The core idea
Our practical workflow is business question → relevant data → analysis → insight → intervention → measured outcome. It is an editorial synthesis, not a proprietary assessment. Begin with the decision that needs support and end by checking what the action changed. An interesting chart is an intermediate result. The workflow should be proportionate: a small, well-defined comparison may answer the question better than a complex predictive system.
Source and attribution [1]Using it in practice
Name the decision owner and agree a population, outcome and time window. Collect only information relevant to the purpose, with appropriate access and handling. Inspect quality before analysing patterns. Separate what the data shows from your explanation of why. Discuss the finding with people who understand the work, then choose a feasible action and an evaluation design. Keep a record of definitions, assumptions and changes so the work can be updated.
An example, not a reported case
Worked example · illustrative
A fictional employer asks why casual workers leave early. HR first defines active work and departure, then compares fully observed 90-day joining cohorts by site, tenure and shift availability. A pattern of low suitable-shift availability prompts interviews and process checks. The team pilots clearer onboarding and allocation communication. Subsequent retention is compared over equivalent windows, with contract mix and seasonal demand considered. The conclusion reports uncertainty and does not claim that the observed pattern proves a manager caused departures.
What to watch for
Do not turn the question into a search for which employees are likely to be disloyal. Avoid collecting sensitive or intrusive data without a justified purpose and appropriate legal assessment. A plausible insight is not yet evidence that an intervention works. Keep the distinction between prediction, explanation and causal evaluation visible throughout.
Define the decision before the analysis
Write the choice the organisation faces, its owner, deadline and affected population. “Produce a dashboard” describes an output; “decide whether to trial a revised induction process” describes a decision. Specify what could change as a result and what is outside scope. If no plausible decision follows, reconsider the request before collecting more data. An analysis brief should remain short enough for the operational owner to challenge. Record disagreements about definitions early, when they can still be resolved without rebuilding the whole analysis.
Use more than one kind of evidence
CIPD’s evidence-based-practice factsheet distinguishes research, organisational information, practitioner expertise and stakeholder perspectives. It also stresses critical appraisal rather than accepting a familiar practice because it is popular. In a local analysis, this means asking what each source can and cannot establish. A pattern in records may reveal where to investigate; employee accounts may explain practical obstacles; research may suggest a mechanism worth testing. None automatically proves that a proposed intervention will work in this organisation. Keep contradictory evidence visible alongside the evidence that supports the preferred explanation.
Create a minimum data specification
For each field, record its purpose, definition, source, coverage period and quality concern. Define the population and denominator explicitly. For an early-retention question, specify whether casual workers with no recent shifts count as active, whether rehires start a new period and when the 90-day observation window ends. These choices can change the answer more than the chart type. Exclude fields that do not support the decision. Agree access, retention and privacy requirements with the appropriate people before extracting personal information; an analytical interest alone is not a justification for collecting it.
Check comparability before ranking teams
A site with a higher exit percentage may have a different role mix, newer starters or a seasonal contract. Start with counts and denominators, then compare like groups over equivalent windows. Flag small samples rather than presenting unstable rankings as precise performance measures. Check missing data and whether reporting practices differ between locations. Describe an association as an association. A manager’s name appearing beside an outcome does not demonstrate that the manager caused it, and an aggregate pattern is not a sound basis for judging a particular employee.
An invented retention example
Suppose 12 of 40 starters in one fully observed cohort leave within 90 days, compared with 6 of 40 in the next cohort. The observed rates are 30% and 15%, a 15-percentage-point difference. Calling this a 50% relative reduction is arithmetically possible but can sound more conclusive than the evidence warrants. Before attributing the change to onboarding, examine season, roles, shift availability and other concurrent changes. Speak with relevant employees and managers. These invented counts illustrate interpretation; they are not evidence for a particular retention intervention.
Specify the intervention and comparison in advance
Describe exactly what will change, who will receive it, who will deliver it and when. Choose an outcome and observation window before rollout, and record a credible comparison where feasible. Also measure whether the intervention was delivered: an ineffective design and an undelivered design are different findings. Include potential harms or burdens, such as extra administrative work or unequal access. Where the design cannot support a causal conclusion, state that limitation. Useful learning can still come from a carefully described pilot without claiming it proved causation.
Present a decision and an honest uncertainty statement
Separate the observed finding, its possible explanations and the recommended next action. Show the population and dates beside the headline number. Explain what would change the recommendation and what remains unknown. Offer a proportionate next step with an owner, resources and review point. Avoid hiding inconclusive results: a decision to stop a weakly supported initiative can be valuable. Use the downloadable evidence-to-action worksheet to record the reasoning in ordinary language, while keeping identifiable employee records in approved systems rather than in a widely shared briefing.
Make the next review reproducible
Keep definitions, calculation steps, source dates and an appropriate record of changes. Record the actual decision and its rationale, including when leaders choose a different option. At the review, check delivery, outcomes and unintended effects, then decide whether to continue, modify or stop. Share a suitable account of what changed with the people who contributed evidence. This workflow is an ADM HR teaching aid; the references provide further reading, not certification that the local analysis or decision is valid.
Take it into your next conversation
Three useful questions.
- Which decision will be different because of this work?
- What evidence would contradict our preferred explanation?
- How will employees know what changed as a result?
Related terms
Go to the evidence
Sources & attribution
Further reading: Nadeem Khan and Dave Millner (2023), Introduction to People Analytics: A practical guide to data-driven HR, 2nd edition, PDF pp. 227, 244 and 294–295. Edition & reading notes ↗
[1] CIPD: People analytics strategy and process ↗
[2] HM Treasury: evaluation principles ↗
[3] CIPD: Evidence-based practice for effective decision-making ↗
[4] ADM HR: editable Word worksheet ↗
The core idea is an original summary of the cited work. Application notes, examples and sketches are our interpretations, not quotations or reproductions of the authors’ figures. Publisher records may require access to read the full original work.
Application guidance expanded on 26 September 2026. These practical suggestions and invented examples are editorial explanations; further-reading sources are not the original authors of the historical models unless expressly identified. Original overview review dates are retained. Reworked on 28 September 2026 to consolidate repeated application notes and add original decision exercises, worked examples and practical guidance. Newly linked CIPD and AIHR resources were checked; existing legal review dates are unchanged.
Published 2026-09-20 · Reviewed 2026-09-22. Editorial approach