Practical guide

How to create a customer health report with AI

A customer health report should help an accountable customer team decide what to investigate or do next. It is not a decorative dashboard and it is not a churn prediction unless the underlying model has been tested for that purpose. The report needs to preserve the difference between observed facts, calculated indicators, human judgment, and AI-assisted interpretation. This guide uses a TRACE.

By Two Prune

A customer health report should help an accountable customer team decide what to investigate or do next. It is not a decorative dashboard and it is not a churn prediction unless the underlying model has been tested for that purpose. The report needs to preserve the difference between observed facts, calculated indicators, human judgment, and AI-assisted interpretation. This guide uses a TRACE method: Target the decision, Reconcile the population, Analyze signals in context, Capture evidence and challenge, and Escalate with a bounded action. The method produces a report that another reviewer can reconstruct rather than a colour that must be trusted on sight.

T — Target the decision and reporting window

Write the decision at the top of the report. Examples include whether to investigate a fall in adoption, resolve an implementation blocker, convene an executive review, or test whether a success plan is still current. Avoid the vague instruction to “improve health.”

Record:

Lifecycle stage matters. An onboarding account should not be judged against the same usage pattern as a mature account. ChurnZero’s first-party health-score handbook recommends tailoring factors by segment and lifecycle stage and combining quantitative and qualitative evidence rather than relying on a single universal score. That is useful design guidance, not proof that any particular factor predicts renewal in your customer population. ChurnZero customer health score handbook

  • account, product, segment, lifecycle stage, contract period, and report date;
  • reporting window and data cutoff for every source;
  • accountable owner and intended decision maker;
  • current success outcomes and their evidence owner;
  • material limitations, missing systems, and known delays;
  • the previous report version and the reason for any method change.

Sources: [1]

R — Reconcile the account and source population

Create a source register before calculating the headline status.

| Source | Population and cutoff | Stable account key | Required fields | Quality check | Owner | | --- | --- | --- | --- | --- | --- | | Product telemetry | named product and active users through date | account plus user | relevant events, timestamps, entitlements | reconcile licensed, invited, and active users | product analytics | | Support | cases opened or active in period | account | severity, age, state, SLA clock | inspect reopened and merged cases | support operations | | Success plan | current approved version | account | outcome, milestone, due date, owner | identify stale or ownerless milestones | customer success | | Relationship evidence | meetings and documented feedback | account plus contact | role, date, topic, disposition | distinguish silence from negative feedback | account owner | | Commercial record | current contract snapshot | account | term, renewal date, products | reconcile amendments and hierarchy | revenue operations |

Do not let an AI tool guess account joins from similar names. Preserve the join rule and exception list. If a parent account has several business units, show whether signals are aggregated, weighted, or kept separate. A report that silently drops a business unit can look healthier than the underlying population.

Gainsight’s published object glossary distinguishes current and previous health score fields and exposes multiple customer-success objects. It supports the practical point that a report should retain dated source objects and score history rather than only the latest displayed state. The exact objects and availability still depend on the customer’s configured product. Gainsight standard objects

Sources: [2]

A — Analyze signals without collapsing their meaning

Use a signal ledger before producing any overall status.

| Signal | Observation | Baseline or expectation | Direction | Confidence | Interpretation limit | | --- | --- | --- | --- | --- | --- | | adoption | exact relevant-feature users / entitled users | stage-specific target | rising / stable / falling | high / medium / low | activity is not outcome attainment | | implementation | milestones completed on time | approved plan | ahead / on track / delayed | high / medium / low | plan may be stale | | support | severity and age of unresolved cases | service expectation | improving / stable / worsening | high / medium / low | volume can rise with adoption | | relationship | documented stakeholder engagement | agreed contact map | broadening / stable / narrowing | high / medium / low | silence has several causes | | outcomes | evidence against named success outcomes | success plan | achieved / progressing / unverified | high / medium / low | customer attribution may be unavailable |

Keep the raw indicator, its denominator, and its interpretation separate. “Twenty active users” is not meaningful without the entitled or intended population. “No support cases” may mean a stable product, low use, or an unobserved channel. A sentiment label should link to the dated interaction and reviewer challenge; it should not replace the source conversation.

If you use weights, publish the formula, effective date, and owner. Show the result both with and without any subjective component. Do not label a score “predictive” unless it has been evaluated against an appropriate outcome with held-out data, error analysis, and ongoing monitoring.

C — Capture evidence, contradiction, and human challenge

For every material signal, preserve a source anchor and a challenge note. The challenge asks what alternative explanation could produce the same pattern.

Examples:

AI can help normalize event labels, compare periods, draft the first narrative, and identify contradictory signals. A human owner should verify the population, material calculations, source quotations, account context, and proposed action. Record which passages were AI-assisted and what the reviewer changed. Never send sensitive customer material to an unapproved service.

  • usage fell because a seasonal workflow ended, not because value declined;
  • meeting attendance fell after the customer changed its operating cadence;
  • support cases rose because a wider rollout began;
  • an executive sponsor left, but decision authority transferred cleanly;
  • a positive meeting summary omitted two unresolved objections.

E — Escalate with a bounded action and verification date

The report should finish with decisions, not generic recommendations.

| Finding | Proposed action | Owner | Due date | Evidence needed to close | Escalation condition | | --- | --- | --- | --- | --- | --- | | relevant-feature adoption fell after entitlement expansion | confirm intended users and inspect onboarding blockers | customer success lead | date | agreed user population and blocker log | executive review if intended users lack access | | critical case remains unresolved | convene support and product owner review | support lead | date | case disposition and customer confirmation | service-risk escalation at agreed age | | outcome evidence is stale | refresh the success plan with customer owner | account owner | date | dated outcome record and next measure | treat outcome as unverified, not achieved |

Choose a verification date appropriate to the action. A completed meeting is not closure; closure requires the specified evidence. Keep the previous report and explain any change in status, factor, weight, or threshold.

Worked report: a false green implementation account

Northstar is a synthetic account in month four of implementation. The draft report marks it green because 42 of 48 licensed users logged in during the month, the customer attended all three scheduled calls, and no critical support case is open.

TRACE changes the decision. Population reconciliation shows that only 18 of the 48 licences belong to the workflow being implemented; 15 of those users logged in, but only six completed the relevant approval action. The other logins came from a separate reporting module. The success plan lists a data migration milestone as complete, while the linked acceptance record is blank. Two medium-severity cases were merged into a problem record, so the open-case count of zero concealed an unresolved dependency. Meeting notes contain a positive general comment and a specific objection about data completeness.

The report therefore records strong general access, incomplete workflow adoption, an unverified migration milestone, and an unresolved implementation dependency. It does not assign a renewal probability. The bounded actions are to validate the intended-user roster, complete migration acceptance, resolve the problem record, and remeasure approval completion in fourteen days. The account owner also records the customer’s objection verbatim and asks for confirmation after remediation.

Copyable customer health report

Two Prune can assess how a participant frames a customer-health question, reconciles account evidence, uses AI, challenges a draft interpretation, and communicates a bounded action in a realistic work simulation. It does not operate the customer’s health programme, predict churn, or make commercial decisions for the customer.

  • Decision and scope: account, lifecycle stage, window, cutoff, decision owner.
  • Population reconciliation: account hierarchy, products, intended users, source completeness.
  • Signal ledger: raw values, denominators, baselines, direction, confidence, limits.
  • Contradictions: signals that disagree and plausible alternative explanations.
  • Outcome evidence: named customer outcome, current proof, missing evidence.
  • AI contribution: transformations or drafts produced, source anchors, human corrections.
  • Decision: investigate, maintain, escalate, or hold pending evidence.
  • Action register: owner, date, closure evidence, escalation rule.
  • Version history: method changes, prior status, reviewer, next verification date.

Sources

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

Explore this topic