Practical guide

How to report AI capability evidence to an HR leader

An HR leader needs a decision-ready view of workforce capability without turning incomplete assessment evidence into a claim about readiness, productivity, or individual employment suitability. A useful report separates what people demonstrated in a bounded task from participation, self-reported confidence, tool adoption, course completion, and business outcomes. It also makes missingness,.

By Two Prune

An HR leader needs a decision-ready view of workforce capability without turning incomplete assessment evidence into a claim about readiness, productivity, or individual employment suitability. A useful report separates what people demonstrated in a bounded task from participation, self-reported confidence, tool adoption, course completion, and business outcomes. It also makes missingness, access conditions, and permitted uses visible.

Start with the decision, not the dashboard

The UK government's 2026 employer guide for AI upskilling defines AI skills across technical, non-technical, and responsible or ethical categories and recommends practical, role-linked learning, accessibility, follow-up, and outcome monitoring. Its companion methodology says employers often struggle to measure whether training changes performance or outcomes. The AI foundation skills benchmark describes foundation knowledge and responsible-use skills for work. These sources support a disciplined evidence report; they do not validate a particular assessment or prove transfer to work. Employer guide: What works for AI upskilling in the UK Research evidence, analysis and methodology AI foundation skills for work benchmark

Classify the HR leader as the owner of learning strategy, access, communication, development routes, and workforce reporting only where that matches the organisation. HR may be a collaborator or reviewer for technology governance, business performance, and operational deployment. Do not imply universal budget or decision authority.

Sources: [1] [2] [3]

Use the SIGNAL report

Use SIGNAL: scope the decision, index the cohort, grade the evidence, narrate bounded findings, assign actions, and log the next evidence gate.

### S: Scope the decision and permitted use

Open with the decision the report supports, its audience, and its deadline. Examples include choosing the next practice module, deciding whether to repeat a pilot, or identifying which work scenarios need clearer guidance. State prohibited uses beside the purpose.

Name the assessment version, scenario family, date window, comparison basis, and responsible owners. If the report combines assessment, learning, HR, or tool data, explain why each source is present and who may access it.

### I: Index the cohort and denominator

Show invited, started, completed, assessable, and included counts separately. Record withdrawals, technical failures, accommodations, language variants, duplicate attempts, and incomplete submissions without converting them into poor performance.

Describe coverage by relevant work context and access condition, not by unnecessary personal categories. Small groups may need suppression or aggregation. A completion rate is a participation measure, not a capability score.

### G: Grade evidence strength

For every finding, grade evidence as strong, moderate, limited, or unavailable using declared rules. Consider task realism, directness of observation, scoring consistency, missing evidence, assessment conditions, and whether multiple signals agree.

Separate final-deliverable evidence from observed work-process evidence. Distinguish assessment scores from course completions, confidence surveys, manager impressions, and tool telemetry. Report reviewer disagreement and calibration notes rather than hiding them in an average.

Where percentiles or benchmarks are immature, label them directional or unavailable. Do not create precision from a small or unrepresentative reference group.

### N: Narrate findings with boundaries

Organise findings around the decision. For each one, state the observed pattern, evidence supporting it, uncertainty, plausible alternative explanations, and what the result does not establish.

Use the six Two Prune assessment pillars only when the underlying assessment actually measured them: Context Framing, Evidence Navigation, AI Orchestration, Verification & Risk Control, Judgment & Synthesis, and Communication & Delivery. Do not translate a pillar into a personality trait or permanent worker label.

Show distributions and case examples rather than only averages. Use de-identified examples that explain the scoring rule without exposing participant evidence beyond the authorised audience.

### A: Assign actions to owners

Tie each proposed action to an observed gap and an accountable owner. HR might own protected practice time, accessible learning routes, communications, and development resources. Assessment owners might revise confusing instructions or recalibrate a rubric. Business leaders might choose priority workflows. Security, privacy, or legal owners might clarify permitted tool and data use.

Prioritise repairs that affect the evidence before scaling. If access failures or scoring ambiguity are material, repeat the bounded exercise after repair rather than declaring a workforce conclusion.

### L: Log the next evidence gate

Record what will be checked next, when, by whom, and what result would change the decision. Separate near-term assessment evidence from later transfer or operational evidence. A stronger assessment result after practice does not by itself prove productivity or sustained workplace behaviour.

Version the report and retain the query, cohort rules, scoring configuration, limitations, approvals, and changes. If the assessment or comparison group changes, mark the comparability break.

HR evidence-report artifact

| Finding | Evidence type | Cohort and coverage | Strength | Limitation | Decision implication | Owner | Next gate | | --- | --- | --- | --- | --- | --- | --- | --- | | | assessment process and output | | | | | | | | | participation or access | | | | | | | | | learning or confidence | | | | | | |

Include a one-page decision summary, denominator flow, pillar distribution, scoring-confidence note, access and accommodation note, observed examples, action register, and next-evidence schedule.

Worked example: synthetic operations cohort

A fictional organisation invites 24 operations staff to a 55-minute AI-enabled incident-brief assessment. Twenty-two start, 20 submit, and 18 attempts are assessable. One submission is excluded after a platform outage removed source access; another used the wrong scenario version after an invitation error. Those exclusions are operational failures, not low scores.

The report's purpose is to choose the next practice module and decide whether the pilot is ready for a larger repeat. It explicitly prohibits use for promotion, disciplinary, selection, or compensation decisions.

Across the 18 assessable attempts, Context Framing and Communication & Delivery show consistent evidence. Verification & Risk Control is weaker: 11 participants repeat a plausible but unsupported incident-cost estimate, while only five retain a source or qualify the uncertainty. AI Orchestration varies widely, but the evidence is moderate because two reviewer pairs disagree on whether a prompt revision demonstrates verification or only stylistic iteration.

The denominator flow shows that completion is 20 of 24 invited, while assessed evidence covers 18 people. The report does not combine those numbers. It also shows that six participants used an approved accessibility adjustment and that their results are not interpreted as a separate capability group because the small count and mixed adjustments would not support that conclusion.

The action register assigns the learning team a source-checking practice module, the assessment owner a clearer verification anchor, and the reviewer lead a calibration exercise using four disputed submissions. HR owns protected practice time and the accessible delivery route. The operations leader identifies two real workflows for a later transfer check.

The disposition is repeat after evidence repair. The next gate is a 20-person bounded repeat with the corrected invitation, verified source access, and reviewer agreement checked before scoring the full cohort. A later workplace review may examine whether source qualification appears in real incident briefs, but the assessment report makes no productivity or transfer claim.

Two Prune boundary

Two Prune provides assessment evidence about observed performance in realistic work simulations. It can help an authorised HR leader understand cohort patterns, evidence strength, scoring confidence, and development areas. It does not certify workforce readiness, measure course impact by itself, infer protected traits, forecast later workplace performance, recommend employment actions, or make automated decisions.

Sources

UK government: Employer guide—What works for AI upskilling

  • UK government: Research evidence, analysis and methodology
  • UK government: AI foundation skills for work benchmark

Sources: [1] [2] [3]