AI Skills Budgets Need Workflow Targets Before Training Starts

ROI & Cost Optimisation

3 October 2026 | By Ashley Marshall

AI training is becoming a default budget line. The harder question is whether anyone can name the workflow, behaviour and commercial result that the training is meant to change.

The economic case is large, but it is not a cheque to buy courses

The latest UK evidence gives leaders a serious reason to invest in AI capability. Learning and Work Institute, working with the Rigby Foundation, estimates that comprehensive AI upskilling could support an economy-wide productivity uplift of 2.3% by 2035 and create a net present value of £80 billion. Its research included a YouGov survey of 1,012 UK HR decision-makers. In that survey, 85% said their organisation used AI in some capacity, but only 18% used it extensively. The gap between access and extensive use is where most of the commercial work now sits.

That headline does not mean any training programme will deliver a return. It means skills are one necessary part of a much larger adoption system. The same Learning and Work Institute research found that 63% of employers believed stronger leadership and management skills would help them adopt AI more fully. That finding should change who owns the budget. AI upskilling cannot be left solely to learning and development, because managers choose priorities, redesign work, remove blockers and decide whether saved time is converted into better service, lower cost or additional capacity.

What this means in practice is simple. Before approving a training budget, ask for a one-page investment case naming the target teams, the workflows they will improve, the measures that will move and the senior owner who will act on the results. A customer service course might target faster case preparation without lowering resolution quality. A finance programme might target shorter month-end commentary preparation with a documented review step. A sales programme might target better account research while prohibiting unverified claims. If a proposal contains only learner numbers, course hours and satisfaction scores, it measures activity rather than value.

The UK's adoption data exposes the training gap

Official statistics show why generic awareness sessions are unlikely to be enough. The Office for National Statistics reported that AI use among UK businesses with 10 or more employees rose from around 12% in late 2023 to around 35% by June 2026. Yet the average number of AI technologies used by adopting businesses increased only modestly, from about 1.4 to 1.6. Only 10% of adopting businesses reported extensive use, and only 11% of businesses said more than half of their workforce had received AI-related training.

This is not primarily an access problem. Many staff already have access, including through tools embedded in Microsoft 365, Google Workspace, CRM systems and specialist software. It is a depth problem. People test summarisation, drafting or search, but the organisation does not convert those experiments into a repeatable way of working. The ONS also found that 55% of employees reported using AI for work or education, compared with 35% of businesses reporting formal use of at least one AI technology. Informal behaviour is moving faster than operating design.

The practical response is to map capability by workflow, not by job title or licence count. Ask each team to identify three recurring tasks where AI is already used, three where it could help, and three where it should not be used without additional controls. Record the current time, error rate, rework, escalation rate or customer outcome for each selected task. Then train against those real examples. This creates a baseline and prevents an enthusiastic minority from defining success for everyone else.

A common counterargument is that measurement will slow adoption. In reality, a lightweight baseline speeds useful adoption because teams stop debating vague productivity claims. Five observed cases before training and five after training can reveal whether a workflow is promising. The goal is not laboratory certainty. It is enough evidence to decide whether to stop, adjust or scale.

Start with workflow outcomes, then design the learning

Most training procurement starts in the wrong order. A provider presents a catalogue, leaders choose a course level, staff attend, and the business later asks what changed. Reverse that sequence. Choose a workflow with enough volume to matter, define a safe improved method, and build learning around the decisions people must make inside it.

For example, an accountancy team might use AI to produce a first pass of variance commentary. The target is not better prompting. The target might be reducing preparation time from 90 minutes to 45 minutes while keeping all figures traceable to an approved source and requiring a named reviewer. A recruitment team might use AI to structure interview notes. Its target could be faster administration with no automated candidate scoring and a clear retention rule. A professional services team might improve proposal research, with citations checked before any claim reaches a client. These are observable operating outcomes.

The Government's Skills England employer guide supports this applied approach. Its evidence base included 23 workshops, 10 case studies and 536 survey responses. The guide says effective training should be practical, reachable, integrated, modular, expandable and sustainable, summarised as the PRIMES framework. It also reports that 51% of organisations identified gaps in training flexibility, 34% in practical contextualised learning and 35% in clear AI skills frameworks.

Turn those principles into a simple training brief. Include one real workflow, approved tools, permitted data, a worked example, a deliberately flawed output, a checking routine and a pass condition. Learners should leave with an artefact they can use, such as a review checklist, an approved prompt structure or an exception route. Completion should require demonstrating the workflow, not merely watching content. This design costs more effort than buying a generic webinar, but it is much easier to connect to operational value.

Protected practice time is part of the investment

Leaders often budget for course fees and licences while treating employee time as free. That produces a misleading return calculation. If 50 people attend two hours of training, complete follow-up exercises and need manager support, the organisation has committed a meaningful amount of paid capacity. It should record that cost and create the conditions for the time to produce a result.

Protected practice matters because people do not build judgement from a slide deck. They need to try an approved tool on realistic material, inspect weak outputs, compare approaches and receive feedback. Skills England's guide identifies cost as a participation barrier for 42% of employers and limited availability as a barrier for 37%. It explicitly recommends paid or protected time, practical scenarios and support for people with different levels of digital experience. Those recommendations are operational requirements, not optional learning features.

A useful pattern is a four-week workflow sprint. Week one establishes the baseline and covers safe use. Week two gives participants a controlled task using non-sensitive or approved data. Week three tests the method on live work with review. Week four compares outcomes, documents exceptions and decides whether the workflow should become standard. Managers should reserve perhaps 60 to 90 minutes each week and cancel lower-value work to make room. Adding training on top of a full workload encourages rushed practice, hidden use and abandonment.

What this means in practice is that the budget should contain four lines: provision, learner time, manager or reviewer time, and workflow change. The last line might include template updates, access controls, integration work or quality sampling. It is perfectly reasonable for those internal costs to exceed the course fee. They are the mechanism through which learning becomes a business capability.

Smaller firms can keep this proportionate. One team, one workflow and one month is enough to learn. A tightly scoped pilot with honest cost tracking is more valuable than an organisation-wide AI day that creates enthusiasm but no durable change.

Measure behaviour, quality and value together

Training dashboards usually stop at attendance, completion and learner confidence. Those measures can help diagnose delivery, but they do not prove that work improved. A useful AI skills scorecard needs three layers: behaviour, quality and value.

Behaviour measures whether the new method is actually used. Track the proportion of eligible cases using the approved workflow, the number of active users, the use of required review steps and the rate of exceptions. Quality measures whether outputs remain reliable. Depending on the workflow, this could include factual correction rates, customer complaints, manager rework, missed citations, policy breaches or escalation frequency. Value measures the business result: time released, cycle time, conversion, avoided external spend, capacity created or risk reduced.

The categories must be read together. High usage with rising rework is not success. Time saved without a plan for that capacity is not automatically a cash benefit. Better quality may justify the investment even when speed is unchanged, but leaders should name that objective in advance. For financial reporting, distinguish between theoretical time saved, usable capacity and realised value. Thirty minutes saved across a team becomes valuable only when that time is consistently available and redirected to work the business needs.

This point matters because the Government's AI Adoption Plan for Professional and Business Services describes a gap between bottom-up use and top-down transformation. It reports that 69% of firms had expanded generative AI training, while three-quarters were not ready on core enablers such as data, orchestration and monitoring, and 70% had made limited progress on process redesign. Training can run ahead of the systems needed to turn individual gains into firm-wide performance.

Review the scorecard at 30, 60 and 90 days. Stop programmes that have no credible workflow pull. Adjust those that show adoption but weak quality. Scale only when the method, controls and ownership are repeatable. That discipline protects the budget and gives staff confidence that the organisation values useful capability over performative activity.

Build a portfolio of skills investments, not one universal programme

The strongest objection to workflow-specific training is that it appears fragmented. Leaders may prefer one standard course for everyone because it is easier to procure and report. A shared baseline still has value. Every user should understand confidentiality, verification, human accountability, approved tools and escalation. But a universal programme cannot teach every team how to redesign its work, because the decisions, risks and evidence differ.

The better model is a portfolio. Create a short mandatory foundation for all users, role-based modules for recurring activities, and deeper pathways for champions, managers, technical specialists and assurance owners. Connect each module to a capability level and an approved workflow. Refresh content when tools, policies or risks change. This balances consistency with practical relevance.

Businesses should also resist treating confident early adopters as automatic trainers. Tool fluency is not the same as the ability to design safe work or teach colleagues. Select champions who can explain limitations, follow governance, observe where people struggle and feed evidence back to process owners. Give them time and a clear remit. Their job is to improve the system, not to promote unrestricted experimentation.

For organisations that already have plenty of licences but shallow use, the next step is not another broad awareness event. It is an adoption review. Compare formal workflows with what employees are doing, choose two or three high-volume opportunities, and fund the skills, process changes and controls together. The depth of AI adoption is a more useful management signal than the number of accounts activated.

The decision rule is straightforward. Approve AI training when there is a named workflow, a baseline, protected practice, a quality check and a benefit owner. If those elements are missing, pause the purchase and design the operating change first. The £80 billion opportunity described by Learning and Work Institute depends on capability becoming productive action. Businesses will capture their share by treating learning as part of workflow transformation, not as evidence that transformation has already happened.

Frequently Asked Questions

How much should a UK business budget for AI training?

Start with the full cost of one workflow sprint: provision, paid learner time, manager review, process changes and any technical controls. A small team pilot is usually more informative than setting an arbitrary per-person course budget.

What is the best first metric for AI upskilling?

Use an operational metric tied to the selected workflow, such as preparation time, rework, factual correction rate or escalation rate. Course completion should be a delivery measure, not the main success measure.

Should every employee receive the same AI training?

Everyone using workplace AI should receive a common foundation covering safe use, verification and accountability. Practical modules should then reflect each role's workflows, data, decisions and risks.

How quickly should AI training show a return?

A tightly scoped workflow should produce useful evidence within 30 to 90 days. Some benefits, especially quality and risk reduction, may take longer to realise, but the leading measures should move during the pilot.

Does time saved count as financial value?

Not automatically. Record theoretical time saved separately from usable capacity and realised value. The benefit becomes credible when the saved capacity is consistently redirected, demand is absorbed or paid cost is avoided.

Can free AI courses be enough for a small business?

They can provide a useful foundation. The business still needs to connect the learning to approved tools, real tasks, data rules, checking routines and an owner who will decide whether the workflow should scale.

Who should own an AI upskilling programme?

Learning and development can coordinate delivery, but an operational leader should own the workflow outcome. Information security, data protection and subject experts should contribute where the use case creates relevant risks.

What should make us stop an AI training pilot?

Stop or redesign when staff do not use the workflow, quality declines, review costs erase the benefit, required data cannot be handled safely or no owner will act on the evidence.