Will AI Replace Your Job? The Careers Most Likely to Change by 2030

A job title is a poor description of a working day. An accountant may spend the morning checking a spreadsheet, the afternoon discussing an ambiguous transaction and the evening explaining a decision to a client. A teacher prepares material, manages a room, notices confusion and builds trust. A software developer writes code, but also decides what the software should do and what must never go wrong.

AI can affect those activities differently. That is why lists of jobs supposedly destined to disappear can be misleading, even when they begin with legitimate research. The ability to automate a task is not evidence that an employer can remove the entire role, still less proof that it will do so by a particular date.

The period to 2030 is best treated as a planning horizon. Workers and employers need to identify which responsibilities may change, where human judgment remains necessary and how the benefits and costs of a transition will be distributed. The answer will vary between organisations, countries and people doing apparently similar jobs.

Key findings

• Occupational exposure measures potential interaction with AI, not a count of jobs already lost.

• Digitised, repeatable tasks are easier to place inside an AI-assisted workflow than duties requiring physical presence or responsibility for unusual situations.

• Productivity gains can change staffing, workloads, service quality or demand; they do not imply one inevitable employment outcome.

• Entry-level learning pathways deserve attention when routine work is automated.

• Domain knowledge, verification, communication and the ability to redesign a process remain useful complements to technical skills.

Automation and job transformation are different outcomes

Four terms need separating. Task automation means software performs an activity previously done by a person. Occupational exposure means some activities within a job match a technology's potential capabilities. Productivity improvement means more or better output relative to the inputs used. Job displacement means a worker or position is actually removed or replaced. Each needs different evidence.

The International Labour Organization's May 2025 exposure index estimated that one in four workers worldwide was in an occupation with some exposure to generative AI. It identified clerical work as especially exposed and emphasised transformation because occupations include tasks requiring human input. The estimate does not mean one in four workers will lose their jobs.

A separate IMF analysis published in January 2024 examined broader AI exposure and complementarity. Differences between its estimates and the ILO's should not be read as a simple disagreement about redundancies: the definitions, technologies and methods are different. Comparing percentages without comparing what they measure produces an unreliable picture.

Which tasks are most exposed?

A useful starting point is whether the input and output are already digital, whether success can be checked and whether the task repeats often enough to justify integration. Drafting a standard response and extracting information from a document are different propositions from negotiating a disputed agreement or repairing unfamiliar equipment on site.

The 2023 paper GPTs are GPTs studied the alignment between US occupational tasks and language-model capabilities. Its authors explicitly did not predict an adoption timetable. That distinction remains important: technical feasibility must pass through costs, organisational choices and practical constraints before becoming a change in employment.

For the reader, the following occupational sections are a task-based analysis, not a ranked forecast of job losses. Ask whether your employer has usable data, a reliable evaluation process and a reason to change the workflow. A capability shown in a demonstration does not answer those questions on its own.

Administrative and office roles

Administrative work often contains repeatable activities: organising information, drafting correspondence, preparing schedules and checking whether a form is complete. In a bounded system, those are plausible targets for assistance. The harder work may be identifying what an unusual request really means and deciding who should handle it.

Consider an illustrative office that automates the first pass over incoming applications. Employees might then handle exceptions, help applicants correct omissions or supervise a larger volume of cases. Alternatively, management might reduce staffing. The technology creates options; demand, budgets and managerial decisions help determine which option is chosen.

Workers can prepare by understanding the whole process rather than only its most repetitive step. Knowing why information is collected, what an exception looks like and how an error affects the next department is useful when the interface changes. Employers should avoid discarding that knowledge while pursuing a faster first pass.

Customer service and sales operations

Customer support is an area where evidence extends beyond hypothetical exposure. Generative AI at Work, in its November 2024 revision, examined 5,172 support agents and reported an average 15% improvement in issues resolved per hour. Effects differed across workers. This was a particular intervention in a particular setting, not a universal forecast for call centres or sales teams.

An assistant can suggest a response, retrieve approved guidance or help an agent navigate a product. A fully automated service must also recognise when it lacks the answer and hand the conversation to someone who can act. Closing a ticket is not the same as resolving a customer's problem.

A useful measure therefore combines speed with repeat contacts, escalation quality and actual resolution. For sales, similar care is needed: producing more messages is not equivalent to building a relationship or winning appropriate business. The durable human contribution may shift towards difficult conversations, judgment and accountability.

Software development and IT

Code is a visible output, but software engineering includes requirements, architecture, testing, security, deployment and maintenance. Generating a function quickly can help a developer; it does not establish that the function belongs in a production system or that its dependencies are safe.

The US Bureau of Labor Statistics occupational profile describes responsibilities beyond writing code, including analysing user needs and designing, testing and maintaining systems. Its occupational outlook is a US labour-market projection, not a verdict on every programming role in every country.

Our practical inference is that workers should become better at checking and integrating generated code rather than measuring their value only by how many lines they produce. Entry-level development also needs a training plan. If simple assignments are automated, employers must deliberately create other opportunities for new staff to learn debugging, review and system design.

Finance, accounting and legal work

These professions combine document-heavy tasks with obligations to interpret evidence and take responsibility for decisions. A tool may help prepare a reconciliation or locate potentially relevant material, while a qualified person still needs to investigate discrepancies, assess context and justify the conclusion.

An illustrative accounting workflow makes the distinction clear. Software could suggest a classification for a transaction, but a missing contract or unusual commercial arrangement may change the correct treatment. The speed of the suggestion is not evidence that the underlying facts are complete. Similar reasoning applies when a legal answer depends on the actual jurisdiction and current authority.

The risk is not only replacement. A professional may become responsible for reviewing a much larger volume of machine-produced work without enough time or support. Good deployment should define what must be checked, who can approve it and when the system should stop. This discussion describes work organisation, not professional financial or legal advice.

Healthcare and education

Healthcare and education involve information processing, but they also involve relationships, physical settings and responsibilities that cannot be reduced to a document. The relevant question is which administrative or analytical activities can be improved without weakening the quality of care or learning.

For example, a system might help organise material for a clinician or generate practice questions for a teacher to review. These are possible workflows, not endorsements of any product's safety or educational effectiveness. In each case, someone must judge whether the output fits the individual and the setting.

Time saved could be used for more direct attention, a larger workload or lower staffing. Those are different organisational choices. Patients, students and staff have an interest in knowing which choice is being made. A deployment that reports only faster paperwork may miss whether the overall service became more useful, accessible or humane.

Which jobs may be more resistant to automation?

Tasks involving varied physical environments, dexterity, interpersonal trust and responsibility for unexpected situations can be harder to automate as a complete package. This is a task-based inference, not a declaration that trades, care or leadership are immune to change. Scheduling, documentation and information gathering within those roles may still be affected.

A technician repairing unfamiliar equipment may use AI to search manuals while retaining responsibility for diagnosis and safe action. A care worker may use scheduling assistance while continuing the interpersonal and physical work of care. These examples show why partial assistance and full replacement should not be confused.

Resistance is also economic. Even if a system can perform an activity, it may be too costly or difficult to maintain for a particular employer. Conversely, a constrained environment may make automation practical earlier than the same task in an unpredictable setting. The conditions around the job matter as much as its title.

New roles and opportunities by 2030

The World Economic Forum's Future of Jobs Report 2025 brings together employer expectations about the period to 2030. Its widely cited projections of 170 million roles created and 92 million displaced concern multiple structural trends, including technology, demographics and the green transition. They are not AI-only totals or guaranteed outcomes.

New responsibilities may include evaluating systems, maintaining approved information, monitoring failures and helping teams redesign work. Some will become distinct jobs; others will become part of existing roles. An organisation does not necessarily need an entirely new job title to take those responsibilities seriously.

Workers should be wary of treating fashionable titles as a secure career plan. A more durable approach is to demonstrate a combination of domain knowledge and measurable capability: identify a real problem, use appropriate tools, verify the outcome and explain the limitations. That combination remains valuable even when the software interface changes.

Skills workers can develop

Build a task inventory

Write down what you actually do over a representative week. Separate recurring information tasks, judgment calls, physical responsibilities and interpersonal work. Note which activities depend on confidential information or require approval. This produces a more useful learning plan than asking whether an entire profession is safe.

Practise verification and communication

For an approved, low-risk task, compare AI-assisted work with your usual process. Check factual claims, calculations and source material. Record review time and errors as well as apparent speed. Then explain the result clearly enough that a colleague can reproduce or challenge it. An unverified time-saving claim is a weak basis for changing a workflow.

Learn the surrounding system

Understanding customers, regulations, handoffs and failure consequences helps a worker recognise where automation belongs. The OECD Employment Outlook 2023 examines both opportunities and risks for job quality and the importance of skills and workplace institutions. Training should address how work is organised, not simply how to operate a new tool.

What employers and policymakers should consider

A pilot should specify what success means before deployment. Include quality, safety, workload and employee experience alongside speed. Give staff a route to report failures without being penalised for contradicting a positive automation narrative. Where people remain accountable for an output, they need enough authority and time to review it meaningfully.

An ILO review published in June 2026 describes emerging productivity effects as uneven and notes that large-scale displacement remains limited in the evidence it reviews. It also highlights risks around younger workers and work organisation. That is a dated assessment of available evidence, not reassurance that future disruption cannot occur.

Policy choices concern more than training courses. They include whether workers can move between opportunities, whether smaller employers can adopt useful technology, and whether people have support during transitions. National averages can hide difficult local changes. A region losing entry-level opportunities may face a serious problem even while aggregate employment elsewhere grows.

Adaptation is a plan, not a promise of safety

No responsible analysis can identify every job that will exist in 2030 or promise that learning a particular tool will prevent displacement. Employers can reorganise work badly, promising technologies can disappoint and gains can be distributed unevenly.

Workers can still act on a clearer question: which parts of my contribution are becoming easier to automate, and which responsibilities must someone continue to own? Employers should ask the matching question: if we automate an activity, who will understand, verify and improve the process that remains?

A credible future-of-work strategy connects those questions. It treats employees as people with knowledge and agency, measures changes in actual work and makes room for learning before removing the tasks through which that learning used to happen.