Practical guide

AI customer health report verification checklist

Use this checklist after a customer health report has been drafted and before anyone acts on it. The VERIFY method checks Version and scope, Evidence traceability, Recalculation, Inconsistent signals, Follow-through to a bounded decision, and Yielded uncertainty. It tests whether a reviewer who did not prepare the report can reconstruct every material finding, calculation, interpretation, and.

By Two Prune

Use this checklist after a customer health report has been drafted and before anyone acts on it. The VERIFY method checks Version and scope, Evidence traceability, Recalculation, Inconsistent signals, Follow-through to a bounded decision, and Yielded uncertainty. It tests whether a reviewer who did not prepare the report can reconstruct every material finding, calculation, interpretation, and next action.

V — Verify the version, scope, and intended decision

Start with the report object, not its headline colour. Record the account or account group, products, lifecycle stage, reporting window, source cutoffs, report version, preparer, reviewer, and intended decision. Confirm that the report is the current approved version and that any prior status is available for comparison.

Check these items:

ChurnZero’s first-party handbook describes customer-health designs that vary by lifecycle and segment and that combine quantitative and qualitative evidence. That supports checking whether the report’s design matches its stated population. It does not establish that a particular factor predicts retention for another company. ChurnZero customer health score handbook

Fail the check when the reviewer cannot tell which population or decision the headline represents. Do not repair an undefined scope by adding a caveat after the recommendation; return the report to its owner.

  • The account hierarchy and included business units are explicit.
  • The relevant user, contract, product, and outcome populations are named.
  • Each source has a cutoff date; “current” is not a date.
  • The lifecycle stage and any segment-specific rule are recorded.
  • The intended decision is concrete, such as investigating a blocker or refreshing a success plan.
  • A method change, threshold change, or source outage is disclosed beside the result it affects.

Sources: [1]

E — Establish evidence traceability

Every material statement should point to a dated source record or a documented human judgment. Use an evidence register rather than accepting embedded screenshots without provenance.

| Finding ID | Claim in report | Source system and object | Population and cutoff | Extract or record ID | Evidence owner | Reproducible? | | --- | --- | --- | --- | --- | --- | --- | | F-01 | relevant workflow use fell | product telemetry event set | intended users through date | saved query/version | analytics owner | yes/no | | F-02 | implementation milestone complete | success plan | current approved plan | milestone and acceptance record | success owner | yes/no | | F-03 | relationship concern increased | dated interaction note | named stakeholders in period | note and reviewer annotation | account owner | yes/no |

Confirm that quotations preserve their context, AI-generated summaries link back to the reviewed material, and judgment labels name the accountable person. Check whether sensitive material entered an approved service. A source link that now resolves to a different version is not traceability; retain the dated object, export, or immutable record permitted by the organisation’s controls.

Gainsight’s official standard-object glossary distinguishes current and previous health-score fields and documents several customer-success objects. This supports preserving dated object identity and score history rather than retaining only the latest display. Exact fields still depend on the customer’s configured product and data model. Gainsight standard objects

Sources: [2]

R — Recalculate material signals

Reperform every calculation that could change the decision. Use source values and the recorded formula, not a copied total from the report.

For each signal, check:

Retain a calculation receipt containing the query or formula version, input snapshot identifier, result, reviewer, and review date. If the report cannot be recalculated, label the affected conclusion unverified and prevent it from driving an irreversible action.

  • numerator, denominator, exclusions, and null handling;
  • account and user joins, including merged and parent-child records;
  • time zone, reporting window, late-arriving events, and duplicate handling;
  • threshold, weight, normalization, and effective date;
  • comparison period and whether its population is genuinely comparable;
  • rounding and whether a boundary case changes the displayed category;
  • subjective inputs with and without the reviewer-supplied value.

I — Investigate contradictions and missingness

A health report is most useful where signals disagree. Create a contradiction log instead of averaging disagreement away.

| Contradiction | Plausible explanations | Evidence needed | Disposition | | --- | --- | --- | --- | | high login rate, low workflow completion | wrong event, access friction, shallow use | intended-user roster and completion funnel | investigate / resolved / open | | no open cases, unresolved customer concern | merged case, external channel, stale sync | problem record and dated customer confirmation | investigate / resolved / open | | milestone marked complete, no acceptance record | administrative closure or missing attachment | signed or recorded acceptance evidence | investigate / resolved / open |

Treat absent evidence as absent. Zero cases is not automatically zero problems, and silence is not automatically positive sentiment. Challenge causal language: a usage change observed after an intervention does not by itself show that the intervention caused it. Record alternative explanations and the evidence that would distinguish them.

F — Follow the finding to a bounded action

Match each actionable finding to a decision, owner, due date, closure artifact, and escalation condition. Generic advice such as “increase engagement” fails this check.

The action should be proportionate to the evidence. A low-confidence indicator may justify investigation, not an executive escalation. A completed meeting is activity, not closure; define what new evidence will close the finding. If the report proposes a status or threshold change, identify who has authority to approve it and which report version the approval covers.

Y — Yield residual uncertainty

End the review with what remains unknown. List source outages, excluded business units, stale success outcomes, subjective inputs, unresolved contradictions, and untested predictive claims. Assign an uncertainty owner and a next verification date where further evidence is feasible.

Use one of four dispositions:

  • Pass: the decision can be reconstructed and limitations do not undermine it.
  • Pass with action: the decision is bounded, but named evidence must be collected.
  • Return for correction: a repairable source, calculation, or interpretation defect could change the decision.
  • Hold: the necessary population or evidence is unavailable, so no supported decision can be made.

Worked check: the green report that fails verification

Harbour North is a synthetic implementation account. Its report is green because 44 of 50 licensed users logged in, all scheduled meetings occurred, and the dashboard shows no open critical cases. The proposed action is to “maintain momentum.”

VERIFY finds that the report is version 7, but the recommendation was copied from version 6 before a product migration. The intended workflow has only 20 users; 17 logged in, while five completed the required approval step. The calculation counted activity from a separate analytics module. The success plan marks data validation complete, but its linked acceptance field is empty. Two medium-severity cases were merged into a problem record that is still open. A positive meeting summary omits a dated objection about incomplete mappings.

The reviewer recalculates the relevant completion rate as 5 of 20, records the unresolved problem, and changes the disposition to return for correction. The bounded actions are to confirm the intended-user roster, complete data acceptance, resolve the mapping defect, and repeat the workflow-completion measure in fourteen days. No churn probability or causal claim is added.

Copyable VERIFY review record

Two Prune can assess how a participant verifies a customer-health report in a realistic work simulation: tracing evidence, recalculating a material signal, challenging an interpretation, and communicating a bounded disposition. It does not operate a customer-success programme, predict churn, or make the customer’s commercial decision.

  • [ ] Version and scope: account hierarchy, products, lifecycle, window, cutoffs, version, decision.
  • [ ] Evidence: every material finding has a dated source anchor and accountable owner.
  • [ ] Recalculation: joins, formulas, denominators, exclusions, thresholds, and comparison periods reproduce.
  • [ ] Inconsistency: contradictions, alternative explanations, and missing evidence are logged.
  • [ ] Follow-through: each action has an owner, due date, closure artifact, and escalation rule.
  • [ ] Yielded uncertainty: residual unknowns and untested claims are visible beside the decision.
  • [ ] Disposition: pass, pass with action, return for correction, or hold, with reviewer and date.

Sources

  1. 1.ChurnZero customer health score handbook · churnzero.com
  2. 2.Gainsight standard objects · support.gainsight.com