doubles every seven months. If the curve holds, it will reach 64 hours by the end of 2027, or a week and a half of work for a project manager (an extrapolation: beyond 16 hours, the measurements are still unreliable).

The list of tasks an agent can carry out on its own will probably grow quickly, and a scope written today (“agents will do this, never that”) risks being overtaken by the end of 2027. The framework can stay stable. It comes down to three names per agent: who designs (instructions and access), who monitors (log, stopping), who signs off what goes out. It also sets, in writing, the level of autonomy of each agent: moving up a level means opening one more access, and that is a decision to be taken.

The sentence on time saved follows the same logic. Written for a first process, it will serve as a reference for the next ones; without it, the question will come back with each new agent, department by department.