The clearest global finding
The International Labour Organization’s refined global analysis still identifies clerical occupations as the most exposed to generative AI. Yet its overall conclusion is not “one in four jobs will vanish.” It says that, because jobs combine many tasks requiring human input, transformation is the most likely outcome for most occupations. [1]
Why “which jobs will AI replace?” gets the question backwards
Job titles hide too much. “Marketing assistant,” “paralegal,” “customer support worker,” “accountant,” and “software engineer” can each describe a huge range of work. One person might spend most of their day moving information between systems or producing standard first drafts. Another might deal with difficult clients, catch errors no one else notices, negotiate priorities, or own the outcome when a decision goes wrong.
AI is usually deployed against a workflow, not a job title. Employers ask whether an activity can be completed more cheaply or quickly with software, whether the inputs are available, whether the result can be checked, and what happens if the output is wrong. A company may automate a routine step while keeping the role. Or it may reduce entry-level hiring because a smaller, more experienced team can now produce the same volume of routine output.
This distinction matters particularly if you are young. The first tasks in a career are often the most structured: research, drafting, data cleaning, formatting, basic coding, standard customer queries, and summary work. Those tasks were also a way to learn how a field operates. Stanford’s 2026 evidence review notes that recent-graduate employment has been weak and that AI may be part of the story in exposed fields, while also stressing that interest rates, pandemic-era hiring and remote-work changes make clean causal claims difficult. [2]
The six task patterns AI can most easily absorb
Use this table to inspect a path you are considering. It is not an exposure score for a career; it is a way to look inside the work.
| Work pattern | Why it is more exposed | What often remains human |
|---|---|---|
| Standard digital input → standard digital output | High exposure AI can read, classify, draft, translate, or summarise when the relevant information is digital and the response has a familiar format. | Deciding whether the output is correct, appropriate, fair, safe, or suited to the real situation. |
| High-volume repetitive processing | High exposure Large volumes make automation easier to justify, especially when each case follows a similar path. | Handling exceptions, improving the process, resolving edge cases, and taking responsibility for errors. |
| Rule-based administration | High exposure Fixed policies, forms, checklists, and known decision trees can be partly automated or routed. | Interpreting ambiguous cases, applying discretion, and communicating decisions to people. |
| First-draft text or code | Meaningful exposure AI is increasingly capable at rough drafts, boilerplate, templates, standard copy, and code suggestions. | Original thinking, system design, fact checking, editing, taste, user needs, security, and ownership of the final result. |
| Scripted first-line interactions | Meaningful exposure Chatbots and voice systems can handle straightforward questions and basic troubleshooting. | Escalations, emotionally charged conversations, negotiation, retention, unusual problems, and trust. |
| Routine analysis from clean data | Meaningful exposure When the data is available and the task has a repeatable method, AI can accelerate the first pass. | Defining the question, spotting bad data, interpreting context, choosing trade-offs, and making a defensible decision. |
Work most exposed does not equal work that will disappear
That warning can sound like a technicality, but it changes the answer. A call centre may use AI to handle routine questions, then employ fewer people on first-line scripts and more people on difficult cases, complaints, retention, quality assurance, and customer insight. A legal team may generate a first draft of a standard document faster, but still need qualified people to advise a client, assess risk, negotiate, and accept professional responsibility. A design team may produce rough concepts in minutes but need stronger judgment about a brand, an audience, accessibility, and whether the idea is actually good.
Economists also care about demand. If AI makes a service cheaper or faster, more people may use it. In that case, a productivity gain can create more work rather than simply reducing headcount. BCG’s 2026 model separates automation potential from substitution and from whether demand for the output can expand. It estimates widespread role redesign in the United States, but it explicitly says its model is not an unemployment forecast. [3]
That is why you should be suspicious of both extremes: “AI will replace every office job next year” and “AI cannot touch my career.” Both avoid the useful part of the question: which tasks are changing, and what higher-value work becomes more important as they do?
Areas to watch closely
Some broad work patterns deserve closer attention because they contain a larger share of structured, digital, or routine tasks. This is not a reason to panic or write them off. It is a reason to examine the role, the training route, and the ladder above the entry level.
Clerical and administrative processing
Data entry, scheduling, form handling, basic record work, document sorting, and routine correspondence are among the clearest places for automation. If this is your route in, look for a path into coordination, client responsibility, operations improvement, or specialist knowledge.
Standard customer support
Basic queries are a natural target for chat and voice systems. Human value tends to move toward complex resolution, relationship management, quality, escalation, and designing better customer experiences.
Routine content production
Generic drafts, simple product copy, basic social posts, summaries, and templated writing can be generated quickly. Research, editorial judgment, audience insight, strategy, reporting, and a distinctive voice matter more—not less.
Entry-level digital production
Some junior coding, design, analysis, and research tasks can now be accelerated substantially. That can make the first rung harder, but it also raises the value of portfolios, practical work, systems thinking, and the ability to review AI output.
Standardised financial or legal support
Routine reconciliation, first-pass document review, standard contracts, and basic research are increasingly assisted by AI. The more contextual, regulated, client-specific, or judgment-heavy work remains distinct.
Transcription, straightforward translation, and simple summaries
These are classic examples of useful AI assistance. The human work shifts to accuracy checking, sensitive context, specialised terminology, confidentiality, and the consequences of mistakes.
If the career you want is exposed, what should you do?
Do not turn exposure into a verdict on your ability or future. A career can be worth pursuing even when AI changes it quickly. The stronger question is whether the course, employer, or apprenticeship gives you a path beyond the most routine output.
Ask what the first two years of work look like. Which tasks are usually given to beginners? Which of those tasks are now being automated, accelerated, or checked by AI?
Look for practice, not just credentials. Placements, apprenticeships, portfolios, clinical work, labs, student publications, real clients, and team projects help you develop judgment that is hard to learn from a generic prompt.
Learn to supervise the tool. Knowing how to ask AI for a first pass is useful. Knowing when it is wrong, biased, insecure, incomplete, or unsuitable is more valuable.
Choose an adjacent capability. Pair a digital skill with domain depth, communication, customer understanding, technical systems, data judgment, or operational knowledge. That gives you more routes if the entry point changes.
AI exposure is not a personal failure and it is not a prophecy. It is a prompt to look more closely at the work. The people who understand a field, make sound judgments, and can use new tools without becoming dependent on them are likely to have more options than people who only repeat a routine workflow.
Explore the question in more detail
Frequently asked questions
References
- International Labour Organization, Generative AI and Jobs: A Refined Global Index of Occupational Exposure (2025).
- Stanford Institute for Economic Policy Research, What is really happening to jobs? Separating AI hype from reality (2026).
- Boston Consulting Group, AI Will Reshape More Jobs Than It Replaces (2026). This is a US-focused scenario model, not an unemployment forecast.
Aifyx provides general career information, not personal education, financial, employment, or professional advice. AI exposure is not a prediction that a particular role will disappear. Labour markets, training routes, and individual circumstances vary by country and change over time.