A Shift Roster With No Time to Train
On August 6, OpenAI said ChatGPT had one billion weekly users. In the same product update, it said Free and Go users would receive unlimited GPT-5.6 Luna text chats the following week, subject to abuse controls. Limits would remain on file uploads, images, and other tools.
A separate OpenAI usage release that day said people using ChatGPT at work were more than twice as likely to create something or complete a task as people using it outside work. Its dataset stops at individually managed Free, Go, Plus, and Pro accounts. Employer-managed accounts sit outside it. On the next morning’s roster, a worker could have more access on a personal account while the manager still had no paid practice hour or cover scheduled.
Britain already has evidence of that gap. The Office for National Statistics reported in July that 55 percent of surveyed workers used AI for work or education, while 35 percent of businesses with at least 10 employees reported using an AI technology. The figures come from different surveys and definitions, so they cannot be subtracted into a precise shadow-use rate. They do show that individual access and formal company adoption are moving on separate tracks.
Training coverage is narrower. Only 11 percent of businesses with at least 10 employees said more than half of their workforce had received AI-related training. Among firms that had adopted AI, 15 percent said more than half of employees used it in daily work. Only 10 percent described their AI use as extensive.
The budget underneath that transition has been moving in the other direction. A July 27 report from the UK’s Industrial Strategy Advisory Council found that real employer training expenditure in 2024 was 19 percent below its 2011 level.
England changed the price and timing of one public training route on August 1. Employers can now use Growth and Skills Levy funds for short apprenticeship units, including three AI leadership subjects. At the same time, the 10 percent government top-up on new funds ended, new funds began expiring after 12 months instead of 24, and the co-investment rate for workers over 25 rose when an employer exhausts its levy balance.
Access to AI is becoming abundant. Open the operating budget, though, and the account or license may be the cheapest item. Someone must release a worker to learn, cover the work, prepare a safe task, and review the result. A free course does not supply that hour.
Britain’s AI skills debate ends up on the shift roster. The budget has to include protected learning time and manager attention, then connect practice to a changed job. If it does not, the employee with the easiest schedule becomes the unofficial AI expert while everyone else learns between tasks.
At work before the training plan
OpenAI’s August product change lowers an access barrier that employers do not control. Unlimited text chats on a free individual account make experimentation easier for a shop supervisor, a care coordinator, an accounts clerk, or a junior analyst. The product update made no workplace claim. The usage report documents work tasks within individually managed account data without counting work users or establishing which employers approved the activity.
The ONS data offer a national view with clear limits. Its Business Insights and Conditions Survey is voluntary. Wave 159 received responses from 38,637 businesses, a 26.7 percent response rate, and excludes industries including finance and insurance, public provision of health and education, public administration and defence, agriculture, and energy supply. ONS calls the results official statistics in development.
Within that frame, adoption has risen quickly. The share of businesses with at least 10 employees reporting at least one AI technology climbed from about 12 percent in late 2023 to about 35 percent in June 2026. Yet the average number of technologies used by an adopting firm moved only from roughly 1.4 to 1.6. ONS described adoption as relatively shallow.
Company size changes the view. Twenty-eight percent of businesses with fewer than 10 employees reported AI use, compared with 49 percent of businesses with at least 250 employees. Sector differences were wider: 58 percent in information and communication versus 13 percent in construction. An online course offered to the whole country lands in workplaces with very different devices, connectivity, margins, data obligations, and available support.
Free software was a common route into manufacturing, wholesale and retail, hospitality, and administrative services. Construction, information and communication, and professional, scientific, and technical firms were more likely to buy external software or ready-to-use services. A personal account can sit beside a corporate system, or fill a gap where the employer has approved nothing.
A line manager may discover that customer notes are reaching a personal account. The immediate choice is concrete: stop the task, move it to an approved system, or create a safe example for practice. Blanket prohibition can hide the activity; open permission leaves the employee without a data rule or reviewer.
The 55-to-35 spread is a warning light, not a calculated deficit. The worker survey includes work or education and may capture informal use; the business survey asks firms about specified technologies in business operations. Those reporting boundaries explain some of the difference. They also leave managers with a basic unknown: which AI already sits inside ordinary work?
A useful discovery exercise starts below the license count. Ask people which tasks they already give to an AI system, which account they use, which information they enter, whether the output reaches a customer or colleague, and where they check it. Do this without turning the first conversation into a disciplinary sweep. Employees will hide experiments if disclosure feels like confession.
Answers to that inventory determine who trains first. A call-centre team rewriting internal notes needs different practice from an engineer reviewing generated code or a care worker handling personal records. General literacy gives them common language; the job itself supplies the test.
A smaller training budget for a larger task
The Industrial Strategy Advisory Council’s July report arrived with a long view of employer investment. UK employers spent GBP53 billion on training in 2024, down 10 percent in real terms from 2022 and 19 percent from 2011. The proportion of employers funding or arranging training fell from 65 percent to 59 percent over the same longer period.
Council project sponsor Phil Smith described the choice he heard from businesses as a struggle to bring in people with the right skills or retrain those already at work. His formulation puts recruitment and internal development on the same budget. Cutting one does not make the other easier when many employers compete for the same occupations.
Small firms do not explain the decline by themselves. Ninety-nine percent of large employers provided training, compared with 61 percent of micro firms, yet businesses with at least 100 employees accounted for about 58 percent of the fall in expenditure. The report says they trained more people for fewer days, perhaps through shorter courses or cheaper online formats.
Lower recorded spending does not prove that capability fell at the same rate. Online delivery, vendor support, and peer learning can reduce the cash price or sit outside a formal course budget. The council’s finding about fewer training days raises a narrower concern: broad participation may coexist with less time for practice.
An AI foundation module can therefore raise a company’s completion count while practical time shrinks. Thousands of people may finish 20 minutes of content, but the supervisor still has no hour to review a live use case.
Skills England identifies a larger demand horizon. Its 2026 annual report groups 150 priority occupations across industrial-strategy sectors, construction, health, and social care. Those occupations represented 7.6 million workers in 2025. The report projects 9.5 million by 2035, an increase of 1.8 million or 23.8 percent. Non-priority occupations are projected to grow 14.4 percent.
The 1.8 million figure is a forecast of additional employment, not a count of empty jobs today. Sector models use different methods, sector definitions overlap, and the report says the estimates indicate scale rather than precise outcomes. Sixty-two percent of the additional demand is expected to be filled by workers holding qualifications at Level 4 or above. The remaining 38 percent is expected at Levels 2 and 3, including roles in construction, clean energy, and care.
There are counter-signals to a simple shortage story. The share of vacancies constrained by skills fell from 36 percent in 2022 to 27 percent in 2024. Achievements in training routes related to priority occupations increased 18 percent between the 2021/22 and 2023/24 academic years, an annualized rate near 8 percent against projected occupational growth of 2.2 percent a year. Skills England estimates that about 230,000 recent education leavers in England were working in priority occupations in 2022/23; 74 percent held a Level 4 or higher qualification.
The supply picture is mixed. Further-education routes tied to priority occupations declined while apprenticeships and higher education grew. A national total can improve while a manufacturer, care provider, or small contractor searches in vain for a local course that fits its work and shift pattern.
The council’s immediate evidence is more specific. Eight industrial-strategy sectors had 85,600 skill-shortage vacancies and 349,000 employees with reported skills gaps. About one in 10 firms in those sectors said skills shortages reduced their ability to adopt a technology or process, or to innovate. Among firms already innovating, 40 percent cited skills as a barrier in 2025.
The report’s GBP9 billion upside comes from a model. Researchers estimated the cumulative gross value added from 2026 through 2029 if 10 percent of workers with reported gaps in selected sectors were upskilled. Financial services and life sciences are excluded, and the calculation depends on historical relationships. A board can use it to size a possibility, never as a promised return.
In the council’s linked firm data, firms in the top productivity quartile were 5 to 18 percentage points more likely than the bottom quartile to invest in management training, depending on sector. The data end in 2019 and establish an association, not causation. Their relevance to AI lies with the manager who grants time, gives feedback, and permits a process to change.
Skills England separately reports GBP44.8 billion of employer training expenditure in England in 2024, down from GBP49.4 billion in 2019 at 2024 prices. The council’s GBP53 billion covers the UK, so the two figures cannot be combined. Both series indicate declining real investment, and neither isolates AI training.
Beyond the completion badge
Skills England’s AI upskilling work makes protected time explicit. Its July employer guide says training should be accessible by time, format, and language, and that learners should receive paid or protected time. It also recommends baseline training before staff use workplace AI with confidential data, regulated activity, safety responsibilities, or professional judgement.
The guide draws on 23 workshops, 10 case studies, and 536 survey responses, a much smaller base than ONS uses for national prevalence. In that sample, 97 percent of organizations offered some AI training. Still, 51 percent reported gaps in flexibility and 34 percent reported gaps in practical or contextual learning.
For a shift worker, fit begins with the roster. A course assigned outside paid hours transfers the cost to the employee. A course scheduled during work without cover transfers the cost to customers and colleagues. A short module may reduce either burden, but brevity does not create practice time by itself.
The manager needs to know what stops while learning happens. On a production line, the answer may be a relief worker or a lower planned output for an hour. In a call centre, it may be a smaller queue assignment. In a clinic, it may require protected administrative time and a training environment without patient data. In a small office, it may mean the owner handles incoming work while two people practise.
Unequal schedules can produce unequal careers. Salaried staff with flexible calendars can experiment, attend office hours, and join an internal community. Hourly staff, part-time workers, field employees, and people with caring responsibilities may see the same course link but have fewer chances to use it. If promotion later rewards AI fluency, the company has converted schedule flexibility into a career advantage.
Workers may also wonder whether a course leads to a better job or helps automate the current one. ONS found that training or retraining existing staff was the most common reported response to AI skills needs, while around 10 percent of businesses reported automating or replacing roles as an integration approach. An employer should tell a cohort which tasks are expected to disappear, which responsibilities may grow, and how pay or progression will be handled.
A badge records completion. Evidence of judgement comes from a real exception: an unsupported answer caught, customer data withheld, or a process improved without shifting errors to someone else. Most people do not need to become model specialists. They need the judgement their role demands.
One operating record can keep the budget honest:
| Decision | Minimum evidence before rollout | Cost often left out | Worker outcome to follow |
|---|---|---|---|
| Give tool access | Named tasks, data rules, prohibited uses, escalation owner | Setup, identity controls, support, incident review | Who receives access and who remains excluded |
| Schedule foundation learning | Role population, starting confidence, language and accessibility needs | Paid hours and shift cover | Completion by role, contract type, location, and shift |
| Move into practice | Safe sample task, reviewer, quality baseline | Manager review time and sandbox data | Ability to complete a task safely without hidden help |
| Change a workflow | Owner, customer impact, fallback, risk threshold | Process redesign, documentation, consultation | Task quality, workload, autonomy, and error recovery |
| Recognize capability | Observable skill standard and assessment | Calibration and appeals | Pay, promotion, internal moves, and retention |
| Refresh training | Tool change, incident pattern, or scheduled review | Repeated release time | Whether capability persists three and six months later |
Completion can be counted immediately. Quality, judgement, and mobility take longer. After practising, a team may reject the tool for a sensitive task; avoiding that error belongs in the return even though usage falls.
Managers need practice and time of their own because they set the quality bar, divide work, and coach failed results. Today’s service level appears on their dashboard while a worker’s capability six months later may not. Without relief in operating targets, postponing the cohort can look like sound short-term performance.
Inside the new levy window
England’s Growth and Skills Levy changed on August 1. New apprenticeship units let employers fund 30 to 140 delivery hours over one to 16 weeks for existing workers aged 19 or older. The initial list includes AI strategy and opportunity, AI adoption, procurement and governance, and AI delivery and organizational transformation.
The modular offer sits inside a much larger ambition. In January, Technology Secretary Liz Kendall announced free foundation courses and a goal to reach 10 million workers by 2030. The levy units address a different layer: employer-backed learning for a role, with funding and delivery conditions attached.
The format lets an employer fund a bounded capability without placing every incumbent worker in a long apprenticeship. A leadership team can select a modular unit tied to procurement, governance, or delivery, then fit it into a work calendar sooner than a multi-year qualification.
The funding mechanics became less generous at the margin. New funds entering an employer’s apprenticeship service account no longer receive the government’s 10 percent top-up. They expire after 12 months instead of 24. When a levy-paying employer exhausts its account, the co-investment rate for a new start involving a worker over 25 is 25 percent. Training for eligible workers under 25 is fully funded, though levy funds are used first if available.
In the government’s example, a GBP3,500 customer-service apprenticeship costs the employer GBP175 under the old co-investment rate and GBP875 under the new one after levy funds run out. It is an illustration of the marginal cost, not a price for an AI unit, and it does not apply while account funds remain.
The expiry change gives employers a shorter window to commit or use available funds. It does not set a course-completion deadline. The council’s report identifies complexity, policy churn, and navigation as barriers already faced by employers, so a shorter window may spur action or increase the cost of delay for organizations without a training department.
The government explains the wider changes as a refocus on young people entering skilled careers. Full funding for eligible workers under 25 supports that aim. Existing workers over 25 are also the people many employers need to retrain as AI changes established jobs. Once funds run out, the higher contribution creates a choice between entry investment and incumbent development. Age, domain knowledge, and the role transition belong in that review.
The changes apply to England’s funding system, not one uniform UK training regime. Employers operating across England, Scotland, Wales, and Northern Ireland need to separate eligibility and budget treatment. A UK-wide training promise can fail when local HR teams discover different funding routes after enrollment has begun.
Practice at Airbus and KPMG
Skills England’s company cases show how employers have joined learning to work. The collection labels them illustrative, and the participating organizations reported the outcomes themselves. None of the performance claims was independently audited, so the cases are better read for design than comparison.
Airbus built AI capability through a longer sequence. It began with data analytics and governance, established structured pathways, used communities of practice, and added AI competencies to workforce planning. Employees could add a competence to an existing role, move into a data or AI job, or change how their current job operated.
Access came with a gate. The case says employees completed mandatory training before using generative AI tools. The training covered what data could enter a system, alongside ethical and sustainability considerations. That approach makes governance part of adoption, rather than an annual compliance reminder detached from the tool.
In a pilot involving about 2,000 employees, primarily in office roles, Airbus reported savings of up to four hours a week for effective users. The cohort, tool, and training conditions make the upper-bound claim interpretable without turning it into an average or a result for production workers.
The internal mobility result may be more durable than the headline hours. Some employees moved into data analyst or data scientist roles. Others added data work to existing jobs. Airbus placed the learning architecture inside its annual competence-planning cycle, giving HR business partners and managers a way to connect training to recruitment, redeployment, and development.
KPMG UK took a different route for a professional-services workforce of roughly 17,000. It embedded an AI learning coach called Spark in Microsoft Teams, supported employee communities, and enrolled more than 130 people across grades in an AI-focused apprenticeship. The apprenticeship evaluated applied business outcomes, rather than course completion alone.
One learner reportedly built a tool that reduced an analysis task from 200 hours to four. Its baseline and visible work product give an assessor more to examine than a confidence score, although one project cannot support a portfolio forecast.
KPMG also gave early-career employees tool access and responsible-use guidance from their first week, while offering deeper pathways for people whose roles required them. More experienced staff could retrain as professional work changed. A common foundation did not require identical depth across every job.
Roche adds a regulated contrast. It tied mandatory training to approved tools, used local champions, and reinforced restrictions on sensitive data. In one function, around half of employees reported saving at least four hours a week. The approved access and data boundary are as informative as the reported hours.
Each employer put something before the course catalog: a business problem, a live project, or an approved system with a data boundary. Airbus can support that order with a digital academy, KPMG with a coach inside Teams, and Roche with a controlled global environment. A 40-person supplier faces the same sequence without the same fixed infrastructure.
Training infrastructure beyond one firm
The council’s central conclusion is that funding alone does not explain the decline in employer investment. Firms also face fragmented programs, changing policy, uncertain returns, lack of staff time, difficulty finding relevant provision, and weak coordination. These costs weigh heavily on smaller employers without dedicated HR or workforce-planning staff.
Poaching creates a rational hesitation. “I can spend three years training an apprentice and at any time they can walk away,” one SME employer told the council. The worker is entitled to move, and mobility spreads skills through the economy. The employer still faces a concentrated cost while other firms share the benefit. That gap encourages everyone to wait for someone else to train.
AI shortens the useful life of some content, adding another reason to delay. A provider needs time to design and approve a course. The tool interface, model behavior, or regulation may change before the cohort finishes. Employers may decide that informal experimentation is faster, even when it produces uneven capability.
Collective infrastructure can distribute the fixed cost. The council recommends employer consortia, stronger local brokerage, and large anchor employers that open suitable training, expertise, or facilities to suppliers. The goal is to aggregate demand and give smaller firms a route through the system without asking each owner to become a funding specialist.
Airbus offers a physical example at Filton. It partnered with Weston College Group and employers including Rolls-Royce, GKN Aerospace, MBDA, Babcock, and regional small businesses to establish the Advanced Manufacturing Technical Excellence College. The curriculum and facility serve a wider aerospace, defence, and manufacturing cluster, rather than one company’s vacancies.
Shared infrastructure does not remove the roster decision. A supplier still has to release a worker, cover production, and decide how the skill changes the job. It can improve course quality, reduce search cost, and let multiple employers support a cohort. Those changes make the release decision easier to justify.
Large employers have another lever: purchasing power. If an anchor firm expects suppliers to use an AI-enabled quality, planning, or reporting system, it can help fund the training needed to operate it. Procurement can reward a credible workforce plan. Without support, a new technology requirement becomes an uncompensated cost pushed down the chain.
Worker voice belongs in that arrangement. Employees know which part of the shift cannot be interrupted, which instruction conflicts with actual work, and where an AI output would create another checking task. Union representatives and frontline workers can help design access, scheduling, assessment, and transition rules before a rollout hardens into policy.
Small firms also need a proportionate measurement standard. They do not need a data warehouse to compare one task before and after training. They can record time, rework, customer complaints, safety exceptions, and the number of cases escalated to an experienced person. They should also record who received the opportunity. A productivity gain limited to the owner and two salaried employees is not workforce adoption.
Paid learning on the operating plan
A CFO can see the software invoice. Payroll, manager time, temporary cover, security, legal review, and process improvement sit elsewhere in the accounts. A 40-person customer-service team makes those hidden lines visible.
Before buying 40 seats, its manager can select one recurring task and record time, rework, customer effect, and escalation. The roster shows who works nights, who is part-time, and who already uses a personal tool. Three hours of learning per employee creates 120 paid hours before any manager reviews the practice. A quieter queue, scheduled cover, or a slower service target has to absorb them.
The task determines the depth. Agents drafting routine follow-up may need an approved tool, a data boundary, source checking, and a stop rule. The person redesigning the workflow needs repeated practice and authority to handle exceptions. Realistic examples can expose a planted failure before sensitive access opens.
Results may be untidy. Faster drafts can produce slower review, and an automated summary can send more exceptions to an experienced colleague. At three and six months, compare those effects with the baseline and ask who gained a new responsibility, pay adjustment, or coaching role. Check access by shift and contract type, not only company-wide completion.
New levy funds now have a 12-month life. Before a balance expires, the employer must decide whether to use it, choose a provider, and identify the role that will change. The funding route can cover eligible training and assessment. The next roster still has to show who gets the hour and who covers the work.