Which Jobs Are Most at Risk From AI? Tasks, Skills and Career Strategies

The short answer: AI changes tasks before it eliminates job titles

Artificial intelligence is already changing how people draft text, write code, classify information, support customers and analyze documents. But a job is a bundle of tasks, not a single activity. When software automates one task, the role may shrink, expand or be redesigned rather than disappear entirely. A credible career guide must distinguish exposure to AI from actual job losses.

The International Labour Organization reported in its 2025 update that roughly one in four workers worldwide holds an occupation with some exposure to generative AI, while most affected jobs are more likely to be transformed than eliminated. That is a measure of potential task-level impact, not a forecast that one quarter of workers will lose employment.

What does “at risk from AI” actually mean?

Three different questions are often confused: Can AI technically perform a task? Will an employer deploy it reliably and economically? Will that deployment reduce headcount? The answers depend on accuracy, regulation, customer expectations, security, integration costs and whether new demand offsets productivity gains.

A role with high exposure may gain productivity and become more valuable. A role with low exposure may still face pressure from other technologies or economic shifts. The sensible approach is to evaluate workflows, not sensational rankings of professions supposedly doomed to vanish.

Administrative and clerical work: high task exposure

Routine scheduling, document formatting, data entry, invoice matching, basic transcription and standard correspondence are comparatively easy to describe as repeatable digital workflows. Clerical occupations ranked among the most exposed groups in the ILO assessment, particularly where work follows consistent rules and digital records are available.

That does not mean all administrative assistants will disappear. Handling exceptions, sensitive information, frustrated customers, ambiguous instructions and cross-team coordination remains valuable. Workers can build resilience by learning workflow automation, quality checks, records management and secure use of AI tools.

Customer support: automation of simple requests, escalation of hard cases

AI chatbots and voice systems can answer frequent questions, summarize interactions, suggest replies and route tickets. Businesses may automate password resets, shipment status and straightforward policy explanations while reserving human staff for disputes, vulnerable customers and high-stakes decisions.

Customer service workers can move toward complex case resolution, knowledge-base design, fraud detection, multilingual quality assurance and customer-success analysis. The risk is greater where an organization measures the role primarily by the volume of repetitive interactions.

Writing, translation, marketing and media production

Generative systems can produce first drafts, translate ordinary text, suggest headlines, resize creative assets and generate variations of advertisements. The ILO notes that improved image, voice and video generation has expanded exposure in media and web-related occupations.

The differentiator becomes judgment: verifying facts, interviewing original sources, understanding audiences, avoiding copyright problems and evaluating campaign outcomes. A generic AI-written article is easy to replicate; a well-researched original investigation or tested tutorial is harder to replace.

Software development: the task mix is shifting

AI coding assistants can generate boilerplate, explain unfamiliar code, propose tests and speed up routine debugging. Yet software engineering also requires requirements discovery, architecture, security reviews, production operations, reliability and decisions about trade-offs.

Entry-level tasks may change because employers expect candidates to use AI effectively. A strong developer portfolio should show the ability to validate generated code, reason about failures, write meaningful tests and ship secure, maintainable systems—not merely prompt a chatbot to produce an app.

Accounting, finance and legal services

Classification, reconciliation, research summaries and document review can be assisted by AI. However, legal advice, financial reporting and regulated decisions carry accountability requirements that cannot be outsourced to a model without governance and professional oversight.

For accountants and paralegals, promising skills include audit trails, exception handling, evidence review, privacy protection and understanding the limits of automated analysis. Human responsibility becomes more visible when the easy drafting work is automated.

Healthcare, education and other high-stakes fields

AI may assist with medical documentation, appointment logistics, lesson planning and administrative workflows, but that does not establish that it can safely replace clinicians or teachers. Patient care and education involve contextual judgment, trust, safeguarding and responsibility for outcomes.

Roles that blend technical tools with direct human interaction can change substantially without disappearing. Training in AI literacy, documentation, error detection and professional ethics is useful, but organizations should not treat unverified model outputs as authoritative medical or educational decisions.

Jobs involving physical work and unpredictable environments

Electricians, plumbers, many care workers, maintenance technicians and field-service specialists operate in variable physical environments. A language model cannot by itself climb a ladder, inspect hidden wiring or safely handle a patient. Robotics may eventually automate some tasks, but cost, dexterity and safety create different adoption barriers.

Physical work is not immune to automation: scheduling, diagnosis, inspection and inventory tasks can still change. A technician who can use digital diagnostics and interpret sensor data may become more effective.

Which workers are most vulnerable within the same occupation?

Risk depends on the share of routine tasks, the employer’s ability to standardize processes, access to clean digital data and the cost of mistakes. Two employees with the same job title may have very different exposure if one handles repetitive forms and the other resolves complex exceptions.

The highest practical vulnerability often comes from combining narrow responsibilities with limited opportunities to learn new systems. The solution is not panic but expanding the range of problems you can solve.

Why AI exposure is not the same as unemployment

Productivity improvements can reduce labor hours per task, but they can also lower prices, increase demand and create new services. The final employment effect varies by industry and time horizon. Evidence about what a model can do in a benchmark is not direct evidence about what companies will deploy or how many workers they will hire.

Public forecasts frequently use different definitions, assumptions and scenarios. Compare reports carefully and avoid treating a single percentage as a guaranteed prediction. The ILO explicitly distinguishes transformation potential from full job redundancy.

A practical self-assessment: map your work into tasks

Write down everything you do in a typical week, then group tasks into repetitive digital work, judgment-heavy work, relationship-based work and physical work. For each task ask whether AI can draft an output, whether a human must verify it, and what happens if the result is wrong.

Identify two repetitive activities that could be safely accelerated, and two judgment-based activities you can deepen. This produces a much more useful career plan than searching for a universal list of “AI-proof jobs.”

Five capabilities that improve career resilience

Develop domain expertise so you can detect errors and understand what matters. Learn data literacy to evaluate evidence. Build communication skills for ambiguous, high-stakes situations. Practice workflow design to combine tools safely. Strengthen cybersecurity and privacy habits so productivity gains do not create new risks.

A sixth useful habit is documenting outcomes: time saved, quality improved, customer problems solved and errors prevented. Employers are more likely to value demonstrated results than a long list of AI tools on a résumé.

How students and recent graduates can prepare

Choose a real project from your field and complete it with and without AI assistance. Record where the tool helped, where it failed and how you checked its output. Publish a portfolio that includes the problem statement, sources, constraints, tests and final result.

For job applications, learn the software and processes used in the target industry. A graduate who can evaluate AI output and communicate decisions may stand out more than someone who merely knows prompt templates. Internships, apprenticeships and supervised projects still matter.

How experienced professionals can adapt

Start with low-risk internal tasks such as meeting summaries or draft templates using approved tools. Measure quality and time, then create review checklists. Seek responsibility for exception handling, process improvement, training and customer outcomes rather than only repetitive production.

Ask employers for reskilling opportunities and clear policies about confidential data. Employees should know which tools are permitted, who owns outputs and when a human must approve an action.

What employers should do before replacing work with AI

Audit processes, data quality and error costs before automating. Involve workers who understand exceptions, measure real outcomes and preserve escalation routes. Assess bias, accessibility, privacy and security. A demonstration that looks impressive in a pilot does not guarantee reliable production performance.

Plan for redeployment and training, not just labor savings. If a tool creates more customer demand or higher output expectations, the organization may need different staffing rather than simply fewer staff.

Common myths about AI and employment

Myth: AI can replace every job it can assist. Reality: assistance, autonomous execution and legal accountability are different. Myth: only low-skill workers are affected. Reality: digitally intensive professional tasks can be exposed too. Myth: physical jobs are completely safe. Reality: planning, diagnostics and paperwork can still be automated.

Myth: learning one chatbot guarantees job security. Reality: durable value comes from expertise, verification, collaboration and continuous learning. Myth: every productivity gain means layoffs. Reality: employment effects depend on demand, adoption, regulation and organizational choices.

The outlook: scenarios, not certainty

Over the next several years, some businesses will automate repetitive workflows while others use AI to expand output without proportional hiring. Different countries and sectors will move at different speeds. Energy costs, regulation, trust and implementation failures may slow adoption; improved models and lower costs may accelerate it.

The most defensible forecast is that tasks and skill expectations will keep changing. Workers, educators and policymakers should prepare for multiple outcomes rather than betting on a single dramatic prediction.

Bottom line: build a career around outcomes, not routine outputs

If most of your value comes from producing standard text, forms or summaries, begin broadening your role now. If you already handle complex judgment, relationships and accountability, learn how AI can support those responsibilities while keeping humans in control.

The goal is not to compete with a machine at generating generic content faster. It is to become the person who understands the real problem, checks the evidence, manages risks and delivers a result people can trust.

Sources and further reading

ILO – Generative AI and jobs: A 2025 update

ILO – Refined global index of occupational exposure

World Economic Forum – Future of Jobs Report 2025

OECD – Artificial intelligence and work