CyphrQ
Menu

Will AI Replace Jobs? What the IMF, ILO and Usage Data Say

Will AI replace jobs? The IMF, the ILO and Anthropic's usage data measure different things. Here is what each number means and how to use it.

By CyphrQ.ai

Written with AI assistance and human editorial review. Every figure linked to its source.

11 min read

Will AI replace jobs? The research says AI will change more jobs than it removes. The IMF and the ILO both measure exposure, meaning how many tasks AI could touch. Neither says how many jobs will disappear. Usage data shows how people use AI today. The three use different methods, so their numbers cannot be added together.

What does "exposure to AI" actually mean?

Exposure means a job contains tasks that AI could do or help with. It is not a forecast of job loss.

A job can be exposed and still grow. If AI speeds up part of the work, the person may do more of it, do it better, or move to the parts a machine cannot do. An exposed job can also shrink, if the exposed tasks were most of the job.

So the first question for any headline number is: what did the authors count? A task score, a survey, or logged conversations with a chatbot? Each answers a different question.

What does the IMF say about AI and jobs?

The IMF's January 2024 analysis is the source of the "almost 40 percent" figure. In a blog post by IMF Managing Director Kristalina Georgieva, the IMF states that "almost 40 percent of global employment is exposed to AI." It gives these figures by income group:

Group Share of jobs the IMF says are exposed to AI
Advanced economies about 60 percent
Emerging markets 40 percent
Low-income countries 26 percent

The post attributes these to "IMF staff analysis." The underlying paper is the IMF Staff Discussion Note "Gen-AI: Artificial Intelligence and the Future of Work", published 14 January 2024. Only its summary page and the blog post were read for this post, so its method is not described in detail here.

The IMF's other findings matter as much as the headline:

  • Half the exposed jobs benefit, half may not. For advanced economies, the blog says "roughly half the exposed jobs may benefit from AI integration, enhancing productivity." For the other half, AI could perform key tasks now done by people. That could lower labour demand, wages and hiring, and "in the most extreme cases" some jobs may disappear.
  • Richer countries feel it first. The summary says advanced economies will see AI's benefits and pitfalls sooner, mainly because more of their jobs are cognitive-intensive.
  • Inequality could rise. The summary says labour income inequality may rise if AI's complementarity with high-income workers is strong, and that rising capital returns would increase wealth inequality.
  • Who is exposed. Women and college-educated workers are described as more exposed but also better placed to benefit. Older workers are described as potentially less able to adapt.

Note the word "may." The IMF is describing a range of outcomes, not a prediction of how many jobs will be lost.

What does the ILO say about AI and jobs?

The International Labour Organization (ILO) takes a task-level view. Its May 2025 research brief, Generative AI and jobs: A 2025 update, updates its 2023 index. It is based on ILO Working Paper 140.

Method. The index is "assessed at the 6-digit occupational level covering nearly 30,000 tasks." It combines task-level data with human and AI-generated input. It sorts occupations under ISCO-08, the international occupation classification, into four exposure gradients, based on each occupation's mean exposure score and the variability of its tasks.

Findings, as the ILO states them:

  • "One in four workers across the world are in an occupation with some degree of GenAI exposure."
  • "3.3% of global employment falls into the highest exposure category," with a gap between female (4.7%) and male employment (2.4%).
  • Exposure by income: "11% of total employment in LICs vs 34% in HICs." LICs and HICs are low-income and high-income countries.
  • "Clerical occupations continue to have the highest exposure levels."
  • The conclusion: "As most occupations consist of tasks that require human input, transformation of jobs is the most likely impact of GenAI."

The brief adds that the average automation score barely moved between the two editions, from 0.30 in 2023 to 0.29 in 2025. What changed was the spread between occupations, which narrowed a lot (standard deviation from 0.30 to 0.14). The brief says the transition should be managed through social dialogue.

What does real usage data show about automation and augmentation?

The IMF and the ILO estimate what AI could do. Usage data records what people do with it. One public example is the Anthropic Economic Index.

In its third report, published 15 September 2025, Anthropic describes how it measures this. It uses "a privacy-preserving classification method" that sorts anonymised Claude.ai conversations into task groups defined by O*NET, a US occupational task database. For business use, it sampled 1 million transcripts from August 2025 from first-party API customers.

What it found:

Measure Result as stated
Claude.ai conversations: automation 49.1%
Claude.ai conversations: augmentation 47%
API conversations: automation patterns 77%
API conversations: augmentation 12%
API traffic mapped to computer or mathematical tasks 44% (36% on Claude.ai)

Anthropic says this is the first time automation has been more common than augmentation on Claude.ai. Its "directive" conversations, where a person hands over a task with minimal back-and-forth, rose from 27% to 39% since December 2024. Here, "automation" means the AI completes the task. "Augmentation" means the person and the AI work on it together.

Three cautions apply, and the authors raise some of them:

  1. It is one vendor's data about its own product. It says nothing about people using other assistants.
  2. It counts conversations, not jobs. A task being automated does not mean a worker was let go.
  3. The authors call the overall trend "noisy" and say their sample size limits how much detail they can explore.

Still, the API figure is a useful signal. When companies build AI into their own systems, the pattern in this sample leans heavily towards handing the task over.

Why can't you add these numbers together?

Because each one answers a different question.

Source What it measures Method Headline figure
IMF (2024) Share of employment exposed to AI IMF staff analysis (method not read in full) almost 40% global
ILO (2025) Share of workers in occupations with some GenAI exposure Task-level index, ISCO-08, nearly 30,000 tasks one in four workers
Anthropic (2025) How conversations with one assistant are used Classified anonymised transcripts, mapped to O*NET tasks 49.1% automation on Claude.ai

"Almost 40%" and "one in four" are not a disagreement to settle by picking the bigger number or averaging them. They use different definitions of exposure, different data and different years. The Anthropic percentages are shares of conversations, not shares of workers.

If you quote one, quote it with its method. That is how this post handles each figure.

Will AI replace jobs in office work first?

Both institutions point to office and knowledge work rather than manual work. The ILO names clerical occupations as the most exposed. The IMF says advanced economies are more exposed because their employment is concentrated in cognitive-intensive roles. Anthropic's API sample leans towards computer and mathematical tasks.

That does not mean those workers lose their jobs. The ILO's finding that most occupations need human input for some tasks is the reason it expects transformation. The practical reading: look at the tasks inside a role, not the job title.

AI's physical and energy footprint is a separate part of the economy story, covered in the post on data centre electricity and the IEA.

Going deeper: for security and IT teams

The labour research measures tasks. For a security team, the useful move is to treat each task handed to an AI tool as a data flow with an owner. Then govern it with the same controls you use for any other system. None of the sources above studies security roles, so the task examples here are an interpretation, not a finding.

Map the work to the NIST AI RMF. The AI RMF Playbook gives the anchors. GOVERN 1.6 asks for "mechanisms ... to inventory AI systems", resourced by risk priority. GOVERN 6.1 asks for policies covering "AI risks associated with third-party entities". That covers almost every assistant staff use at work. GOVERN 4.3 asks for practices that enable "AI testing, identification of incidents, and information sharing." On the Manage side, MANAGE 3.1 asks that third-party AI risks be "regularly monitored", and MANAGE 4.1 asks for post-deployment monitoring plans covering user input, "appeal and override", incident response and change management. MANAGE 2.4 asks for assigned responsibility to "supersede, disengage, or deactivate" a system that behaves outside its intended use. NIST's generative AI profile, NIST AI 600-1, adds GenAI-specific actions. Its PDF was not read for this post, so none of its action IDs are cited here.

Build the inventory from tasks. For each AI-assisted task, record five things: the tool, the account type (enterprise tenant or personal), the data classes allowed in, where the output goes, and the named human who signs off. A task that drafts a customer reply is a different risk from one that closes a ticket on its own.

Data-handling settings to check per tool. Is sign-in through your identity provider, so access ends when someone leaves? Are prompts used for model training, and who in your tenant controls that? How long does the vendor keep prompts and files, and can admins export them for an investigation? Document the answers from the vendor's own admin documentation, not a sales deck.

Logging and detection signals. - DNS or proxy traffic to generative AI services from accounts outside your enterprise tenant. - New OAuth consent grants that give an AI app access to mail, files or calendars. - DLP events where labelled or classified content is pasted or uploaded to an AI domain. - API keys for AI services used from unexpected hosts, or at volumes no workflow explains. - Actions by an agent's service account that no ticket or change record accounts for.

Test each signal at least once. Paste a harmless canary string carrying a sensitive label into an unapproved tool, and confirm an alert reaches a person. For finding the tools in the first place, read how to discover shadow AI at work.

In practice: a regulator finding on one worker's prompt. In September 2024, Victoria's privacy regulator, OVIC, published its finding that the state's Department of Families, Fairness and Housing breached Information Privacy Principles 3.1 and 4.1. A child protection worker had used ChatGPT while drafting a report for the Children's Court. The worker entered "a significant amount of personal and delicate information", including names and risk-assessment details, and the generated text downplayed risks to the child. OVIC's compliance notice set six actions. One requires the department to block "ChatGPT and other similar tools" for child protection workers. A single task, drafting, created both an accuracy failure and a disclosure.

In practice: how often unsanctioned AI is involved. IBM's 2025 Cost of a Data Breach release quotes the headline directly: "One in five organizations reported a breach due to shadow AI." Fewer than four in ten of the breached firms (37%) had any policy to govern AI use or spot unapproved tools. Where unsanctioned use was heavy, the average breach cost USD 670,000 more than where it was low or absent. The sample is breached organisations only, so read it as a picture of incidents, not of every workplace.

Questions to ask vendors and your own team. - Which of our tasks does this tool complete on its own, and which does it only assist? - Can we export prompt and action logs into our SIEM, and how long are they kept? - Who can switch the tool off, and how fast? - Where the EU AI Act applies, how do we evidence its AI literacy duty? The EU AI Act timeline covers what applies when, and ISO/IEC 42001 explained covers the management-system option.

What can you do this quarter?

You cannot predict the labour market. You can make your team's position clearer.

  1. List the tasks, not the titles. For each role, write five to ten recurring tasks. Mark each as AI-assisted, AI-automated or human-only today.
  2. Pick one low-risk task to assist. Summarising a long ticket thread is a reasonable first one. Measure time saved and errors found.
  3. Teach checking. A short session on how to verify AI output beats a ban that people route around.
  4. Write the rule for pasting data. Start from the checklist of what not to paste into AI at work.
  5. Revisit in six months. The sources here are dated 2024 and 2025. Check for newer editions before quoting them.

Learn more

Key takeaways

  • Exposure means a job contains tasks AI can do. It is not job loss.
  • The IMF says almost 40% of global employment is exposed, about 60% in advanced economies and 26% in low-income countries. It says roughly half of exposed jobs in advanced economies may benefit.
  • The ILO says one in four workers are in an occupation with some exposure, and that transformation is the most likely impact.
  • Anthropic's usage data found 49.1% automation and 47% augmentation on Claude.ai, and 77% automation patterns in its API sample. It is one vendor's data on conversations.
  • Never merge figures from different methods. Quote each with its source and method.
  • Look at tasks inside roles. For security and IT teams, govern the tools and build the skill to check AI output.

Sources