Will AI replace jobs, and which ones?

Which jobs are most exposed to AI?

Office jobs top every study. The International Labour Organization (ILO), in an index published in May 2025 with Poland's NASK research institute, ranks clerical work as the most exposed to generative AI. Much of it is reading, sorting, keying in and writing text.

By that index, 25% of jobs worldwide sit in a potentially exposed occupation, rising to 34% in high-income countries. Only 3.3% of global employment falls in the highest exposure band. The ILO scored close to 30,000 job tasks, combining a worker survey, expert review and an AI tool to extend the scoring.

Exposure is also climbing in highly digitized skilled jobs in media, software and finance. It weighs more on women: in high-income countries, the most exposed jobs make up 9.6% of female employment against 3.5% of male employment.

Employers say the same. In the World Economic Forum's Future of Jobs Report 2025, over 1,000 large employers (more than 14 million workers across 55 economies) expect the steepest declines by 2030 in secretarial and admin roles, data entry, bank tellers, cashiers, and bookkeeping and payroll clerks.

Does exposed mean replaced?

No. An exposed job is one where part of the work can be done by AI. The ILO is explicit that its figures measure potential exposure, not actual job losses. Fully automating a job remains limited, because many tasks, even when done faster, still need a person.

An exposed role can go three ways. Some tasks move to AI and the job refocuses on the rest. AI helps the person do more, or do it better. Or nearly every task gets handled and the role goes away. In January 2024 the IMF estimated that about 40% of jobs worldwide, and 60% in advanced economies, are exposed; roughly half could benefit, while for the other half AI could take over key tasks and reduce demand for labor.

In France, the national AI Commission (report delivered in March 2024) reached the same view: AI replaces tasks, not jobs. In 19 jobs out of 20 there are tasks AI can't do, so jobs AI could replace outright would be about 5% of French employment. That figure is the Commission's estimate.

The Commission also used an Insee survey to compare French firms that adopted AI between 2018 and 2020 with firms that didn't. Total employment grew faster at the adopters. But firms that adopted AI for administration or marketing cut mid-level admin and sales staff. That data predates generative AI.

What does the data show so far?

Observed effects, as opposed to projections, are still scarce and mostly American. They show slower hiring of young people in some exposed occupations, by an amount still in dispute, and no wave of layoffs.

Stanford's Digital Economy Lab (Erik Brynjolfsson and co-authors) tracks payroll records for millions of US workers processed by ADP. In the August 2026 revision, using data through June 2026, employment of 22- to 25-year-olds in AI-exposed jobs is 19% below where it would be had it tracked less-exposed jobs. Experienced workers show no such gap. The effect runs mainly through fewer young hires rather than more departures. Declines cluster where AI substitutes for tasks; where it complements workers, employment is flat or rising. The authors call these early, descriptive indicators, not proof of cause.

The OECD is more cautious. Its Employment Outlook 2026, released in July 2026, finds the evidence of AI's impact on younger workers limited so far, and says the business cycle and longer-term shifts in skills demand matter more. It notes that studies conflict, and that exposed jobs are concentrated in sectors (information, finance, professional services) that cut hiring quickly when the economy slows.

In Denmark, Anders Humlum and Emilie Vestergaard (NBER, May 2025) linked usage surveys to government employment records. Two years after ChatGPT's launch, they find no effect on earnings or hours worked and rule out anything above 2%. New tasks do appear, though, such as producing content with AI, checking it, or wiring it into existing tools.

Which jobs are least exposed?

Jobs built on physical work, human contact or being on site. Employers surveyed by the World Economic Forum expect the largest headcount growth in frontline roles: farmworkers, delivery drivers, construction workers, salespeople, food processing workers. Care jobs (nurses, social workers, personal care aides) and teaching follow.

Even in these jobs, a tradesperson or field technician spends hours on quotes, reports and orders, and that part of the job is as exposed as any office role.

A degree is no shield either: the ILO flags rising exposure in digitized jobs in software, media and finance. Exposure tracks how much of the work is handling text and data.

What changes in a small or mid-sized business?

According to Insee, 18% of French companies with 10 or more employees used at least one AI technology in 2025: 31% of those with 50 to 249 employees, 58% of those with 250 or more. Among non-users, 71% say they don't see the point.

In a small business, jobs rarely vanish overnight. AI changes chunks of the week for the office manager, the bookkeeper, the account manager or the technician:

• Data entry: copying an order, invoice or work order from a PDF into a spreadsheet or business software.
• Searching documents: finding a client's last quote, a contract clause, or how a similar job was handled.
• First drafts: email replies, quotes, meeting notes, reports, presentations.
• Sorting: filing incoming email, flagging urgent requests, organizing the documents in a case file.

Who reviews, who decides, who learns?

Review and decisions stay with people. An AI model can be confidently wrong, and the person who signs is accountable (see AI reliability). What you do with the freed time is a management call: more clients served, shorter turnaround, less overtime.

The US data points to a risk: if the tasks you give beginners move to AI, you hire fewer young people and nobody learns the trade. What your senior staff know needs to be written down and passed on before they leave (see passing on know-how).

What does an AI agent do, and what does it not do?

An AI agent chains tasks across your tools: it reads an email, looks up the answer in your files, and drafts a reply or a document (see AI agents at work). It takes on the exposed part of the work, and your teams keep the rest.

• It reads and sorts email, attachments and case documents.
• It finds a past project, a price or a clause, and cites the document it came from.
• It writes a first draft: a reply, a quote, meeting notes, a Word file or a PowerPoint.
• It doesn't decide: taking on a client, setting a price, signing or hiring stay with a person.
• It shouldn't send anything on its own: a person reviews and approves before anything goes out.
• It can't know what was never written down: the history of a client relationship, a promise made on the phone, how a site is really going.

How should you talk about it with your teams?

People worry even when nobody says so. The simplest move is to say early what you're automating and what you're leaving alone. The ILO itself calls for employers and workers to work through these changes together.

• Start from tasks: list with each team what eats their time without adding value (data entry, searches, follow-ups).
• Involve the people who do the work: they know the exceptions no procedure covers.
• Say who signs off: every document AI prepares is checked by a named person.
• Train your teams: in France, short courses can be funded by your OPCO training fund (see AI training through your OPCO).
• In France, companies with 50 or more employees must inform and consult their works council (CSE) on new technologies (Labor Code, article L2312-8), and an AI tool can fall under that. This page is not legal advice.
• Say what the freed-up time will go to, then follow through.

How Saména works with your teams

Saména builds custom AI agents that automate the work that runs on data: email, quotes, case files, documents. If you come with your processes, Saména starts by mapping them to see which tasks an agent can prepare.

Agents prepare, and someone at your company approves before anything is sent. Nothing goes out without them. Agents cite the documents behind every answer, write Word and PowerPoint files, and read your files in SharePoint, OneDrive or Dropbox, read-only, in the folders you choose.

It starts with one call. Two weeks later, your first agent is running, and a dedicated contact stays with you after that.

Frequently asked

Will AI replace accountants?

Not the whole job. Data entry and routine bookkeeping are exposed: employers surveyed by the World Economic Forum in 2025 put bookkeeping and payroll clerks among the fastest-declining roles through 2030. Advice, analysis and responsibility for what gets signed stay with a person.

Which jobs won't AI replace?

No study guarantees any job. The least exposed rely on physical work, human contact or being on site: construction, healthcare, personal care, teaching, sales, delivery. Their paperwork (quotes, reports, orders) is still exposed.

Are young workers hit hardest?

In the US, a Stanford study of payroll data finds a 19% employment gap for 22- to 25-year-olds in exposed jobs, mostly because fewer are hired. The OECD, for its part, finds the impact on young workers limited so far and points to the economy as the bigger factor.

How many jobs will AI eliminate?

Nobody knows. France's AI Commission estimated in 2024 that about 5% of jobs could be replaced outright, and expected the overall effect on employment to be positive. The World Economic Forum projects, worldwide and from all causes combined, 170 million jobs created and 92 million displaced by 2030: a projection based on what employers report.

Should you cut admin staff when you adopt AI?

The research doesn't point there: it shows tasks changing before headcount does. You can also use AI to handle more work with the same team, or to give people time back for clients. It's a management decision, and one to announce clearly.