What the UK AI Labour Market Survey 2025 Measures

The short answer: a focused employer signal, not a national headcount

The most consequential fact about the UK AI Labour Market Survey 2025 is not its largest percentage. It is who and what its figures count. The report describes answers from organisations associated with the UK AI market: what they say about skills in their workforces, recruitment, vacancies, training and related matters. It does not test individual workers, count every vacancy in the country, or survey a probability sample of the entire UK labour market. That boundary changes what a reader can responsibly conclude.

The report says 97% of respondents identified at least one skills gap. Properly stated, that is a very high share of answering organisations reporting a gap under the survey’s categories. It is not evidence that 97% of AI workers are unqualified, that 97% of UK employers have a shortage, or that a particular worker is likely to lose a job. The first interpretation preserves both the respondent unit and the self-reported nature of the answer; the others change the measure into something the survey did not make.

There is still practical value here. A structured sector survey can reveal the capabilities employers say they need, the recruiting problems they encounter and the routes they use to develop staff. Those details can help educators, policymakers and employers formulate better questions. They can also help a worker identify a skill worth checking against real openings. But the answer is a signal for investigation, not a national shortage counter or a personal forecast.

The evidence comes from 119 organisations responding after 3,940 were approached. The report’s method section describes selection through commercial channels, self-identification as an AI provider and identification by the Department for Science, Innovation and Technology (DSIT). It also cautions about self-selection. A response rate around 3% does not tell us exactly how nonrespondents differ, nor does it prove their answers would be lower or higher. It does mean the responses cannot simply be assumed to represent all organisations in the UK AI economy.

That makes the report neither worthless nor definitive. It is more informative than an anecdote because it asks a set of employers a common series of questions; it is narrower than a representative national statistic because participation and sector boundaries limit generalisation. A careful headline keeps the denominator and verb intact: ‘responding organisations reported’. The distinction is as important for a worker considering a course as for a journalist writing about shortages.

This formulation is not just cautious wording. It tells a reader what kind of next question to ask. If the claim concerns what participating employers experience, the survey is directly relevant. If it concerns how widespread a skill gap is across all UK employees, the reader needs a worker or representative employer measure. If it concerns whether an applicant can get hired, postings, recruitment outcomes and role requirements are more pertinent. Evidence should match the decision being made.

Sources: AI Labour Market Survey 2025; AI Labour Market Survey 2025 report: executive summary

What did the survey ask, and how was it assembled?

The survey was designed to describe several connected parts of the AI labour market, rather than to administer one test of skill. The report covers organisational skills gaps, technical and non-technical capabilities, recruitment and vacancies, training and routes into roles, workforce composition, and use or planned use of selected technologies. The research also includes interviews intended to add context. In combination, those components give a broad picture of what participating organisations say is happening inside and around their workforces.

The distinction between a questionnaire and an assessment matters. An organisation’s answer about a capability is a report of its view, not a standardised examination of employees. A hiring manager may describe a gap because a team needs deeper expertise, because a tool or product has changed, because a vacancy has remained open, or because work is being reorganised. Those situations can overlap, but they are not interchangeable. The survey records employers’ accounts; it does not directly determine which individual meets an independently defined proficiency threshold.

According to the report, the survey launched in January 2025 and was sent to 3,940 organisations. The 119 respondents form the survey’s analytical base, while 20 one-to-one interviews contribute qualitative material. The combination permits descriptive percentages alongside examples of how respondents understand the issues. Interviews can explain language, context and mechanisms that a fixed response list may miss. They do not turn a selective sample into a representative one, and they cannot establish how common a view is across the sector on their own.

The report also says not every question was mandatory. As a result, denominators can vary from chart to chart. If one item has fewer responses than the overall survey, its percentage should be read against the valid answers for that item, not automatically against all 119 organisations. Several questions allow multiple responses, meaning the listed categories can overlap. A company might report a technical gap, a non-technical gap and difficulty recruiting at the same time. Adding those percentages as if they were separate portions of a pie would manufacture a total the instrument did not report.

Fieldwork timing is another boundary. The survey launched in January 2025, but the report was published on 28 January 2026. A date printed on the publication page is not the date every response was collected, and a later publication date does not make the data a description of the exact current market. Readers should distinguish the period respondents were asked about from the later date on which the findings became public.

The study was commissioned by DSIT and conducted by Gardiner & Theobald. The report states that its findings and recommendations are the contractor’s views, not government policy. Commissioning provides institutional context; it does not mean each interpretation has been adopted as an official government position. A sound reading therefore keeps three things visible: the question asked, the responding organisations who answered, and the date or reference period each answer concerns.

Executive summaries have to compress methods and caveats to foreground findings, while technical sections describe how evidence was assembled. A strong phrase such as ‘critical gap’ is not itself a statistical definition. The operational definition comes from the instrument and method. Here, the method reveals an employer-reported survey with a small responding group and accompanying interviews. That is the evidence behind the rhetoric, not a separate national census hidden in the headline.

Sources: AI Labour Market Survey 2025; AI Labour Market Survey 2025 report: executive summary

Whose experience sits behind the percentages?

The responding organisations have a recognisable profile. The survey report says 93% were small or medium-sized enterprises, 66% developed commercial AI solutions, and 77% were located in Greater London and southern England. London alone accounted for about half of responses. These are descriptions of the respondents, not estimates of the entire UK AI sector’s size, geography or business mix.

That composition makes the survey especially useful for understanding the kinds of concerns that arise among participating, often smaller organisations involved in AI development or provision. A small product company trying to recruit machine-learning specialists may face a different labour problem from a large bank adopting external tools, a public body buying an AI service, or a regional employer with no AI-specific team. The report includes organisations with different forms of involvement, but the concentration among SMEs and commercial developers affects which experiences are most visible in the results.

This is not a reason to dismiss the findings. A sample need not represent every possible organisation to reveal useful patterns. The key is to describe the population to which a claim most directly applies. ‘Among the responding organisations, most of which were SMEs’ is more faithful than ‘UK companies say’. The former gives readers enough information to judge relevance; the latter quietly widens the inference.

The phrase ‘AI sector’ also carries a boundary choice. Does it mean firms building AI models or products? Businesses selling services that rely on AI? Organisations employing specialist AI staff? Or any employer beginning to use a purchased tool? The report’s groups encompass varying relationships with AI, so the answer depends partly on its operational categories. A firm that procures a third-party system may have a skills need around procurement, data handling or oversight without being an AI developer. A developer may need technical research and engineering skills. Pooling both can be appropriate for a broad ecosystem question, but it should not be mistaken for one uniform occupation or labour market.

DSIT’s separate Artificial Intelligence Sector Study 2024 illustrates why this distinction matters. That publication’s purpose is to estimate the scale and scope of firms developing AI products and services and examine sector contribution, investment, products and market dynamics. It addresses what counts as an AI-sector firm from a different angle. It is not a worker-proficiency or skills-shortage measure. Business population, workforce skills and advertised vacancies are separate objects of measurement, even when each uses the label ‘AI’.

For a person in a particular region or role, the respondent profile is a cue to check local evidence rather than assume the survey mirrors their situation. A developer in London may find its recruitment findings closer to home than a finance administrator in rural Wales. Yet even a close industry match cannot answer whether that individual’s own tasks or competencies are scarce. That requires evidence at a more specific level than this survey supplies.

Geography matters because a national label can conceal very different access to jobs and training. The respondent concentration in London and southern England cannot establish that a person elsewhere faces the same mix of employers, specialist vacancies or development routes. The OECD posting analysis also finds higher advertised AI-hiring intensity in London and research hubs, but postings are a separate measure with their own limits. Together these observations justify checking regional evidence; they do not establish that moving is necessary or feasible for someone whose family, health, salary needs or care responsibilities tie them to a place.

Sources: AI Labour Market Survey 2025; Measuring the demand for AI skills in the United Kingdom; Artificial Intelligence sector study 2024

What counts as a ‘skills shortage’ in this report?

The report uses several employer-facing indicators that are related but answer different questions. It asks whether organisations see skills gaps in their current workforce, what kinds of skills they name, whether they have had AI-related vacancies, whether vacancies were hard to fill, what reasons employers give for difficulty, and whether a gap affects business goals. Each indicator has a distinct unit and reference point. There is no single master percentage that combines them into a national skills-shortage measure.

A reported skills gap concerns a capability an organisation believes its existing workforce needs. A vacancy concerns a role the organisation says it tried to fill. A hard-to-fill vacancy concerns its recruitment experience. A stated effect on business goals concerns the employer’s attribution about consequences. These can form a plausible chain: a company needs expertise, lacks it internally, recruits, struggles to find a suitable applicant and says the gap impedes a project. But the survey does not prove that every case follows this sequence or that one factor caused another.

The UK Employer Skills Survey (ESS) uses more explicit distinctions. The 2024 ESS describes a skills gap as an employer judging at least one current employee not fully proficient. It defines a skill-shortage vacancy as a hard-to-fill vacancy attributed to applicants’ lack of skills, qualifications or experience. Those definitions show why a skills gap inside a team and a shortage among applicants are different concepts. An organisation can have one without the other: existing staff might need development while new recruitment is straightforward, or an employer might struggle to hire even though current employees are fully proficient.

The AI Labour Market Survey is designed around a sector-specific subject and asks employers to describe their circumstances; ESS is a UK-wide employer statistics programme with defined incidence, density and volume measures. The reports are therefore complementary rather than interchangeable. It would be wrong to silently import ESS’s technical definition into every use of ‘gap’ in the AI survey, just as it would be wrong to treat the latter as if it had ESS’s sampling frame.

This distinction also matters for education decisions. A worker hearing that companies report a gap should ask: gap in which capability, for which activity, at what level, in what region, and according to what observation? ‘Understanding AI concepts’ might mean different things in model development, product implementation and responsible procurement. ‘Programming’ may be a prerequisite for one role but irrelevant to a worker trying to use an approved assistant to summarise internal documents. Broad labels are useful for a first map, not a course prescription.

A practical reading keeps the evidence layers separate: perceived in-house capability, attempted hiring, applicant fit, business consequences and planned technology use. Combining them into the single word ‘shortage’ can be convenient shorthand, but it overstates what is known unless the article immediately specifies which measure it means.

One useful diagnostic is to ask what action a measure could rationally inform. A perceived current-workforce gap may support a conversation about training or work allocation. A skill-shortage vacancy may support recruitment redesign or a review of experience requirements. A future technology plan may support scenario planning. An individual proficiency gap requires an assessment at that level. When one percentage is asked to answer all these questions, interpretation has outrun the data.

Sources: AI Labour Market Survey 2025; Employer Skills Survey 2024

What does ‘97% reported a gap’ tell us, and what does it leave out?

The report says 97% of respondent organisations identified at least one skills gap; 57% reported a technical gap and 30% a non-technical gap. Among technical categories, the report lists AI concepts and algorithms at 60%, data management at 38%, programming at 36%, data analytics or statistics at 36%, and software or systems engineering at 34%. These figures show which categories respondents selected. They do not measure the proportion of workers who failed a skills test.

The denominator is organisations that answered the relevant question, not employees. Imagine a participating company with a hundred employees and a gap in one specialist role. It may answer that it has a skills gap, just as a company with a much smaller workforce could. The survey’s organisation-level percentage gives each response its place in the respondent count; it does not weight that answer into a national estimate of affected workers unless the methodology explicitly does so. The report does not present the 97% as a weighted prevalence measure among individual workers.

Nor does the figure mean that nearly all AI tasks are beyond current workers. ‘At least one gap’ is a broad threshold: one named need is enough for a positive answer. The categories may overlap, and the survey items are not a standardized competency examination. Respondents may identify gaps because they want to deepen an existing capability, because a specific project calls for a rare specialism, or because the organisation is growing. Those explanations can be important to policy, but they differ from a general claim that workers cannot perform their jobs.

The 57% technical and 30% non-technical figures also should not be added to get a total. They may have different bases and can overlap with each other. A company can need technical expertise and also report needs in communication, management or related areas. Multi-select questions are often designed precisely to capture that overlap. Their percentages describe selection frequency, not mutually exclusive slices.

A useful worked interpretation is to translate each number into its full sentence before reacting. Instead of ‘there is a 97% AI skills shortage’, write: ‘97% of the organisations answering this question said they had at least one skills gap, according to the survey’s categories.’ Then ask what kinds of gaps they selected and whether another measure supports a related claim. This sentence is less dramatic, but it preserves what the evidence actually observes.

For an individual, the relevant next step is not to enrol in a broad AI course simply because the number is high. Identify the work you want to do and the tasks involved. A data scientist might inspect current advertisements for statistical modelling, data engineering or evaluation requirements. A procurement analyst might need to understand data protection, system limits and vendor assessment. A manager integrating a tool may need workflow design and quality checks. The same headline statistic points to different learning choices because the jobs and tasks differ.

The report’s numbers are valuable as a map of participating employers’ stated concerns. They become misleading when the map is relabelled as a census, a proficiency test or an individual risk score.

The distinction between technical and non-technical labels also deserves care. The report describes non-technical skills as those not requiring specialised tools, techniques or knowledge such as programming, engineering or modelling. That category is not a synonym for unimportant or generic. Communication, judgment and coordination may be central to whether technical work is understood and used. Still, a category label alone cannot tell a learner which behaviour a particular role requires or how to demonstrate it.

Sources: AI Labour Market Survey 2025

What does the vacancy evidence add?

The survey reports that 58% of respondents had AI-related vacancies in the previous two years and 35% reported hard-to-fill AI vacancies. It identifies lack of work experience and technical skills or knowledge among the reasons employers gave. This adds a recruitment dimension to the skills-gap answers: participating organisations were not only describing needs among existing staff; many also described attempts to hire.

But an employer reporting a hard-to-fill role is not the same thing as a national count of unmet demand. The survey’s figures describe retrospective experiences among selected respondents. They do not enumerate every open UK role, record a common vacancy duration for the full labour market, or establish how many candidates were considered. A vacancy may be created by expansion, replacement, a temporary project or a new organisational structure. Those possibilities all represent hiring activity, but they do not all imply the same net increase in employment.

‘Hard to fill’ also has to be understood in relation to the survey question and its respondent. Employers may face a narrow combination of requirements, salary limits, location constraints, recognition problems or slow recruitment processes. A position can remain open because the needed candidate pool is small, but also because an employer’s terms do not attract available candidates. The survey records stated reasons, not an independent diagnosis that isolates supply from employer choices or other frictions.

ESS provides a useful contrast. Its 2024 release says 6% of UK employers had a skill-shortage vacancy, defined as a hard-to-fill vacancy because applicants lacked skills, qualifications or experience. That release draws on 22,712 employers and states the results are representative of employers across the UK, using random-probability sampling. The AI survey’s 35% hard-to-fill share cannot be placed beside ESS’s 6% as if one were a more precise version of the other: the populations, sector focus, questions and denominators differ.

The comparison does clarify the kind of additional evidence needed. A national shortage claim should specify the vacancy definition, how many employers or vacancies are represented, and whether the problem is applicant competence, location, pay, experience or some combination. Job postings can add an observable advertised-demand signal, but they also miss jobs filled through networks or internal promotion and cannot show whether applicants were qualified. Survey and posting data answer different questions.

For a worker, vacancies are more actionable when narrowed to a target occupation and place. Inspect a sample of recent listings that genuinely match the desired level, not just the broad label ‘AI’. Record repeated requirements, required experience, location and whether the role expects development, implementation, governance or domain use. This is a local check, not a guarantee of demand, but it keeps learning decisions tied to jobs that could plausibly fit the reader’s existing experience and constraints.

The survey’s vacancy findings therefore strengthen the conclusion that respondents describe real recruiting friction. They do not settle how large the UK-wide shortage is, what caused the friction, or whether a particular applicant will find work.

There is a practical distinction between ‘there are jobs’ and ‘this is a viable move for me’. A role can be advertised and still be unsuitable because it requires relocation, a salary reduction, credentials a candidate cannot obtain quickly, or conditions incompatible with health or family needs. Conversely, someone may already have adjacent experience that meets much of the role’s needs. Vacancy evidence is most useful when filtered through real constraints, not treated as a promise that demand will overcome them.

Sources: AI Labour Market Survey 2025; Employer Skills Survey 2024

When employers say gaps affect business goals, how strong is that evidence?

The report says 28% of surveyed organisations identified existing employees’ technical skills gaps as affecting their ability to achieve business goals. That is meaningful evidence of employer perception: a portion of respondents connected their own reported technical gaps with objectives they said mattered. It helps explain why the issue may receive attention inside a business, rather than remaining an abstract training concern.

The sentence must stop before it turns into a measured productivity loss. The report does not describe a counterfactual comparison in which otherwise similar organisations had different skill gaps. It does not audit foregone output or estimate how much revenue, innovation or delivery time was lost because of a specific capability. The result records an employer attribution, not a causal effect size. The business may face other constraints at the same time, such as investment, data access, staffing, regulation or project prioritisation.

This difference between attribution and causal evidence is not peculiar to AI research. Organisations usually know their goals and can report what they believe is blocking them. That knowledge is useful, especially when planning training or recruitment. But identifying a reported obstacle does not demonstrate that removing it would produce the intended result. A team might acquire a technical skill and still lack usable data, authority to deploy a system, or a customer need for the proposed product.

The 2025 report also describes planned technology use, including plans to use agentic AI. A plan is evidence of intent at the time of response, not an observed rollout. Implementation may be delayed, narrowed, cancelled or changed. Even a realised deployment would not by itself show that headcount rose or fell: adoption, task redesign, labour demand and displacement are separate outcomes. A respondent can expect a tool to affect work without the survey proving what happened later.

A careful reader can preserve the importance of the finding without inflating it. ‘Some respondents said technical gaps affected business goals’ signals an operational concern worth studying. ‘Skills shortages are costing the UK a defined amount of productivity’ would require data this measure does not provide. The former can justify asking employers which task or project is impeded and what capability would help. The latter needs output measures, a clear comparison and a design that supports a causal claim.

For workplace decisions, ask for a concrete example: Which deliverable was delayed? What work could not be completed? What evidence would show that training changed the bottleneck? If the answer concerns a repeated task, a bounded project can test whether a course, mentorship or tool practice addresses it. If the problem is a missing role or poor recruitment design, individual upskilling may not be the right remedy.

The figure is strongest as an indicator of what respondents perceive as consequential. It is weakest when used to imply the survey measured an economy-wide cost or proved that a specific intervention will work.

The next step for a policymaker is concrete: distinguish a training need from a missing hiring channel, an access barrier or a management decision. If employers say experience is hard to find, supervised practice may be worth examining; the survey alone does not prove that a new course would supply it. If the barrier is a mismatch between formal degree requirements and demonstrable capability, recruitment criteria may matter. These are hypotheses to test, not interventions whose success is established by the 28% finding.

Sources: AI Labour Market Survey 2025

Can the 2020 comparison show that shortages are getting worse?

The 2025 report compares some results with the earlier UK AI labour-market study. In that comparison, the share reporting a technical skills gap rises from 49% in 2020 to 57% in 2025, while the non-technical figure moves from 32% to 30%. The respondent counts for the organisational surveys are similar: 118 in the 2020 study and 119 in 2025. These repeated snapshots offer context, but the differences do not establish a controlled national trend.

A repeated question can make two studies more comparable, yet matching wording does not make their samples the same. The 2025 respondents are not a panel of the same employers followed over time. The 2020 report used a mixed-method design with different research strands, and the 2025 report had 20 interviews compared with 50 in the earlier study. The composition of organisations, business conditions and AI market itself can also change. A difference between percentages could reflect real change, a different mix of respondents, changed interpretations or more than one factor.

The reports do not present these comparisons as a controlled experiment or a panel estimate. Nor does the evidence ledger indicate a statistical significance test for the observed survey percentages. Without that information, a reader should not describe a nine-point difference as a proven increase in the national shortage. It is accurate to say the later respondent sample reported technical gaps more often than the 2020 respondent sample on the comparison presented by the report.

There is a second issue: even a genuine rise in employer-reported gaps would not alone explain why it happened. The field may have grown, job definitions may have shifted, respondents may have become more aware of new capabilities, or organisations may now expect a different level of proficiency. A skills gap is relative to a task and employer standard; changing the work or the standard can alter the answer even if workers’ underlying abilities remain stable.

The best use of the comparison is to generate testable follow-up questions. Are the same kinds of technical capabilities appearing in current vacancies? Do employers in the same region and firm type describe similar recruiting problems? Are training routes changing? Are the questions and respondent populations sufficiently alike to support a sharper trend analysis? A new study with a defined population, stable instrument and transparent weighting could help answer those questions more firmly.

The 2020 report remains relevant because it shows this topic was investigated before the current wave of public attention. Its purpose included examining skills gaps, recruitment, training, geography and workforce diversity. That continuity helps frame the 2025 report as part of a research series. It does not mean every result is directly comparable or that current tools caused the differences.

For decisions, avoid treating the trend line as a reason to abandon a current occupation or rush into a degree. The change is descriptive and sector-level. A worker should still inspect the skills demanded for a target role, determine whether the gap is learnable within their constraints, and assess whether the role is geographically and financially viable.

A responsible trend claim would state the comparison explicitly: among respondents in the two editions, the share reporting technical gaps was higher in the later survey. It would then note that the studies are not a controlled panel and that the report does not establish population-wide change. That gives readers the observed direction without pretending to know its cause. A consistent series, representative sample or validated workforce measure could revise the conclusion.

Sources: AI Labour Market Survey 2025; Understanding the UK AI labour market: 2020; Understanding the UK AI labour market: 2020 full report

How is this different from wider employer statistics and vacancy studies?

The most useful comparison is not a contest over which number is ‘right’. It is a map of what each source observes. The AI Labour Market Survey asks a selected group of AI-connected organisations about their experiences and perceptions. ESS measures employer-reported skills gaps and shortage vacancies across a representative UK employer survey. The OECD’s analysis estimates the share of online postings related to AI skills. A separate government-commissioned analysis classifies AI-related vacancy postings using another taxonomy and dataset. These sources illuminate different stages and populations.

ESS 2024 surveyed 22,712 employers across the UK and reports a random-probability sampling approach, with results representative of UK employers. It defines skills-gap incidence as the share of sites judging at least one employee not fully proficient; density is the share of employees judged not fully proficient; volume is the estimated number of such employees. It reports that 12% of employers had at least one skills gap and that the equivalent workforce share was 4.0% in 2024. This is a broader employer baseline, but it does not isolate AI-sector needs in the same way as the specialist survey.

ESS also reports a distinct shortage-vacancy measure. Six percent of employers had a skill-shortage vacancy in 2024, and 27% of all vacancies were skill-shortage vacancies. The definition links a hard-to-fill role to applicants’ lack of skills, qualifications or experience. Those details make ESS valuable for national context. They do not convert the AI survey’s respondent reports into comparable quantities, because one covers a specific selected set of AI-connected organisations and the other uses a different sample, question design and population.

The OECD working paper ‘Measuring the demand for AI skills in the United Kingdom’ analyses Lightcast online postings from 2012 to 2022. It applies a natural-language-processing classification rule to estimate AI-hiring intensity, meaning the share of job postings related to AI skills. It reports an economy-wide average of 0.6% over 2017–2022, with higher AI-related posting intensity in London and research hubs and changing demand across occupations and industries. This supports a claim about advertised demand in that time window, not about current employees’ proficiency or how many vacancies employers could not fill.

The Department for Science, Innovation and Technology’s ‘AI Skills for Life and Work’ programme takes a wider view still. Its Drivers analysis describes a multi-method project with evidence review, forecasting, public and employer surveys, expert engagement and scenarios. Its vacancy analysis examined Lightcast postings from January 2021 to December 2023, reporting 448,484 postings in scope, including 118,017 AI specialist and 54,805 AI expert roles. Those are classified advertisements, not a count of people hired or currently competent. The research also acknowledges that vacancy analysis shows demand for new roles and does not capture skills used in existing jobs.

These sources cannot simply validate or refute one another. More postings may indicate that employers advertise more AI-related roles, while a survey may find that responding organisations perceive internal gaps. A national employer survey may report lower broad skills-gap incidence because most employers are not specialist AI firms. The signals can all be true at once. Their differences are expected when the unit, time period, definitions and sampling frames differ.

There is also a timing issue. The OECD posting analysis ends in 2022, the broader vacancy analysis ends in 2023, the ESS refers to 2024, and the specialist survey launched in 2025 and was published in 2026. They do not describe one synchronized market snapshot. A reader should avoid narrating them as if all were observed in the same month.

For an individual or employer, triangulation means asking whether the same specific signal appears across evidence suited to the question. For national prevalence, use representative employer data with explicit definitions. For advertised role requirements, inspect postings and understand classification limits. For the reported concerns of AI-connected companies, use the specialist survey. For a personal learning choice, narrow the job and task first. No one measure answers all four questions.

Triangulation is not majority vote. If a survey reports perceived gaps, postings show advertising activity and a broad survey reports workforce proficiency, do not count which source has the biggest percentage. Ask whether their separate observations fit a coherent explanation. Specialist firms may report needs concentrated in a small part of the economy, while postings distribute across occupations and broad surveys describe different training issues. The measures can complement each other without collapsing into one score.

Sources: AI Labour Market Survey 2025; Employer Skills Survey 2024; Measuring the demand for AI skills in the United Kingdom; AI Skills for Life and Work: Drivers analysis; Artificial Intelligence sector study 2024

A practical reading rule: from survey result to a useful next step

A responsible interpretation can be built in four steps. First, name the unit: organisation, employee, vacancy, posting or task. Second, state the observation: perceived gap, judged proficiency, hard-to-fill recruitment, advertised demand or planned tool use. Third, preserve the population and period. Fourth, say what the measure cannot establish. This simple discipline stops an employer survey percentage from quietly becoming a national worker statistic or a forecast of job loss.

Consider an organisation that says it lacks data-management capability. The survey supports the conclusion that the respondent perceives that need. It does not show whether the cause is too few analysts, poor data systems, unclear responsibility or a new project. A recruiter may report a hard-to-fill role, but that observation does not tell us whether the pay, location, experience requirement or candidate supply is decisive. The next useful question is diagnostic: what work is blocked and what evidence would show that a change solved it?

For a worker, translate the broad label into a work sample. If a target role repeatedly asks for data analysis, can you show a small, relevant analysis with documented assumptions and checks? If postings require model evaluation, do they expect coding and statistical depth, or the ability to compare outputs in a domain workflow? If the work is implementation, does the employer need software engineering, configuration, security review or stakeholder coordination? The distinction can change whether the right next move is self-study, a project, a short course, employer-supported training or a longer formal programme.

Compare learning options against the outcome rather than the label. A degree can provide structured foundations, time for sustained practice and a recognised signal, but it takes longer and may not suit someone who needs to preserve income or cannot relocate. A focused course can introduce a defined skill more quickly, but completion alone does not prove workplace capability. A certificate may signal completion to some employers, yet its value depends on the provider, content and role. A project can show applied work, though it needs feedback and should not expose confidential data. Self-study costs less but can leave gaps in sequencing and assessment. No route is universally best.

A bounded next step is to collect a small, current sample of relevant openings in the geography and salary range that are realistic for you. Mark recurring requirements and distinguish essential skills from preferred ones. Then choose one requirement that connects to your existing experience and can be tested through a practical exercise. Set a short review point: did the project improve your ability to perform the target task, and do the openings still support the move? This is a learning experiment, not a promise of hiring.

If your aim is to use AI within an established profession, you may need applied literacy, verification habits, and knowledge of the rules and risks in that field, rather than machine-learning engineering. If you want to build AI products, software and data foundations may matter more. If you want to conduct machine-learning research, deeper mathematics and formal study may be relevant. Clarifying the intended outcome prevents a sector shortage headline from directing everyone toward the same curriculum.

The report’s most durable contribution is therefore not a forecast about an individual’s fate. It is a set of employer-reported questions about capabilities, hiring and development that can be checked against a reader’s own work and local market.

A useful conversation can begin at work: ‘Which recurring task or deliverable is limited by the capability we call an AI skills gap, and what would we observe if that changed?’ That invites a manager to name actual work, distinguish training from resourcing, and agree on a practical test. It avoids turning a sector headline into pressure on an employee to retrain without evidence. The final verdict is clear: the survey details what selected AI-connected organisations reported, but does not support a national headcount of missing skills, a causal estimate of business loss or an individual probability of displacement.

A practical discussion can stay modest: choose one workflow, identify where judgment or checking remains essential, and agree how to review the result after a defined trial. If the task changes but the role remains viable, improvement within the current occupation may be enough. If evidence points to a different skill or adjacent role, compare that path with the cost, time and income implications before committing. The survey can prompt this inquiry, but the answer must come from the work and circumstances in front of the reader.

Sources: AI Labour Market Survey 2025; Employer Skills Survey 2024; Measuring the demand for AI skills in the United Kingdom

Questions readers ask