Practical guide
AI capability measurement guide for a people analytics leader
A people analytics leader should begin an AI capability programme with a decision agenda, not a dashboard request. The role can help frame questions, design evidence, test alternative explanations, and report uncertainty. It should not be assumed to own learning delivery, technology controls, employment decisions, or every investment decision. This guide uses the QUESTION agenda: qualify the.
A people analytics leader should begin an AI capability programme with a decision agenda, not a dashboard request. The role can help frame questions, design evidence, test alternative explanations, and report uncertainty. It should not be assumed to own learning delivery, technology controls, employment decisions, or every investment decision. This guide uses the QUESTION agenda: qualify the decision, understand the work, separate constructs, test evidence, include safeguards, organize responsibilities, and name the next decision.
Role basis and limits
CIPD defines people analytics as using data about people and the business to inform organizational decision-making. Its profession map includes analytical consulting, research design, data analysis, data science, and data governance, while its practitioner guide treats organizational data as one of four evidence types alongside professional expertise, scientific literature, and stakeholder views. The current UK employer guide to AI upskilling recommends training connected to real tasks, roles, systems, governance, and responsible use. These sources support the agenda below. They do not establish that a people analytics leader owns a universal mandate or that a particular programme produces workforce outcomes. CIPD Profession Map: People analytics CIPD: People analytics guide UK employer guide to AI upskilling
Sources: [1] [2] [3]
Q — Qualify the decision
Write the decision before selecting metrics. Examples include whether to investigate a workflow, run a bounded capability pilot, repair a learning pathway, or collect more evidence. State who owns that decision and what the people analytics leader contributes.
Use a decision card:
| Field | Question | | --- | --- | | Decision | What choice must be made, by whom, and by when? | | Population | Which roles, locations, access conditions, and workflows are in scope? | | Evidence use | Planning, learning, programme improvement, assessment reporting, or another stated purpose? | | Alternatives | Continue current practice, change process, provide support, run a pilot, or stop? | | Consequence | Who benefits, who carries risk, and what is reversible? | | Minimum evidence | What would be sufficient for this bounded decision? | | Stop condition | Which finding would prevent or narrow the action? |
Do not begin with “prove the workforce is AI-ready.” That combines an undefined construct, an unbounded population, and a predetermined conclusion.
U — Understand representative work
Map the actual work before measuring capability. Select a small set of consequential tasks across roles rather than a list of tool features. For each task, record the audience, inputs, constraints, expected artifact, quality criteria, verification route, and failure consequence.
Interview workflow owners and practitioners. Review existing artifacts, error patterns, policies, support requests, and access conditions. Include people who use different tools, require accessibility support, or work under different information restrictions. A high-frequency task may be a poor pilot if its outcome is difficult to observe or if safe practice cannot be reproduced in a simulation.
The output is a task-to-evidence map, not a catalog of occupations.
E — Establish separate constructs
Do not collapse exposure, confidence, use, skill, judgment, artifact quality, and business outcome into one readiness score.
Define each construct in observable terms:
For every construct, specify what it excludes. Self-reported confidence may help identify support needs, but it is not direct evidence of applied capability or safe judgment.
- access: whether approved tools and required accommodations are available;
- exposure: whether someone has encountered the relevant workflow and guidance;
- use: what actions occurred in a stated period and context;
- applied capability: how a person frames, executes, verifies, and communicates in a representative task;
- artifact quality: whether an output meets task-specific acceptance criteria;
- control operation: whether required review, escalation, privacy, or security steps occurred; and
- outcome: a downstream change that may have multiple causes.
S — Specify the evidence design
Use a mixed evidence plan matched to the decision. A plan might combine a representative work simulation, artifact review, process evidence, a short survey, access records, workflow interviews, and operational quality data. Record why each source is included and what it cannot show.
Create a metric contract for every reported measure:
| Field | Required definition | | --- | --- | | Construct and decision | What is being represented and why? | | Numerator and denominator | Which events or people count, and which do not? | | Population and period | Who and when? | | Source and owner | Where does the evidence originate? | | Missingness | What is unavailable, and how is it reported? | | Confounds | Which plausible alternatives could explain the result? | | Review | What independent or specialist check is required? | | Interpretation limit | What conclusion is out of scope? |
Predefine the analysis and critical exceptions before looking at favorable results. Do not fill missing denominators with the number of dashboard users or infer non-use from absent telemetry without checking access and configuration.
T — Test alternative explanations
For every interpretation, ask what else could produce the pattern. Differences between groups may reflect task mix, tool access, prior training, manager support, language, accessibility, time available, or the reliability of the measure. Improvement after a programme may reflect selection, practice effects, concurrent process changes, or a changed scoring standard.
Use comparison only when the groups, tasks, conditions, and measurement are sufficiently comparable. Otherwise describe the evidence without causal language. Preserve unusual critical failures even when an average looks strong.
I — Include rights, safeguards, and purpose controls
People data can be sensitive personal information. Confirm the applicable privacy, legal, security, employee-relations, accessibility, retention, and information-governance review before collection. State the purpose at collection and prevent silent reuse for an incompatible decision.
Collect the minimum evidence needed for the stated programme decision. Apply access controls, retention dates, and aggregation rules. Review whether small groups could be reidentified. Give participants an accurate explanation of what is observed and how it is used. Provide an accessible route to question or correct relevant records where the governing process requires it.
Two Prune is an assessment product. Its evidence should not be repurposed as an automated employment recommendation.
O — Organize owners and challenges
Use a responsibility map instead of assigning every action to people analytics.
| Work | Typical accountable party | People analytics contribution | | --- | --- | --- | | Workflow quality | Workflow or service owner | Define evidence and examine patterns | | Learning pathway | Learning owner | Diagnose needs and evaluate evidence | | Tool and data controls | Technology, security, privacy, or legal owner | Specify measurable control evidence | | Assessment design | Assessment owner with subject specialists | Advise on construct, sample, and interpretation | | Programme decision | Named executive or programme owner | Present evidence, alternatives, and uncertainty |
Name an independent challenger for consequential analysis. Record disagreements and unresolved evidence rather than converting them into consensus wording.
N — Name the next decision
Report a bounded disposition: proceed with a pilot, repair the evidence design, narrow the population, address access first, gather more evidence, or stop. Pair it with an owner, date, evidence request, and review condition.
The report should answer:
- What decision was examined?
- What evidence was available and missing?
- What patterns and critical exceptions were observed?
- Which alternative explanations remain plausible?
- What can and cannot be inferred?
- What is the smallest defensible next action?
Worked agenda: a fictional finance-operations pilot
A people analytics team is asked to show whether 140 finance-operations staff are “AI ready.” The leader rejects that formulation and reframes the decision: should the organization run a six-week pilot for preparing monthly variance-commentary drafts?
The work map shows that the task requires source selection, calculation checks, explanation of material variances, handling of confidential information, and escalation of unresolved anomalies. The evidence plan includes approved-tool access, a representative assessment simulation, blind review of the draft artifact, process evidence about verification, and interviews about workflow constraints. Confidence is collected only as a support signal.
Initial results show stronger average communication than calculation verification. Twelve participants lack access to the approved reference system, and three outputs contain unsupported explanations despite polished prose. The team does not publish one readiness percentage. It reports access separately, retains the critical unsupported-claim cases, and identifies tool access and verification practice as prerequisites.
The programme owner approves a 24-person pilot after access is repaired. The learning owner adds an exercise on source traceability; finance owns artifact acceptance; security owns data-handling controls; people analytics owns the measurement plan and uncertainty note. The next decision is whether evidence from the bounded pilot supports expanding to another team. No claim is made about productivity, workplace performance, or the whole workforce.
Agenda quality check
[ ] The decision, owner, population, time, and alternatives are explicit.
- [ ] Representative work and acceptance evidence are defined.
- [ ] Access, confidence, use, capability, quality, controls, and outcomes remain separate.
- [ ] Every measure has a reproducible contract and missingness rule.
- [ ] Alternative explanations and critical exceptions are visible.
- [ ] Privacy, security, accessibility, employee-relations, and legal owners reviewed within their mandates.
- [ ] Role responsibilities reflect the local organization rather than a generic title.
- [ ] The disposition is bounded and names the next evidence gate.
Two Prune boundary
Two Prune can provide a realistic assessment simulation and evidence about how a participant framed a task, navigated supplied information, used AI, verified work, synthesized a recommendation, and communicated a deliverable. It reports observed assessment performance and evidence strength. It does not prove workforce outcomes, make employment decisions, replace a people analytics strategy, or determine whether a programme should be funded.
Sources
CIPD Profession Map: People analytics
- CIPD: People analytics guide for people professionals
- GOV.UK: Employer guide to AI upskilling
Sources: [1] [2] [3]
Sources
- 1.CIPD Profession Map: People analytics · www.cipd.org
- 2.CIPD: People analytics guide · prod.cipd.org
- 3.UK employer guide to AI upskilling · UK Government