What OECD Skills Outlook 2025 Says About AI Skills at Work

What does the 2025 Outlook actually answer about skills changing at work?

OECD Skills Outlook 2025 says that skills and opportunity at work should be examined through several kinds of evidence; it does not give workers a universal AI skills checklist or calculate an individual's chance of job loss. Its useful answer is at the population and market level: capabilities, advertised skill demand, employment context, firm reports and learning access each illuminate a different part of change. None alone tells you what will happen to your position. For your own decision, use the report to frame questions about changed tasks and the conditions for learning, then check those questions against your actual work and constraints.

That boundary matters because exposure and displacement describe different things. Exposure is a relationship between what tasks involve and what AI systems may be able to do; it indicates potential interaction or change, not an employment outcome. Displacement concerns whether work or workers are actually removed, and requires evidence about adoption, redesign, demand and staffing beyond technical capability. The Outlook contains AI-related material, but it is distributed across measures with different populations, time periods and purposes. Reading those signals as if they formed a single personal risk score would claim more than the report measures.

The sections that follow therefore keep each lens in its lane. First, the adult assessment describes broad capabilities measured among adults. Separately, the occupational chart combines changing skill demand with employment projections and context. An OECD vacancy analysis examines skills requested in AI-exposed occupations, while a survey asks SME respondents about perceived staffing needs. The Outlook also discusses how AI can substitute for some tasks while complementing others, and barriers to adult learning. These findings can guide what to investigate, but they are not interchangeable: a measured capability, a job advertisement, a firm response and a projected occupation trend answer different questions. The report's breadth is useful for locating the kind of evidence a question needs, not for collapsing unlike findings into a forecast. If the concern is whether a particular duty is changing, market-level findings can suggest what to inspect, while the worker's own task mix and employer's choices remain to be established locally.

Sources: OECD Skills Outlook 2025: Full Report

Which capabilities does the adult skills assessment measure?

The 2023 Survey of Adult Skills measures literacy, numeracy and adaptive problem solving. OECD Skills Outlook 2025 describes standardized, low-stakes, untimed assessments administered by trained interviewers in respondents' homes across participating countries. These are broad capabilities relevant to handling information, quantitative demands and unfamiliar problems. They are not a direct test of using generative AI, prompting a particular tool, or performing a specific occupation's tasks to its required standard.

That distinction determines what the results can help a worker decide. Population-level capability evidence can inform policy and identify broad learning needs; it cannot tell an individual which capability is missing from a changed workflow. A person deciding what to learn still needs to name a recurring work output, the step AI or another workplace change has affected, and the quality or safety standard the finished work must meet. A broad assessment may help explain why literacy, quantitative reasoning or adapting to new problems can matter, but it does not diagnose whether those foundations, a role-specific skill, tool fluency or access to feedback is the immediate constraint. It likewise cannot rank the value of alternative training routes, since that requires evidence about a target role, its entry requirements, and the time and resources a learner can commit. A measured skill level is context for policy, not an instruction to enroll in a credential.

The assessment also measures performance under its own survey conditions, rather than the full combination of knowledge, judgment, relationships, tools and accountability used at work. An employee might score well on a general capability and still need training in a regulated process or a company's verification rules. Conversely, a general assessment result cannot establish that the employee is unable to adapt to a new system. Neither conclusion follows from the measure because the assessment was designed to compare adult skills, not to evaluate a particular worker's job readiness or forecast what an employer will automate.

Foundational skills remain relevant precisely because new tools do not remove the need to understand inputs, interpret outputs and notice when an answer fails. But that general relevance is not evidence that everyone should take the same AI course. The practical sequence is narrower: identify the changed task and its standard, determine which capability is needed to complete or check it, and then choose practice or instruction that addresses that gap. This is an application of the assessment's scope, not a personal recommendation produced by the survey.

Sources: Widening opportunities by investing in 21st-century skills: OECD Skills Outlook 2025

What does the Skills Disruption Index combine—and what can it not establish?

The Skills Disruption Index in Figure 4.4 of OECD Skills Outlook 2025 is best read as a comparison of two different kinds of occupational change, with additional context—not as a measure of AI exposure or an estimate of anyone’s job risk. The chart pairs change in the skills requested in online vacancies between 2021 and 2024 with projected employment change for U.S. occupations between 2023 and 2033. Earnings and worker counts provide further context for interpreting the occupations. The inputs answer separate questions: what advertised skill demand has been shifting, what U.S. employment projections anticipate, and how large or remunerative an occupation is in the chart’s context.

The first axis comes from Lightcast online job advertisements and tracks changes in the skills employers request for occupations. A vacancy is evidence about what an employer chose to state in a posting; it does not inventory every duty performed by current employees. Some requirements may be implicit, omitted, or described differently across postings. Nor does a changing share of advertised skills by itself show that incumbent workers have adopted those skills, that their work has already been redesigned, or that AI caused the change. It is a labor-market signal about stated demand, useful for noticing movement in the advertised skill mix and asking what that movement could mean for a particular role.

The second axis is projected occupational employment change from the U.S. Bureau of Labor Statistics for 2023–2033. It is a forward-looking U.S. projection, not a record of jobs already added or removed. Its geography matters: the projection may help a reader outside the United States understand one comparison used by the OECD, but it should not be substituted for local employment evidence. A projection of employment levels also differs conceptually from a change in requested skills. An occupation can have changing skill requirements while projected employment rises, falls, or remains comparatively stable; those combinations are not contradictions because one describes the composition of demand and the other its projected scale. A skill-demand shift can occur within a growing occupation, while projected growth can occur without a large measured shift in advertised skills. Reading both dimensions helps avoid treating either one as a substitute for the other.

Figure 4.4 also places earnings and worker counts alongside these measures. Those details help readers distinguish occupations that may show similar directional changes but differ in their economic context or the number of people working in them. They do not turn the chart into a personal forecast. Earnings are context for the occupations represented, not a promise about an individual’s pay, and an occupation’s worker count does not indicate how many particular positions an employer will retain. The combined display gives a broader comparison than either change series alone, but each input keeps its own definition and measurement period.

The chart is useful when read as a map of occupational context: it can show where advertised skills are moving and set that movement beside a U.S. employment projection, with earnings and workforce size in view. But the label “disruption” should not be mistaken for observed job change or an AI-specific exposure score. Figure 4.4 does not establish that AI produced a particular occupation’s skill-demand pattern, that every worker in that occupation performs the advertised tasks, or that the projection will occur as shown. Its inputs come from different sources and periods, so a reader should compare their direction and context while retaining those boundaries. To understand a specific job, the chart can suggest a question about skill demand; actual duties, workflow changes, and employer decisions still need evidence from that work setting.

Sources: From skills to labour market opportunities: OECD Skills Outlook 2025; Occupational projections and worker characteristics: U.S. Bureau of Labor Statistics

What skills are employers advertising in AI-exposed jobs?

A separate OECD vacancy study, Artificial intelligence and the changing demand for skills in the labour market, examines occupations highly exposed to AI that do not require specialist AI skills. It reports increased demand for management and business skills and an 8 percentage-point increase in the share of vacancies requesting at least one emotional, cognitive, or digital skill. The useful interpretation is that advertised skill combinations in this group of occupations are changing. The finding does not say that every worker must acquire every skill in those categories, or that the study tested a universal retraining prescription.

The study’s unit of evidence is online vacancies and the skills requested in them. Its results therefore describe advertised requirements in the analysis, rather than a direct assessment of what workers currently know or what employers later hired them to do. The report’s result groups occupations by their high exposure to AI and distinguishes those not requiring specialist AI skills. That framing makes the finding relevant to ordinary roles where AI may interact with existing work, without making the advertised skill change an outcome for each incumbent. The evidence record supports describing this as an OECD analysis of vacancies; it does not establish a specific country coverage, observation window, or coding procedure here, so those details should not be inferred.

Management and business skills are broad categories, so the result is more informative as a pattern than as a ready-made course list. An increase in demand for such skills could reflect employers asking for coordination, planning, or business-related capabilities alongside technical change, but the vacancy result alone does not determine which interpretation applies in a given occupation. Likewise, the 8 percentage-point figure concerns the share of postings asking for at least one skill from the emotional, cognitive, or digital group. It does not mean each of those three categories rose by eight points, that the same postings requested all three, or that a worker’s skill level changed by that amount. Keeping the denominator and wording in view prevents the percentage-point change from being turned into a claim about individual proficiency.

Vacancy evidence has a practical strength: postings can reveal what employers are choosing to specify, including shifts that may be relevant before they appear in broad employment totals. It also has a boundary: formal advertisements may omit skills treated as implicit or learned on the job, and stated requirements do not confirm actual hiring or day-to-day use. The OECD finding is consequently a market signal to investigate, not a prescription to collect credentials in every listed category. A reader can use it to compare the advertised requirements of a target role with their own recurring tasks and identify a specific capability gap; whether that gap merits a course, practice, or a role search requires evidence about the work and local opportunities.

The central distinction is between a change in what vacancies request and a demonstrated change in every workplace. The OECD analysis helps correct a narrow assumption that AI-exposed work is defined only by specialist AI skills: the reported demand also includes management, business, and grouped emotional, cognitive, or digital skills. That broad pattern may help a worker frame a better question about how responsibilities are being combined. It cannot show that AI alone caused the change, that all employers share it, or that mastering every category will protect a job. Treat the reported shift as a reason to inspect role-specific evidence and the actual standard for work, not as a universal skills checklist.

Sources: Artificial intelligence and the changing demand for skills in the labour market

What have SME respondents said about staffing needs so far?

No: the available SME survey does not show that AI had already reduced staffing everywhere. In Generative AI and the SME Workforce: New Survey Evidence, Chapter 3, 83.0% of surveyed SMEs using generative AI said it had no effect on their firm's overall staff need; 9.1% said the need had decreased, and 5.5% said it had increased. The largest response was therefore no perceived change, not a cut. That is meaningful counterevidence to a claim that immediate staffing reductions were universal among the firms surveyed. It is not evidence that no firm reduced staffing, and it says nothing definitive about every small or medium-sized enterprise outside this survey.

The finding comes from a late-2024 survey of more than 5,000 SMEs in Austria, Canada, Germany, Ireland, Japan, Korea and the United Kingdom. A respondent considered best placed to describe technology use answered on behalf of the firm. The result is thus an employer-side report of perceived change in overall staffing need across seven national settings. Its value is to show what those respondents said at that point in time; it is not a count of workers who entered or left jobs, nor a before-and-after payroll comparison. The three percentages are reported separately and do not add to 100%, so they should not be treated as a complete distribution of every possible response or as if the remaining share had a known direction.

“Staff need” also needs to stay distinct from nearby outcomes. A firm can experience a change in how much work its employees handle without changing total staffing need. It can alter the mix of tasks, delay a planned hire, use contractors differently, or shift work among current employees. Conversely, an assessment that employees can complete work faster does not itself establish that the firm's total need for staff has fallen. The OECD survey result concerns respondents' perceived overall staffing need; it does not report that the 83.0% experienced no workflow change, or that the 9.1% necessarily dismissed employees. Treating workload, contractor use, hiring plans and headcount as interchangeable would turn one survey question into several unsupported outcomes.

The snapshot also cannot settle what happened later or what will happen next. Adoption may be early, uneven across activities, or concentrated in a subset of firms; staffing implications can follow workflow changes with a delay. Perceived need may also differ from realized staffing, and firm-level answers can conceal different experiences among workers or teams. These are reasons to limit the conclusion, not reasons to disregard the observed majority reporting no effect. On the surveyed firms' own account, immediate reduction was not the dominant reported outcome, and reported increases show that perceived effects were not exclusively downward.

The defensible reading is narrow but useful: these AI-using SME respondents did not report a uniform immediate cut in overall staff need. A majority reported no effect, while smaller shares reported decrease or increase. The survey offers employer-perception evidence about one late-2024 moment, not proof of zero displacement or a forecast of future staffing. For a worker deciding what this means for their own position, it cannot substitute for checking whether assignments, hiring, or responsibilities are actually changing at their workplace.

Sources: Generative AI and the SME Workforce: New Survey Evidence, Chapter 3

How can AI substitute for one task while complementing another?

AI can reduce the effort or demand attached to one activity while increasing the importance of another because a job is a bundle of linked tasks, not a single operation. Widening opportunities by investing in 21st-century skills: OECD Skills Outlook 2025 describes task-level effects as involving both substitution and complementarity. A system may perform or assist with a repeatable component, yet the workflow still needs people to define the problem, assess whether an output is fit for use, resolve exceptions, communicate with others, or accept responsibility for the result. The capability is applied to parts of work; how those parts are recombined depends on the task and how the work is organized.

Substitution describes a case where technology performs a task or reduces the human effort required for it. Complementarity describes a case where technology supports a task in a way that raises the value, reach, or volume of human contribution elsewhere. These are not mutually exclusive descriptions of an entire occupation. The same implementation can displace some routine activity and create or intensify connected work: an automated first pass may leave more cases for review, make errors more consequential, or allow a service to handle more requests. Whether those effects occur depends on reliability, the cost of checking, workflow design and whether demand responds to changed productivity. OECD's chapter synthesizes evidence across studies and task settings; it should not be read as one experiment finding a common outcome for all workers.

Consider a hypothetical workflow illustration, not an observed study or a named workplace: an AI tool drafts a routine first pass from supplied material. The worker's task may shift from composing every line to checking source evidence, correcting unsupported claims, handling unusual cases and deciding what can responsibly be sent to a client. Drafting effort could fall, while verification and exception handling become more central. If the draft is dependable and demand expands, the team might process more work; if checking is extensive or demand stays fixed, the team may organize the same work differently. Nothing in this illustration establishes a staffing result. It makes visible how task-level substitution and complementarity can coexist within one workflow.

The mechanism is conditional at each step. Automating a repeatable component can free time only if the tool actually works well enough in that context and the remaining checks do not consume the savings. A larger volume of outputs can increase review load; a lower cost per output can also change how much work customers request. Exceptions may require contextual judgment that the system cannot reliably supply, but an employer could assign those exceptions to fewer specialists, distribute them across a team, or redesign the service. Client communication and accountability may grow in importance when outputs need explanation or someone must own the decision, yet that importance does not guarantee more positions or higher pay.

This is why productivity and staffing cannot be inferred from the task mechanism alone. If a system reduces time per task, an organization may use the capacity for additional output, shorter turnaround, a different quality standard, reduced costs, or some combination. Demand, budgets, implementation expense, risk tolerance and management decisions shape which response follows. Complementary tasks may expand while total headcount stays flat or falls; a new review responsibility may replace an old duty rather than add to the workday. The OECD synthesis helps explain possible patterns, but it does not promise that workers doing complementary tasks will be retained or rewarded.

For an individual, the practical inference is to map the changed steps in a task bundle instead of classifying the whole job as automatable or safe. Identify which activity the system can perform, what inputs and conditions it needs, what must be checked, which cases still need judgment or conversation, and who is accountable if the result fails. Then ask whether your workplace is changing the volume, ownership or standard of those steps. That map can reveal where experience transfers into verification, exception handling or coordination, while keeping the key uncertainty visible: a changed task mix is not by itself evidence about net staffing, wage outcomes or the future of the role.

Sources: Widening opportunities by investing in 21st-century skills: OECD Skills Outlook 2025

What does the Outlook say about learning while already employed?

For a worker already employed, the OECD’s Widening opportunities by investing in 21st-century skills: OECD Skills Outlook 2025 treats adult learning as both a question of what is useful and who can realistically take part. Its discussion identifies limited time and financial resources, care responsibilities, and unequal access as barriers. Those conditions shape which learning options are usable: a course can be relevant in theory yet unavailable in practice if its schedule conflicts with work or care, its cost competes with essential expenses, or its prerequisites assume knowledge the learner has not had a chance to build. Feasibility therefore belongs in the initial comparison, alongside relevance, rather than being left until after a worker has selected an idealized course.

Relevance and quality matter because participation alone is not evidence of capability. A course should connect to a task or responsibility the learner expects to perform, and offer a way to practice, receive feedback, and judge whether the work meets its actual standard. The Outlook discusses course relevance and quality as concerns for adult-learning systems; it does not show that enrolling in a particular course produces a hiring, retention, or earnings gain. A completion record may document participation, while a work sample, supervised practice, or demonstrated ability to handle an exception can provide different evidence. Which evidence matters depends on the target work and whether an employer, client, or regulator requires a recognized qualification.

The OECD chapter also discusses modular learning and recognition of prior learning as policy approaches. A module can let someone address one defined gap without immediately committing to a full program; recognition of prior learning can make existing experience count toward a qualification or learning pathway. These options may reduce barriers or avoid repeating material, but the chapter does not establish that every provider offers them, that every module is sufficiently deep, or that an employer will recognize prior learning. A worker should check the actual course structure, assessment, entry requirements, and recognition rules rather than infer them from the format’s name.

A practical comparison can begin with four routes. A focused module may fit a narrow tool or knowledge gap when the learner has prerequisites and can get useful feedback. Employer-supported practice may connect learning to real tasks and provide supervision, but depends on permission, time allocation, and access to a suitable workflow. Longer formal study can supply sequence, depth, and a credential where the destination requires one, while demanding more time, money, and sustained participation. No immediate course can be a sound choice when the target requirement is unclear, safe practice is unavailable, or current duties do not yet show a meaningful gap; the next step may be to gather evidence about the task or target role first.

For each route, compare the exact capability sought, prerequisites, quality of instruction and feedback, how improvement could be demonstrated, and whether the destination actually asks for a credential. Then check schedule against paid work and caregiving, total cost and available support, and compatibility with health and energy. A short format may fit a tight schedule but leave someone unprepared for a consequential responsibility; a longer credential may provide stronger structure but impose a burden the learner cannot absorb. These are trade-offs to investigate, not outcomes ranked by the OECD. The Outlook’s account makes access part of the learning decision; this comparison is practical guidance derived from that account, not a decision instrument tested by the report.

Sources: Widening opportunities by investing in 21st-century skills: OECD Skills Outlook 2025

How can you identify the capability worth building first?

A useful learning target starts with a recurring output, not a broad label such as “AI skills.” Write down what the work must produce and which step has changed or may change: gathering information, drafting, checking, deciding, or handing work to someone else. Then name the quality or safety standard that output must meet. The standard might be factual accuracy, completeness, traceable sources, a required format, or a rule about what information can be used. This makes the question concrete: what must someone be able to do reliably for this output to be acceptable?

Next identify a meaningful failure mode and who is expected to catch it. Could a missing source, a misleading summary, an overlooked exception, or a privacy breach make the output unusable? Does the worker check it, a colleague review it, or a specialist approve it? The answer shows where responsibility sits and what capability is actually missing. If the task has changed from writing every draft to reviewing generated drafts, for example, the first gap might be evaluating evidence against a standard rather than learning another drafting interface. That is an inference about what to investigate in that workflow, not evidence that the role or its staffing will be preserved.

Consider a hypothetical illustration, not a report finding or a named workplace: someone responsible for preparing a weekly summary finds that a tool can produce a first draft from approved source documents. Before choosing training, they would define what makes the summary reliable, such as whether each claim can be checked against a source and whether important exceptions are retained. They would identify who verifies the draft and what error would matter most. If the missing capability is tracing claims back to source material, practice might focus on checking provenance and spotting unsupported statements; if the harder gap is interpreting the material, domain knowledge or data literacy may matter more. The example supplies no measured result and does not imply that any tool can safely perform the task.

From that analysis, select the smallest capability whose improvement could be shown in practice. A demonstration might be a correctly checked sample, an explanation of a decision, or a task completed against a defined review standard. In editorial synthesis, it helps to distinguish durable foundations—problem framing, data literacy, domain knowledge, and verification—from knowledge of a particular interface, which can change quickly. This is a useful way to organize learning questions, not an OECD taxonomy or a guarantee that an employer will value a given demonstration. The distinction can prevent a worker from confusing familiarity with a current product’s buttons with the ability to judge whether its output is fit for the job.

Before trying a live tool, check confidentiality rules, workplace policy, and whether the appropriate person has authorized the use; do not enter sensitive work material into an unapproved system. If safe, permitted practice is available, use it to test the smallest capability gap against the stated standard. If it is not, the next useful evidence may come from training research or from examining current target-role requirements: what prerequisites, credentials, and demonstrated tasks do relevant openings actually specify? A clear learning target helps choose among those inquiries, but it cannot show that an employer will fund training, recognize the evidence, or retain the role.

When is an upgrade enough, and when should you investigate another role?

An upgrade is a proportionate first route when meaningful responsibilities remain in the role and there is a feasible, supported way to learn and verify the changed work. Investigate an adjacent move when your experience transfers to work with a more workable task mix. Explore a larger transition when assignments are being removed and you can identify a destination that fits your constraints. This is practical editorial reasoning, not a validated decision model or forecast: the trigger is evidence about work and feasible options, not an occupational exposure score or a headline about AI.

For a within-role upgrade, look for evidence that the work is actually being redesigned: a manager or team has named new responsibilities, the relevant workflow is being introduced, or quality expectations now include checking, exception handling, or another task. Then ask whether you can practice on appropriate work, get feedback, and verify performance against a clear standard. A training plan without access to the workflow or a way to judge the result may not answer the real need. If redesign is only being discussed, a small inquiry—what task is changing, when, and what good work will require—can test whether investing in a course is timely.

An adjacent move deserves investigation when your existing knowledge, relationships, or accountability could carry into another role whose recurring tasks appear more workable for you. The useful evidence is not that a job title sounds safer; inspect several current openings in the location or remote market you could realistically access. Compare recurring duties, required tools, credentials, experience, and whether employers expect a formal qualification or demonstrable work. Repeated requirements give a stronger basis for a gap assessment than one unusual listing. They still describe advertised criteria, not a promise that a candidate will be hired, so check whether the transition cost and preparation are acceptable before treating it as a plan.

A larger change becomes more worth exploring when concrete assignments, hours, or responsibilities are being removed and a plausible destination has been checked against real openings and your circumstances. The evidence might include a documented staffing or workflow change, not simply that a tool can perform one task. Urgency matters: if a confirmed change is imminent, waiting for a lengthy within-role pilot could leave too little time to prepare. Even then, separate what is known about the change from what remains uncertain about future work, and compare more than one feasible route if time allows.

Use the same constraint test for all three paths. What is the minimum pay you need, and can the route meet it during training and afterward? Does it require relocation, a credential, unpaid study time, or costs you cannot absorb? Can the schedule fit work and care, and is the pace compatible with your health and family responsibilities? Which experience could shorten the transition, and what evidence would show you are ready? A role with a promising task mix may still be inaccessible if its prerequisites or location do not fit. Conversely, a modest upgrade may be more workable if it builds on experience and has employer-supported practice. These questions do not calculate a score; they expose trade-offs that an exposure label hides.

A within-role pilot may be too slow if staffing changes are documented and near, while leaving solely because tasks are exposed may discard useful experience before a local alternative is tested. Set a review point: after you confirm the change and the learning support, or after you compare a small set of relevant openings and constraints. If neither work redesign nor a realistic destination is yet evidenced, gather that information before committing money or making an irreversible move. The decision can then change as concrete assignments, requirements, or personal constraints change.

What is the next useful step this week?

This week, choose one recurring task and write three lines: what the task produces, what has observably changed, and how you or a reviewer will decide whether the result is acceptable. If you cannot name a reliable change yet, ask a manager or teammate what is changing and when; do not buy training to solve a gap you have not identified. If the task and standard are clear, choose one next inquiry: find a feasible way to practice and get feedback, or compare current openings for an adjacent role against your pay, location, credential, time, cost, health, and family constraints. Keep the inquiry small enough to finish: clarify one workflow expectation or examine a few relevant listings, then note what evidence would change your choice. This is a reversible step, not a commitment to a course or departure. The free task-level checker at `/ai-job-risk-checker` can organize change-pressure signals and first actions; it does not estimate a validated probability of displacement. The paid career roadmap compares scenarios and builds a plan around stated constraints, but it cannot guarantee employment or income. Use either after you have named the task question you need help deciding.

Questions readers ask