What Jobs Will AI Replace? Exposure and Job Forecasts

Current evidence cannot reliably identify which specific jobs AI will eliminate. It can show which occupations have tasks that overlap with AI capabilities and how employment in those occupations is projected to change. Those are different measures. For example, the U.S. Bureau of Labor Statistics (BLS) projects computer programmers to decline 7% from 2025 to 2035, while its separate group for software developers, quality assurance analysts, and testers is projected to grow 10%. Neither figure proves that AI caused the change or predicts an individual worker’s prospects.

The practical question is not whether a job title is safe. It is which tasks make up the role, what the available evidence measures, and where AI assistance can be tested while a person remains responsible for consequential decisions. This guide compares BLS relative AI exposure with U.S. employment projections and gives you a task-level way to assess your own work.

The figures below are tied to their source, geography, occupation definition, and forecast period. BLS projections cover 2025 to 2035. The World Economic Forum (WEF) figures discussed later are global estimates through 2030 and reflect broader labor-market change.

What jobs will AI replace? The short, evidence-based answer

There is no definitive, evidence-based list of occupations that AI will replace. A job title groups together many tasks, and automating or assisting with some of those tasks can change a role without eliminating the occupation. Current evidence is more useful for identifying task overlap and projected employment direction than for proving future job losses.

BLS publishes relative AI-exposure categories and occupation-level employment projections. These answer different questions. Exposure concerns the relationship between occupational tasks and AI capabilities. An employment projection estimates how the number of jobs in an occupation may change over a stated period. Neither is a personal layoff forecast.

Exposure is not the same as replacement

BLS classifies occupations as having Low, Moderate, High, or Very high relative AI exposure. The categories describe the degree to which occupational task content overlaps with AI capabilities and observed AI use. BLS’s methodology draws on mapped research and external data, including analyses involving Anthropic’s Claude and Microsoft Copilot.

The categories do not measure the probability that AI will automate an occupation, that a worker will be laid off, employment growth, wages, productivity, or the prospects of a particular person. BLS also cautions that low exposure does not mean an occupation is unaffected. High exposure does not mean employment must decline.

BLS does not directly measure how much workers in each occupation personally use AI. The classification is a relative, occupation-level measure based on its methodology, source data, occupational crosswalks, and limitations. Read the BLS explanation of AI exposure categories for the definitions and method.

Read exposure alongside employment projections

BLS employment projections estimate how occupational employment may change over a stated period. They are not causal estimates of AI replacement. The selected examples below use the official BLS 2025 to 2035 percentage projections.

Occupation	BLS 2025-35 projection	What the figure can tell you
Word processors and typists	-34.4%	A steep projected decline, not an AI-caused job-loss estimate.
Data entry keyers	-25.5%	Projected occupational employment change.
Telephone operators	-27.6%	A separate occupation from switchboard operators.
Switchboard operators	-26.0%	A separate occupation from telephone operators.
Bookkeeping, accounting, and auditing clerks	-6%	BLS identifies software automation as one factor changing demand and duties.

A declining occupation can still have annual openings as workers retire, transfer, or leave the labor force. For example, the BLS profile for bookkeeping, accounting, and auditing clerks discusses both projected decline and continuing openings. A decline does not mean there will be no opportunities to enter or work in that occupation.

The useful two-axis reading is simple: record the exposure category separately from the employment projection. High exposure with a declining projection indicates more occupational pressure than either measure alone. High exposure with growing employment may indicate task transformation rather than simple replacement. A declining projection with low exposure is not automatically reassuring, because factors unrelated to AI may be reducing demand.

Why job titles can mislead: programmers and developers

“Programming jobs” is too broad a label for comparing forecasts. BLS projects computer programmers to decline 7% from 2025 to 2035. Its separate group for software developers, quality assurance analysts, and testers is projected to grow 10% over the same period.

These are different occupational groups with different definitions and duties. The developer figure cannot be used as a forecast for computer programmers, and the programmer figure cannot be used as a forecast for all software work. Before applying a headline to your role, check the official occupation title, SOC definition if known, duties, and projection period.

What the projections say about other changing roles

Projected growth is not proof that AI has no effect on a role. Projected decline is not proof that AI is the cause. The more useful question is which repeatable outputs can be assisted and which decisions still require context, accountability, stakeholder trust, or judgment.

BLS projects graphic designers and receptionists each to decline 2% from 2025 to 2035. It projects market research analysts and marketing specialists to grow 7%, and meeting, convention, and event planners to grow 6%. Writers and authors are projected to show little or no employment change. BLS also says increased use of AI for writing is expected to dampen demand for writers, while other sources of demand continue to support some jobs.

These figures do not make any role safe or immune. A receptionist may use software for scheduling while handling exceptions and visitors. A market research analyst may use AI to organize comments while deciding whether themes fit the sample. A writer may use AI for a draft while retaining responsibility for accuracy, voice, and claims. The task mix matters more than a simple safe or unsafe label.

Audit your tasks before changing your career plan

Start with a task inventory, not a guess based on your title. For each recurring responsibility, record its frequency, inputs, expected output, exception handling, sensitivity, physical requirements, and whether the task affects customers, employees, finances, safety, compliance, or public claims.

Next, match the work to an official occupation definition and record the BLS projection period, metric, exposure source, methodology date, and date checked. Do not ask an AI system to invent an official BLS rating. An AI classification of your task descriptions can be advisory, but a person must verify the occupation match and evidence.

A useful record can include occupation title, SOC code if known, task, frequency, proposed AI role, human owner, evidence references, source and revision dates, reviewer, observed quality, and correction rate. Use allowed values such as assist, do not automate, or needs review to keep the record consistent.

If concurrent systems store these records, do not rely on lookup before create as the only duplicate control. Use a database-enforced unique key or transactional upsert. The key must match the row grain. For example, an occupation projection can be unique by source organization, SOC code, projection base year, projection end year, and metric. An exposure record needs the methodology version and classification date. A survey statistic needs its survey wave, population definition, and task question. These are proposed recordkeeping patterns, not documented features of BLS or any vendor.

If you are considering designing AI agents around bounded tasks and human review, use the audit to define where assistance ends, which checks are deterministic, and who approves the outcome.

A practical example: AI-assisted market-research analysis

Consider a hypothetical analyst organizing open-ended survey comments for a report. The source set contains comments, respondent segment identifiers, the survey question, and sampling notes. AI may propose themes and return the comment IDs supporting each theme. It should not invent percentages, infer representativeness, or publish a recommendation without analyst approval.

A proposed process sequence is:

• Input: a defined survey wave, source comment IDs, segment metadata, and sampling notes.
• AI output: theme labels, supporting comment IDs, uncertainty notes, and a pending review status.
• Deterministic validation: confirm that every cited ID exists, detect duplicate IDs, check that theme labels are present, and block unsupported numerical claims.
• Human decision: the analyst checks whether the themes fit the sample and approves any finding or recommendation.
• Destination: an analysis workspace or draft report that retains the survey wave, source set, run identifier, reviewer, and approval status.

The output row grain matters. A theme result should not be uniquely identified only by the survey wave because the same source may be analyzed more than once. A proposed key could include source_comment_set_id, analysis_run_id, survey_wave_id, and theme_label. If the system stores citations separately, a citation record can use the analysis run, theme identifier, and comment ID. This prevents a later run from overwriting an earlier run or merging raw observations with reported summaries.

{
  "source_comment_set_id": "ILLUSTRATIVE-SET-01",
  "analysis_run_id": "ILLUSTRATIVE-RUN-01",
  "survey_wave_id": "ILLUSTRATIVE-WAVE-01",
  "themes": [
    {
      "theme_id": "THEME-001",
      "theme_label": "Delivery updates",
      "supporting_comment_ids": [
        "C-001",
        "C-014"
      ],
      "uncertainty_note": "Check segment context before reporting"
    }
  ],
  "review_status": "pending_human_review"
}

This is an illustrative editorial workflow, not a documented BLS, Census, HubSpot, or vendor integration. Its value is the separation between organizing evidence and deciding what the evidence means.

What global forecasts do, and do not, say

The WEF estimates that 170 million jobs could be created and 92 million displaced globally by 2030, for a net increase of 78 million in its outlook. These estimates combine employer expectations with ILO employment data and reflect broader labor-market transformation. They are not an AI-only count, and displaced roles do not mean every current jobholder will be laid off. See the WEF’s 2025 jobs outlook for its scope and method.

Keep the units attached to every forecast: geography, time period, population, occupation definition, and metric. The WEF’s global estimates through 2030 and BLS’s U.S. occupation-specific projections through 2035 answer different questions and should not be combined into a single replacement number.

A grounded takeaway for career planning

Use current evidence to investigate how work may change, not to declare a profession safe or doomed. Match your role to an official occupation definition, compare exposure and employment direction separately, and identify tasks that can be tested with a checkable output.

Skills in domain knowledge, communication, reviewing AI outputs, judgment, and accountability may help people work effectively as tasks change, but no source here guarantees an employment outcome. The cited sources support task-level change and varied occupational outlooks, not a definitive list of jobs AI will eliminate by 2030 or 2050.

For a practical next step, choose one low-consequence, verifiable task for a supervised trial. Track corrections and quality, keep ownership of consequential decisions with a person, and use the result to decide whether to expand, revise, or stop.