What Changed in the December 2025 UK AI Skills Framework?

What changed in the December 2025 consultation version?

The December 2025 consultation record supports a high-level answer: the framework was presented as having updated competency domains and duty-based articulations, while the Alan Turing Institute described refined personas, clearer duty statements, defined competencies and new proficiency levels. These summaries identify the kinds of revision made; they do not provide an item-by-item diff showing which individual competency statements changed. The December 8 date belongs to the Zenodo record titled “AI Skills for Business Competency Framework (Version 3.0.0-consultation)”; the Institute’s page says the consultation closed on January 19, 2026. Together, the records establish the consultation’s timing and reported themes, not a complete account of every textual edit.

That distinction matters because “what changed?” can mean two different things. At the record level, the answer is that the consultation version was described as clarifying how roles, duties, competencies and proficiency should be represented, alongside changes to competency domains. At the statement level, answering whether a particular duty or capability was added, removed, renamed or rewritten requires the actual relevant text from both versions. Neither the Zenodo summary nor the Alan Turing Institute’s consultation description supplies that comparison. A phrase such as “new proficiency levels” should therefore be attributed to the publisher’s summary, rather than treated as proof that all levels or their boundaries were newly designed.

The date also fixes the scope of the claim. This is a historical description of the version put out for consultation in December 2025, not a claim that its proposed wording became the final published wording. Zenodo’s later “AI Skills for Business Competency Framework | Zenodo (Version 3.0.0)” record is dated March 9, 2026, so readers checking the framework now need that later release for current wording. Release sequence alone does not reveal which consultation proposals carried through. The strongest defensible conclusion is thus specific but bounded: the publisher and record report substantial structural and competency-domain refinements, while the available summaries cannot establish the exact change history of any named competency.

In practical terms, the record is useful for identifying the consultation’s stated direction and finding the version whose wording needs inspection; it is not a substitute for that inspection. A team discussing a particular role should avoid translating the broad summary into a claim that a named task or competency was newly required, removed, or materially altered. The consultation label itself signals proposed text under review, and the later release date signals that version succession occurred. Neither label, on its own, tells a reader whether a proposal survived unchanged. This is why a careful account can report the update themes confidently while leaving item-level conclusions open until both texts are checked.

Sources: AI Skills for Business Competency Framework (Version 3.0.0-consultation) | Zenodo; AI Skills for Business Competency Framework | The Alan Turing Institute

Which ideas were already present in the predecessor?

Several organizing ideas in the framework were already visible in the accessible predecessor, “AI Skills for Business Competency Framework (2023 public consultation draft).” That draft organized its material around five dimensions and included role personas, a lifecycle orientation and four named proficiency levels: Awareness, Working, Practitioner and Expert. These features matter when assessing novelty: the broad ideas of describing capabilities through dimensions, locating them in role or lifecycle context, and distinguishing levels of proficiency did not first appear in the December 2025 consultation account. The 2023 document is a public consultation draft hosted by the Alan Turing Institute; the Zenodo predecessor record titled “AI Skills for Business Competency Framework | Zenodo (predecessor record)” anchors its public record and version context.

This comparison changes how to read the December publisher summary, but only at the level the sources support. When the Institute describes “new proficiency levels,” the predecessor shows that proficiency levels as a concept were already present. A reasonable, limited inference is that the December version introduced a revised or more explicit proficiency scheme, or new levels within it, rather than inventing the practice of describing proficiency from scratch. The summary does not say which of those possibilities applies, and the accessible predecessor cannot resolve it by itself. The same reasoning applies to personas and duties: their presence in an earlier architecture means they are inherited organizing concepts, while the reported refinement or clarification may still represent meaningful revision.

The predecessor should be treated as a baseline for novelty, not as a complete bridge between every intervening edition and the December consultation. It is explicitly a 2023 consultation draft. Its existence establishes that the named architecture and levels appeared in that draft; it does not establish that their wording, count, boundaries or role remained unchanged in a later final version and then changed again in December 2025. Likewise, the Zenodo predecessor record helps identify the prior public record but provides no verified item-by-item comparison. So the safe historical distinction is between concepts already present in the 2023 draft and refinements reported for December. To claim a specific later change, a reader would need to verify the relevant passage in the December file and the applicable predecessor or intervening published file directly.

This is a useful guardrail for anyone summarizing framework history: novelty can refer to a concept, its wording, its placement, or its level of detail, and those are different claims. The predecessor supports the first comparison for the structures it visibly contains. It does not settle the latter comparisons across versions. In particular, a four-level scheme in the predecessor makes it inaccurate to say that proficiency gradation as an idea began with the December consultation, but it does not license a conclusion about how many levels December used or what each required. Those exact questions remain dependent on direct document comparison.

Sources: AI Skills for Business Competency Framework (2023 public consultation draft); AI Skills for Business Competency Framework | Zenodo (predecessor record)

How does the framework structure a capability conversation?

The framework gives a team a sequence for describing capability in context: start with a persona and role setting, identify the duties in that role, connect each duty to competencies, then express the relevant Knowledge, Skills and Behaviours (KSBs), lifecycle phase and proficiency level. The official programme page, “AI Skills for Business Competency Framework & Resources,” describes these elements as parts of the framework. The structure helps specify what work is in scope, what capability a duty calls for, and how deeply that capability is expected to be demonstrated.

A persona sets the broad role context. It can help a manager or worker decide whose work is being discussed and avoid treating every employee as if they had the same responsibilities. A duty then narrows the discussion to a responsibility or unit of work associated with that role. In practical terms, moving from role label to duty asks, “What does this role actually need to accomplish?” That is a more useful starting point for a capability conversation than asking whether someone generally needs to become “good at AI.”

A competency links a duty to a capability area. The KSB statements break that area into components: what a person needs to know, what they need to be able to do, and what behaviours matter when applying that knowledge and skill. These categories can make a broad label more discussable. For example, a team can ask whether its concern is understanding a concept, carrying out a task, or applying judgment and responsibility while doing it. This is an interpretation of the framework’s design, not evidence that teams using it achieve better results.

Lifecycle phase adds a process location: it can distinguish where in an AI-related activity or system lifecycle a capability applies. Proficiency adds expected depth. Together, they help distinguish a capability needed for awareness of a process from one needed to perform or oversee work at a higher level. The practical implication is that two roles may share a competency label while needing different KSBs, lifecycle involvement or proficiency. The framework provides a vocabulary for making those distinctions explicit; it does not itself establish the level required in a particular employer’s job.

That boundary matters for career planning. A framework entry can help someone formulate a question about a task and the capability it appears to involve. It cannot show whether the worker’s employer expects that capability, whether AI is currently used in that workflow, or whether an AI tool can perform the task reliably in that workplace. Nor does a competency label identify which course or certificate would produce readiness. Those answers require local information about actual duties, approved tools, standards, supervision and available learning support.

More explicit links from roles to duties and capabilities may make a planning discussion concrete because participants can point to a particular responsibility and ask what knowledge, action and behavior it calls for. That is a reasoned implication of the structure, not a usability result reported by “AI Skills for Business Competency Framework & Resources.” There is also a tradeoff: a detailed taxonomy can create more distinctions than a small team has time or expertise to apply consistently. The framework describes an organizing mechanism; the source does not compare its usability with simpler alternatives or measure decision quality.

A worker can therefore use the structure as a prompt for investigation rather than as a verdict. Take one recurring duty, identify the competency and KSBs that seem relevant, then check the fit against the work actually done and the organization’s expectations. If the mapping is unclear, that uncertainty is useful: it signals a need to clarify the task, standard or local use of AI before selecting training. The framework can organize those questions, but evidence about the workplace must supply the answers.

Sources: AI Skills for Business Competency Framework & Resources | Innovate UK Business Connect

What does UK business adoption show—and what does it not show?

Official UK business surveys show adoption rising across two snapshots, but the measures are not a clean time series and neither gives an individual worker’s probability of displacement. Here, adoption means a business reporting that it uses AI in its methods or processes. The Office for National Statistics (ONS) reported 9% of sampled firms using AI in 2023; its late-September 2025 Business Insights and Conditions Survey (BICS) reported 23% of surveyed businesses currently using AI. The populations, question wording and survey designs differ, so the figures indicate separate snapshots, not a precise comparable growth rate.

The 2023 result comes from the ONS report “Management practices and the adoption of technology and artificial intelligence in UK firms: 2023.” Its Management and Expectations Survey ran from November 2023 to March 2024 and sampled about 55,000 businesses. The analysis covered production and services businesses with at least 10 employees, excluded agriculture, forestry and fishing as well as financial and insurance businesses, and weighted its results. The reported 9% therefore describes that defined business population and survey question; it is not a count of workers who personally used AI or of tasks technically exposed to it.

In the same 2023 survey, firms reported practical barriers: 39% cited difficulty identifying a use case, 21% cited cost, and 16% cited a lack of AI expertise or skills. These responses help explain why technical capability does not automatically become routine workplace use. They also suggest that adoption depends on organizational choices and constraints, not simply on whether a tool exists. The ONS analysis reports variation by sector and firm characteristics, but those aggregate patterns do not identify the conditions in a particular team or establish which occupations experienced a change in work.

The later ONS snapshot uses BICS rather than the Management and Expectations Survey. The bulletin “Business insights and impact on the UK economy: 2 October 2025” reports Wave 141 findings for late September 2025, when 23% of surveyed businesses said they were currently using AI. That percentage concerns businesses answering a current-use question in a different survey instrument. It should not be subtracted from or directly compared with the 2023 estimate as if the same firms had answered an unchanged question under an identical sampling frame. Its value is as a dated signal of reported business use in that survey.

The BICS bulletin also asked AI-using businesses about workforce outcomes. Among businesses already using AI, 4% reported that overall headcount had decreased as a result. The denominator is therefore businesses reporting AI use, and the outcome is a business’s report about its overall headcount. It does not say that 4% of workers lost jobs, identify which occupations or tasks changed, or estimate an individual employee’s chance of redundancy. It is an observed business-level response in one survey snapshot, not a causal estimate for a particular role.

Read alongside the 2023 barriers, the 2025 figure offers context rather than a guarantee. Reported use can coexist with limited workforce change at the whole-business level, while the survey does not measure every subtler response: a task may be reorganized, a vacancy may not be filled, or demand may shift without an immediate reported net headcount reduction. Those are possible distinctions in what a business-level headcount question can capture, not outcomes established by the 4% result. Equally, a business reporting AI use does not prove that a specific employee’s work is affected.

For a worker, the useful inference is narrow: UK business use was reported by a minority of firms in the 2023 MES and by a larger share in the late-September 2025 BICS snapshot, while reported headcount reduction remained uncommon among the AI-using businesses in that latter survey. The surveys do not establish the level of AI use in the worker’s occupation, the quality or autonomy of a system, or how an employer may redesign one role. Those questions need task-level and workplace evidence rather than a national business percentage.

This distinction qualifies how to use a role-and-duty framework. The framework can help specify which capabilities relate to a duty, but business adoption data cannot fill in local expectations or prove displacement. Conversely, low reported aggregate headcount reduction should not be used to dismiss a worker’s observed task change. A sound decision keeps the levels separate: survey results describe businesses, exposure analysis concerns tasks or occupations, and an individual’s next move depends on the actual work, local adoption, and practical constraints. The ONS snapshots inform that assessment without resolving it.

Sources: Management practices and the adoption of technology and artificial intelligence in UK firms: 2023 | Office for National Statistics; Business insights and impact on the UK economy: 2 October 2025 | Office for National Statistics

What do hiring signals say about demand?

The UK government’s Assessment of AI capabilities and the impact on the UK labour market reports an association between occupational AI exposure and lower job-posting volume in one analysis: a one-standard-deviation increase in exposure was associated with 3.9% fewer postings. The result is a hiring signal, not evidence that AI caused a corresponding fall in employment. It describes advertised openings across occupations; it does not count filled jobs, establish how many people were hired or made redundant, or estimate an individual worker’s chance of displacement. The number is therefore relevant to labour demand, but its meaning depends on the measure and comparison behind it.

The assessment presents the posting change as occurring over a period when postings in more exposed occupations fell relative to the comparison, then later moved back toward baseline. That timing complicates a simple forecast of lasting contraction: a temporary divergence in advertisements is different from a persistent reduction in employment. It does not make the initial movement disappear. Employers can adjust recruitment before changing existing headcount, and a short-lived change in advertised demand may matter to people entering a field or seeking a new role. The assessment’s point is narrower: the observed pattern did not continue as a steadily widening gap in that analysis, so the initial association cannot by itself support a claim of continuing decline.

The government assessment also cautions that the pattern cannot be attributed confidently to AI. Occupations with higher measured exposure may differ in other ways, and the assessment discusses broader economic conditions as a competing explanation. In particular, exposed occupations overlapped with sectors affected by interest-rate changes. If sector composition and macroeconomic conditions move alongside exposure, a posting gap may reflect some combination of those conditions rather than the effect of AI alone. The assessment does not establish how much each explanation contributed. This is an attribution problem, not proof that AI had no influence: the observed association is compatible with an AI-related hiring adjustment, but it does not isolate that effect from the alternatives.

There is a second measurement boundary. A posting is an employer’s advertisement for an opening, not a direct record of a completed hire or of the stock of people employed. Employers may advertise fewer roles while retaining current staff; they may also recruit differently without changing total employment. Conversely, a change in postings that later reverses could still have affected the timing or availability of opportunities during the period. The assessment’s finding can therefore inform a discussion of hiring conditions, but cannot resolve whether a particular task will be automated, whether a worker’s role will be redesigned, or whether a vacancy will ultimately be filled. Those are distinct outcomes and require evidence beyond the posting series.

The comparison also does not tell a worker what would have happened to postings in those occupations over the same period if measured exposure had been different. That counterfactual is central to a causal claim, and the assessment’s attribution cautions leave it unresolved. A percentage association can summarize how two measured patterns moved together under the study’s approach; it is not an estimate of how many openings AI removed. Nor does a standard-deviation change translate directly into a particular employer’s decision or a specific worker’s task mix. It also cannot identify whether any posting change came from fewer new roles, replacement hiring, or another shift in recruitment. The estimate should be read at its reported occupational and study level, with its timing and competing explanations attached, rather than carried over as a portable forecast for an individual career.

For someone using the framework to plan learning, keep this demand signal beside the role-and-duty map rather than treating it as part of the competency labels. The framework describes capabilities associated with duties; the posting analysis describes an association between occupational exposure and advertised openings over a particular period. Neither substitutes for local information about the actual role, employer plans, or current vacancies. The 3.9% estimate is worth taking seriously as a reported short-term market signal, while its later return toward baseline and the assessment’s alternative explanations limit what can be inferred from it. The evidence does not establish that exposure caused weaker labour demand overall, and it does not establish that no short-term adjustment occurred.

Sources: Assessment of AI capabilities and the impact on the UK labour market | GOV.UK

How is the England foundation benchmark different?

Skills England’s AI foundation skills for work benchmark is a separate, England-scoped resource. The GOV.UK publication describes basic technical and non-technical skills, including responsible and ethical use, for working with simple AI tools. Its stated scope is foundational capability: what someone may need to use straightforward tools at work. That makes it different in geography and depth from the AI Skills for Business Competency Framework, which organizes capability through organizational personas, duties and competencies. One describes a basic foundation; the other offers a role-and-duty structure for discussing more contextual requirements.

The Skills England benchmark should not be read as a universal standard for every UK nation or as a complete specification of what a particular occupation requires. The publication identifies England as its application, so readers elsewhere should not assume that it sets their national policy or training requirements. Nor does a basic benchmark decide whether a specific duty in someone’s job calls for deeper technical knowledge, oversight, judgment or accountability. The business framework’s role-and-duty organization can help make those contextual questions more specific, while the foundation benchmark can describe simpler capabilities that may be relevant across work settings. The sources do not establish a single progression that every worker should follow.

The GOV.UK page says industry partners use the foundation benchmark to design AI foundation courses. That is a description of its intended use, not an evaluation showing that courses built from it improve work outcomes, employment prospects or productivity. The benchmark can still be useful as a reference for basic learning needs; the limitation is about what the cited page demonstrates. Similarly, the business framework explains a way to organize role-related capability, but its structure alone does not establish training effectiveness. The two resources can complement one another when a worker needs both simple-tool foundations and a clearer account of capabilities tied to a particular duty. The available sources do not test which is more effective.

The practical distinction is therefore one of the question being asked. If the question is whether someone needs a starting point for using simple AI tools responsibly, the foundation benchmark is scoped to that kind of basic capability in England. If the question is what capabilities connect to the duties of a particular business role, the role-duty framework provides a different organizing structure. A basic foundation does not settle role-specific expectations, and a role map does not itself establish that a worker has mastered the basics. Treating the resources as complementary references preserves those different uses without turning either into a universal checklist or an evaluated route to a job outcome.

Sources: AI foundation skills for work | GOV.UK

Is the December consultation still the version to use?

No. The December consultation is a historical version to consult when you need to understand the proposal and its stated revision themes; for current wording, start with the published Version 3.0.0 file. The Alan Turing Institute’s “AI Skills for Business Competency Framework” page records that consultation responses closed on January 19, 2026. The Zenodo record “AI Skills for Business Competency Framework | Zenodo (Version 3.0.0)” is dated March 9, 2026. The sequence matters: a consultation document records material put forward for comment, while a later publication is the reference to use when describing the framework as released. Neither date alone tells you the fate of a particular statement.

If a decision turns on whether a duty or Knowledge, Skills and Behaviours (KSB) statement changed, compare the actual text in both documents. First identify the specific role, duty, competency and KSB relevant to the work; then locate the corresponding passage in the December consultation and in the Version 3.0.0 published file. Compare the wording and its surrounding labels, including any proficiency level or lifecycle context that affects how the statement is read. Record the document version and page or section for each passage so that a team can revisit the comparison. If the structure or labels differ enough that the passages are not clearly equivalent, mark the match as uncertain instead of treating a similar heading as proof of continuity.

This method answers a narrower question than asking whether the framework changed overall. The official consultation summary and the later release record establish that a later version exists, but the records cited here do not supply a complete consultation-to-release diff. They therefore cannot establish that an individual entry was retained unchanged, revised, removed or newly added. A consultation theme may have carried through, been edited in response to feedback, or been dropped; only comparison of the relevant source passages can resolve that point. Where the files do not expose a clear match, report the uncertainty and seek clarification from the publisher before relying on an exact revision claim.

For ordinary orientation, readers can use the published file without reconstructing the entire history. Keep the consultation as historical context when the question concerns what was proposed or discussed, and cite the release when stating current framework wording. For a training specification, internal role profile or other decision that depends on precise wording, preserve both versions and make a focused comparison before changing requirements. This is a verification rule, not a claim that every post-consultation detail is materially different. Version succession establishes which file is later; it does not reveal the editorial path between them.

Sources: AI Skills for Business Competency Framework | The Alan Turing Institute; AI Skills for Business Competency Framework | Zenodo (Version 3.0.0)

What learning move follows without overcommitting?

Illustration, not a reported job study: imagine an analyst who repeatedly prepares a monthly performance summary. The inputs might be approved spreadsheets; repeatable steps might include cleaning fields, calculating totals and drafting a first-pass narrative. The judgment lies in choosing meaningful comparisons and explaining anomalies; accountability includes checking figures and ensuring the summary is fit to share. A relevant framework duty or capability can help name the work to discuss, but its label does not show whether AI is used locally or what standard the employer requires.

A bounded next step is to ask a manager or team which tools are approved, what review is expected, and whether protected practice time is available. If support exists, compare a small guided exercise on one low-consequence step with a larger course by relevance, prerequisites, feedback, time and cost, alongside health and family constraints. The framework does not establish that a credential is necessary or sufficient. If policy, time or task allocation is outside the worker’s control, unpaid study or a major career move need not be the default. First observe the real task mix and local adoption; choose learning only when it addresses a specific gap the workplace can use.

Questions readers ask