Practical guide
How an L&D leader can build the case for applied AI capability evidence
An L&D business case should compare realistic ways to answer a defined capability decision, including continuing current arrangements. It should not begin with a preferred assessment product and manufacture benefits around it. The learning and development leader may own programme design, influence investment, review evidence quality, or collaborate with finance, operations, technology,.
An L&D business case should compare realistic ways to answer a defined capability decision, including continuing current arrangements. It should not begin with a preferred assessment product and manufacture benefits around it. The learning and development leader may own programme design, influence investment, review evidence quality, or collaborate with finance, operations, technology, accessibility, data protection, and workforce representatives; local governance determines the role. Use BUILD to create a decision artifact: baseline the problem, understand users and constraints, identify alternatives, ledger costs and benefits, and define evidence and decision rules.
B — Baseline the capability problem
State the case for change in observable terms. Examples include inconsistent verification after AI training, a new tool rollout without evidence of applied use, or managers lacking a common language for feedback. Record:
Do not treat course completion, tool usage, confidence, assessment performance, and operational outcomes as the same construct. Each can support a different question.
The 2026 UK employer guide on AI upskilling recommends aligning learning with organisational goals, roles, starting points, practical work, inclusion, leadership, and evaluation. Its methodology also describes variation and limits in the available evidence. The business case should preserve those uncertainties rather than turn them into universal effect claims.
- the decision and accountable owner;
- affected work, population, and timing;
- current learning and evidence methods;
- direct evidence of the problem and its missingness;
- what happens if current arrangements continue;
- exclusions, such as employment decisions or broad workforce surveillance.
U — Understand users, governance, and delivery constraints
Map who is affected and who must contribute. Include participants, line managers, L&D operations, accessibility, technology, privacy, security, procurement, finance, and the decision authority.
For each group, capture the information they need and the risks they can identify. Practical constraints may include paid learning time, shift coverage, permitted tools, representative tasks, data access, accommodation, language, reviewer capacity, and evidence retention. A low licence price can still be a poor option if administration, scoring review, support, or workflow disruption is ignored.
Classify the L&D leader's authority explicitly. If finance owns investment approval or operations owns workflow standards, say so. The business case should make collaboration visible rather than implying that one role controls the whole decision.
I — Identify and narrow alternatives
Start with a broad option set, then use stated criteria to narrow it. Common alternatives include:
Compare alternatives on relevance to the decision, coverage of the capability construct, accessibility, administration, evidence quality, privacy, interpretability, scalability, and reversibility.
HM Treasury's 2026 Green Book is public-sector appraisal guidance, not a universal corporate template. Its useful transferable principle is to compare costs, benefits, risks, uncertainties, and alternatives—including business as usual—in a balanced way, and to plan monitoring and evaluation before approval.
- continue existing learning and manager observation;
- improve the current knowledge check or work-sample process;
- add a lightweight practice task with structured feedback;
- run a realistic simulation with process and deliverable evidence;
- combine participation, simulation, and later workplace observation;
- defer action while collecting better baseline evidence.
L — Ledger costs, benefits, risks, and uncertainty
Use ranges and evidence grades where precision is weak.
Costs can include design, licences, integration, participant time, manager time, reviewers, accommodations, support, security and privacy review, change management, monitoring, and evaluation.
Potential benefits should connect to the decision. Examples include better visibility into a defined behaviour, more targeted practice, a shared feedback vocabulary, or fewer unsupported inferences from completion data. Do not convert these into money unless the causal link, unit, baseline, volume, and time period can be supported.
Risks can include inaccessible tasks, unrepresentative scenarios, inconsistent scoring, missing evidence, inappropriate reuse of individual results, low participation, technical failure, and overinterpretation of early benchmarks.
For every material estimate, record source, date, owner, range, dependency, and confidence. Show which assumptions could reverse the preferred option.
D — Define the evidence plan and decision rules
Before approval, specify what the organisation will learn and how it will decide. The evidence plan should include:
The supporting UK AI-upskilling case studies provide practice examples, not controlled proof that the same outcomes will occur elsewhere. Use them to generate hypotheses and design questions. The local pilot or staged implementation must produce evidence for the organisation's own decision.
- a bounded population and representative work context;
- measure definitions and data sources;
- preset exclusions and missingness rules;
- accessibility and technical acceptance criteria;
- scorer or reviewer checks where applicable;
- comparison with current evidence;
- stop conditions and incident handling;
- reporting audiences and access boundaries;
- possible outcomes: stop, repair, retest, continue, or expand with conditions.
Business-case artifact
| Field | Required content | |---|---| | Decision | One named choice, owner, and date | | Baseline | Current method, observed problem, missing evidence | | Alternatives | Business as usual plus credible options | | Stakeholder roles | Owner, influencer, reviewer, collaborator | | Costs | Range, source, timing, and owner | | Benefits | Defined outcome, evidence route, and limitation | | Risks | Trigger, mitigation, residual uncertainty | | Evidence plan | Measures, exclusions, acceptance and stop rules | | Recommendation | Preferred next step and why alternatives were not selected | | Approval boundary | Named authority and conditions |
Worked example: verification practice after training
A synthetic L&D team must decide how to improve verification practice for 120 analysts who completed introductory AI training. Completion is 91%, but the organisation has no consistent evidence that people check sources or calculations in realistic work.
The BUILD case records:
The recommendation is a bounded combined pilot, not a full rollout. The business case includes the cost of review and evaluation, preserves the do-nothing comparator, and requires repair and retest if acceptance criteria fail.
- Baseline: completion and confidence surveys show reach and perception, not applied verification. Manager examples are useful but inconsistent across teams.
- Users and constraints: finance owns spend approval; operations defines task standards; L&D owns learning design; accessibility and privacy leads review the pilot. Quarter-end limits participant time.
- Alternatives: keep current reporting; add a knowledge check; add a two-hour simulation; or combine a shorter simulation with later manager observation.
- Ledger: the two-hour simulation offers more process evidence but creates substantial participant and reviewer cost. The combined option uses a 45-minute scenario, structured evidence review, and a later observation prompt. Benefits are expressed as decision-useful coverage, not promised productivity.
- Decision rules: pilot with 18 participants. Stop for a material accessibility defect, more than two incomplete evidence records, or unresolved scoring ambiguity. Proceed only if the evidence distinguishes verification behaviours that current methods miss and reviewers can explain the basis consistently.
Reporting boundaries for Two Prune
Two Prune can support a realistic simulation, capture process and deliverable evidence, and report observed assessment performance with confidence and calibration limits. It does not prove learning transfer, readiness across every role, productivity or financial return, or later workplace performance. It does not make employment decisions. Small or early comparison groups should be described as directional or unavailable rather than presented as stable benchmarks.
Primary sources
UK Government, Employer guide: what works for AI upskilling in the UK
- UK Government, Supporting case studies: what works for AI upskilling in the UK
- UK Government, Research evidence, analysis and methodology: what works for AI upskilling in the UK
- HM Treasury, The Green Book 2026
Sources: [1] [2] [3] [4]
Sources
- 1.UK Government, Employer guide: what works for AI upskilling in the UK · UK Government
- 2.UK Government, Supporting case studies: what works for AI upskilling in the UK · UK Government
- 3.UK Government, Research evidence, analysis and methodology: what works for AI upskilling in the UK · UK Government
- 4.HM Treasury, The Green Book 2026 · UK Government