Write an Attrition Hypothesis Memo from HR Data

Workflow area: HR & Recruiting · Tier: Advanced · Hands-on time: ~45 minutes

Attrition analysis is where correlation-as-causation does the most organizational damage. "Everyone who left had manager X" might mean manager X is the problem — or that manager X inherited the team with the below-market pay band. The memo format here forces every claim to carry its evidence and its confidence level, so leadership debates causes instead of anecdotes.

When to use — and when not to

• Use it for aggregate pattern analysis30+ exits over a period, sliced by department, tenure, level, manager span, compensation band, exit reason category.
• Use it before the retention budget conversationa ranked-hypothesis memo is exactly the pre-read that makes that meeting productive.
• Skip it for individual caseswhy one specific person left is a conversation, not an analysis; small-n "patterns" about named individuals are gossip with charts.
• Skip AI entirely for decisions about individualsno output of this workflow may feed a decision about any specific employee's performance, pay or role. This is aggregate organizational diagnosis only.

The recipe

1. 1.

   Build a truly anonymized extract. From your HRIS, export exits for the period plus a comparison snapshot of stayers: department, level band, tenure band, compensation quartile (not salary), manager span band, location, exit reason category, and exit month. No names, no employee IDs that HR could reverse, no free-text exit comments with identifying detail. Small-cell rule: if any combination describes fewer than five people, coarsen the bands until it doesn't — "the only senior designer in Pune" is identifiable no matter how many columns you deleted.

2. 2.

   Compute the base rates first. Upload to ChatGPT (data analysis) or Claude:

   Copy

   Two CSVs: employees who exited in the period, and the workforce
   snapshot. Using real computation:
   
   1. Overall attrition rate for the period.
   2. Attrition rate by: department, tenure band, level band,
      compensation quartile, manager span band. Show n for each
      group and flag any group with n < 5 as "too small to report".
   3. Which two or three slices deviate most from the overall rate?
   4. Exit-reason category distribution overall and for the
      highest-attrition slices.
   
   Tables only. No interpretation yet.
   

   Reconcile the overall rate and one slice against your own pivot table before proceeding. Rates, not raw counts — twelve exits from a 200-person department and twelve from a 30-person team are different emergencies.

3. 3.

   Generate hypotheses with evidence grades.

   Copy

   Based only on the verified tables above, list candidate hypotheses
   for the elevated attrition. For each:
   
   - Hypothesis (one sentence, causal claim made explicit).
   - Supporting evidence from the tables (with numbers).
   - Evidence grade: STRONG (large rate gap, adequate n, consistent
     across slices) / MODERATE / WEAK (small n or single slice).
   - Confounds: what else could produce this same pattern?
   - What data would confirm or kill it (exit interviews, engagement
     survey cut, market pay benchmark)?
   
   Rank by evidence grade. Include at least one hypothesis the data
   does NOT support, that leadership probably believes, and say why.
   

   The confound line is the one to read hardest. A good entry looks like:

   Copy

   H2: 1-2 year tenure cohort attrition (31% vs 14% overall, n=22)
   reflects below-market pay progression after year one. Grade:
   MODERATE. Confound: this cohort was largely hired remote in the
   2024 surge — could be hiring-quality or remote-isolation effect.
   Test: compare exits against market pay benchmark for the band.
   

4. 4.

   Draft the memo.

   Copy

   Write a 1-page attrition hypothesis memo for the leadership team.
   
   Structure:
   - Headline: overall rate, change vs prior period, cost framing.
   - What the data shows: the 3 strongest patterns, with rates and n.
   - Hypotheses: ranked, each labelled STRONG/MODERATE/WEAK, one
     line of evidence, one line of confound.
   - What this memo does NOT claim: explicitly note correlation vs
     causation and the small-n slices we declined to interpret.
   - Next steps: the 3 cheapest tests to confirm or kill the top
     hypotheses, with owners and dates.
   
   Tone: neutral, precise, no blame language toward any team or
   manager. Never name or identify an individual.
   

5. 5.

   Run the identification and blame sweep. Read the draft asking two questions: could any sentence identify a person (a slice small enough to point at someone)? Does any sentence assign blame rather than describe a pattern? Fix both. Then verify every number against the step-2 tables — a memo this sensitive cannot afford a transposed digit.

Make it yours

• Add the time dimension. Include exit month and ask for the trend — a rate that spiked after a specific reorg or policy change is a far stronger evidence line than a static snapshot.
• Triangulate with engagement data. If you run engagement surveys, compute (separately, same anonymization rules) whether the high-attrition slices also scored low — converging evidence upgrades a MODERATE to STRONG.
• Regrettable-only cut. Rerun the analysis on regretted exits only; the hypotheses that survive both cuts are the ones worth budget.

Quality check

• Every rate has its n, and small cells are suppressednothing reportable describes fewer than five people.
• Every hypothesis carries a grade and a named confoundno naked causal claims.
• The memo contains a "does not claim" sectioncorrelation/causation stated in writing.
• No individual is identifiable and no manager is blamedpatterns, not persons.
• Numbers verified against your own pivot tablethe model computed; you reconciled.