Practical guide
How an HR leader can build the case for applied AI capability evidence
An HR leader's business case should not begin with a promise that assessment technology will raise productivity or identify “AI-ready” people. It should compare realistic ways to answer a bounded workforce decision, make resource and consequence assumptions visible, and purchase the next useful evidence at proportionate cost. The artifact is an option-and-evidence case: a decision record.
An HR leader's business case should not begin with a promise that assessment technology will raise productivity or identify “AI-ready” people. It should compare realistic ways to answer a bounded workforce decision, make resource and consequence assumptions visible, and purchase the next useful evidence at proportionate cost. The artifact is an option-and-evidence case: a decision record covering mandate, alternatives, costs, benefits, risks, pilot design, reporting boundaries, and the next funding gate.
Build an option-and-evidence case with PROVE
Use Purpose, Roles, Options, Value evidence, Exit gate. This is different from an HR decision agenda, which frames questions and stakeholders. A business case appraises alternative investments and states what evidence would justify the next commitment.
The UK government's 2026 employer guide to AI upskilling is based on workshops, case studies, and a 536-response employer survey. It emphasizes practical, contextualised, accessible, integrated, and sustainable learning while reporting evidence limitations and varied organisational contexts. The 2026 Green Book requires proportionate appraisal of options, costs, benefits, risks, uncertainty, and monitoring rather than selecting a preferred option first. NIST's AI Risk Management Framework Core calls for documented roles, training, monitoring, and risk responses. These sources inform the case structure; they do not prove that a particular assessment or learning investment will produce a return. UK employer guide to AI upskilling HM Treasury: The Green Book 2026 NIST AI RMF Core
### P: Purpose and decision boundary
State the decision, sponsor, affected population, horizon, and consequence. Examples include whether to fund a small applied-learning pilot, replace a weak proxy, or investigate a repeated verification failure. Avoid a generic objective such as “measure AI skills.”
Define HR's role for this decision: owner, influencer, reviewer, or collaborator. HR may own the learning portfolio, access and accommodations, participant communication, or use boundary. A functional leader should normally own task relevance and work quality. Security, privacy, technology, finance, and assessment owners retain their mandates. The sponsor owns the investment decision unless local governance says otherwise.
List non-negotiable boundaries: no automated employment decision, no inference from tool use alone, no reuse of development evidence for a new high-consequence purpose without review, and no claim beyond the task and evidence observed.
### R: Roles, current state, and evidence gap
Map the current process before proposing a purchase. Record how the organisation now learns about capability: course completion, self-report, manager observation, tool analytics, work quality review, simulations, or no consistent method. State the cost, delay, coverage, and blind spots of each.
Translate the concern into an evidence gap. “Managers lack confidence” is not yet a purchasing requirement. “The learning team cannot tell whether participants apply source verification to a defined policy-briefing task, so it cannot decide whether to repeat or repair the programme” is decision-ready.
Document baseline evidence and uncertainty. Do not invent productivity loss, error frequency, or programme impact. Where current data are missing, price the discovery work or carry a range. Missing evidence weakens confidence; it does not prove poor capability.
### O: Options and whole-resource cost
Compare at least three credible options, including the status quo. A useful set might be:
For each option, estimate setup time, participant time, manager and reviewer time, accessibility work, content design, security review, system costs, evidence operations, support, change management, and exit costs. Separate committed cash, internal capacity, and assumptions. A low software price can conceal substantial reviewer or task-design work.
Assess reversibility and dependencies. Prefer a smaller step when the principal uncertainty can be resolved before a broad commitment. Identify what would make an option infeasible: inaccessible delivery, unclear purpose, insufficient reviewer capacity, unrepresentative tasks, weak source controls, or consequences that exceed the evidence.
### V: Value evidence and risk-adjusted comparison
Connect benefits to decisions rather than adjectives. Possible benefits include replacing a weak proxy, identifying a specific learning need, producing a traceable evidence packet, reducing ambiguity in a repeat decision, or revealing that a task or programme needs repair. Do not count the same benefit under learning, productivity, and risk.
Use a decision-evidence table:
| Claimed value | Evidence needed | Earliest credible test | What would refute it | | --- | --- | --- | --- | | identifies actionable learning needs | repeated task-specific evidence patterns and owner acceptance | bounded pilot | results too sparse or generic to change learning | | improves decision clarity | sponsor can choose repair, repeat, stop, or expand from the evidence packet | decision review | sponsor still relies on unsupported proxy | | is workable | completion, accessibility, reviewer effort, exceptions | operational pilot | unequal access or unsustainable review load |
Record risks by option: purpose drift, participant trust, accessibility, privacy, reviewer inconsistency, task validity, source quality, workflow burden, and overinterpretation. Assign an owner and mitigation. Do not convert all risks to money when the conversion would be speculative.
### E: Exit gate and reporting boundary
Approve a pilot only with preset success, repair, stop, and escalation conditions. Name the cohort rationale, task, materials, support, evidence captured, reviewer calibration, access checks, reporting audience, and decision date. Specify what the pilot cannot establish.
The exit gate should release the smallest justified next step. Options include stop; repair the task or access; repeat with targeted practice; test another work context; or expand cautiously. A successful completion rate alone is not an expansion gate.
Report participant-level evidence only to authorised people for the stated purpose. Group summaries should show denominator, evidence gaps, variation, and uncertainty. Avoid league tables, universal capability labels, or employment recommendations. Preserve a route for correcting material evidence errors.
- continue current manager observation and course-completion reporting;
- improve existing work-sample review with a common evidence form;
- run a bounded applied assessment pilot with calibrated human review;
- combine a pilot with targeted practice and a repeat task.
Sources: [1] [2] [3]
Worked business case: synthetic service-operations cohort
A fictional organisation has 60 service-operations staff completing a general AI course. HR owns the learning portfolio and is an influencer on funding; the service director owns task relevance and the sponsor owns the investment decision. Course completion is 92%, but managers cannot tell whether participants can verify AI-drafted customer summaries against source records.
The case compares three options. The status quo costs little additional cash but leaves the decision unresolved. A common manager checklist requires about 24 manager-hours and improves consistency, but observation opportunities vary. A 12-person applied pilot requires a synthetic case, accessibility check, participant time, six calibrated reviewer-hours, and an evidence review. The team does not assign invented productivity savings to any option.
The sponsor chooses the pilot because it is reversible and directly tests the evidence gap. The task contains a customer history, conflicting notes, an unsupported generated statement, and a required escalation. The pilot's gate is not an average score. It asks whether participants have comparable evidence opportunities, whether reviewers can support dispositions, whether repeated failure modes lead to a concrete learning change, and whether review effort is sustainable.
Results show that eight participants catch the unsupported statement, five preserve the unresolved conflict, and four escalate correctly. Reviewers disagree on two cases because the escalation criterion is ambiguous. The disposition is repair and repeat: clarify the rubric, add contrastive practice, and run a second bounded case. The case does not claim improved productivity, return on investment, or broad workforce readiness. It has bought enough evidence for the next learning decision without funding an organisation-wide rollout.
Two Prune boundary
Two Prune can provide task-bounded assessment evidence and traceable scoring support for human interpretation in a defined pilot. It does not run HR processes, determine investment value, prove workplace transfer, predict future performance, or make employment decisions.
Sources
UK Government: Employer guide to AI upskilling
- HM Treasury: The Green Book 2026
- NIST AI Risk Management Framework Core
Sources: [1] [2] [3]
Sources
- 1.UK employer guide to AI upskilling · UK Government
- 2.HM Treasury: The Green Book 2026 · UK Government
- 3.NIST AI RMF Core · National Institute of Standards and Technology