Join Qual and Quant into Churn-Driver Hypotheses

Workflow area: Analytics & Experimentation · Tier: Core · Hands-on time: ~40 minutes

Churn analysis done from quant alone produces correlations wearing trench coats: "users who never used integrations churn 3× more" — is that a driver or a symptom of never having had a real use case? Qual alone produces loud-minority bias: the angriest ticket writers are rarely representative of the silent leavers. The join is where the insight lives, and it used to take a week of manual coding. AI collapses the qualitative mining to an hour and drafts the join — while you guard the step everyone fumbles: keeping "hypothesis" and "validated driver" rigorously separate.

When to use — and when not to

• Use it for quarterly churn or win-loss reviewswhen you have both a body of qual (tickets, NPS verbatims, exit-survey answers, sales call notes) and cohort data for the same period and segments.
• Use it when retention moved and pb-28's tree points at "real behavior change"this is the deep-dive that branch deserves.
• Use it for win-lossthe identical mechanics apply to lost deals: call notes plus opportunity data, joined into ranked loss-driver hypotheses.
• Skip it if one evidence stream is missingquant-only or qual-only analysis is fine, but call it what it is; the ranked-drivers framing of this playbook requires both legs.
• Skip consumer AI tools for the qual miningchurned-customer communications are full of names, account details and grievances; this work happens in your approved tenant only.

The recipe

1. 1.

   Assemble the two evidence packs for the same window. Qual: exit-survey verbatims, churn-adjacent support tickets, cancellation-flow comments, sales/CS call summaries from Gong for accounts that left. Quant: churned vs. retained cohort comparison — tenure, plan, feature adoption, seat utilization, support-contact frequency, price events. Same period, same segmentation, or the join will manufacture artifacts.

2. 2.

   Mine the qual into counted, quoted themes. Use pb-17's structured-output mechanics with a churn-specific schema:

   Copy

   Classify each churned-customer verbatim below into stated
   departure reasons. Use ONLY these fields per item:
   { "id": "...", "reason_theme": "...", "verbatim_quote": "...",
     "confidence": "high|med|low", "segment": "..." }
   
   Rules: one primary theme per item; coin a new theme only if
   nothing fits; never infer a reason the text doesn't state —
   use "unclear" freely; quotes must be verbatim.
   After classifying, output a count table: theme, count, % of
   items, 2 representative quotes with IDs.
   
   Verbatims: [PASTE BATCH]
   

   Then sample-audit: pick 15–20 classified items and check them yourself. If more than a couple are miscoded, tighten the theme definitions and rerun — counts you haven't audited are decoration.

3. 3.

   Pull the quant cuts that could corroborate or contradict each theme. For each qual theme, ask: what would the cohort data show if this were a real driver? "Too expensive" (31% of verbatims) predicts churn concentrating after price events or at renewal, in low-utilization accounts. Run those cuts in your analytics tool — inspecting queries per pb-27.

4. 4.

   Have AI draft the join — as hypotheses with both evidence types linked.

   Copy

   Join these two evidence sets into ranked churn-driver hypotheses.
   
   For each hypothesis output:
   - DRIVER (one line, causal phrasing flagged as hypothesis)
   - QUAL EVIDENCE: theme, count, one verbatim quote with ID
   - QUANT EVIDENCE: the specific cohort cut and number
   - AGREEMENT: corroborate / contradict / qual-only / quant-only
   - CONFOUNDS: what else could produce this pattern
   - TEST: the cheapest way to validate (experiment, targeted
     interviews, save-offer pilot)
   
   Rank by (count × revenue at risk), but list contradicted and
   single-source items too — do not discard disagreements, they
   are findings. Do not present any hypothesis as a conclusion.
   
   Qual themes table: [PASTE]   Cohort cuts: [PASTE]
   

5. 5.

   Interrogate the disagreements yourself. The joins where qual and quant disagree are the most valuable output. "Missing integrations" dominates verbatims but churned accounts barely attempted integrations? Maybe it's a socially acceptable exit excuse — or an evaluation-stage loss your churn data can't see. Write your reading of each disagreement by hand; this is the judgment part.

6. 6.

   Ship it as a hypothesis memo, not a findings memo. One page: top 3–5 ranked driver hypotheses, each with both evidence links, its confound, and its proposed test. The deliverable's title says "hypotheses" — the validation (a save-flow experiment, ten targeted exit interviews) is the follow-up work this memo argues for.

Make it yours

• Gong/call-library variant: query the call library for churn-adjacent moments ("pricing pushback," "champion left," "evaluating competitor") and link clips as the qual evidence — clip-level links beat summaries because reviewers can check the source in one click.
• Win-loss edition: same recipe with lost-deal notes plus CRM opportunity data; loss-driver hypotheses ranked by pipeline value, with the same corroborate/contradict discipline.
• Standing pipeline: once the schema and cuts are stable, template both prompts in a Claude Project and run the join monthly — trend lines on driver themes are worth more than any single snapshot.

Quality check

• Both evidence types link from every ranked drivera driver with only one leg is labeled qual-only or quant-only, never silently promoted.
• Qual counts were sample-audited15–20 items checked by hand before any count was believed.
• Quotes are verbatim with IDstraceable to the source ticket, survey or call.
• Confounds are named per driverespecially symptom-vs-cause (low adoption may be a result of leaving-intent, not a reason).
• Disagreements survived into the memowith your hand-written interpretation, not smoothed over.
• The memo says "hypotheses" and proposes testsnothing in it reads as a validated conclusion.