Customer Health Scoring

A customer health score is an attempt to compress the state of an account into a single number or color. Done well, it is the most useful operational instrument an AM team has — it triages attention, makes risk legible across the org, and turns retention from a reactive scramble into a managed portfolio. Done badly, it is theater: a green number on a dashboard while the account quietly slides toward churn.

The difference between a real health score and a vanity score is whether it has been validated against actual outcomes (renewal, expansion, churn) and whether it actually changes behavior. If neither is true, it is decoration.

A health score is a weighted composite of inputs that together estimate the probability of the account renewing and expanding. The output is typically a single number (0–100), a color (red/yellow/green), or both.

Two design philosophies exist:

• Predictive scores — built from historical data on what actually predicted churn and expansion in your customer base. Higher accuracy, harder to build, requires meaningful data on past renewals and churns.
• Heuristic scores — built from informed judgment about what *should* predict outcomes (adoption, engagement, etc.), with weights set by experienced CSMs and AMs. Easier to build, faster to deploy, but should be migrated toward predictive over time.

Most organizations start heuristic and evolve toward predictive as they accumulate enough renewal/churn data to validate the weights. The mistake is treating a heuristic score as if it were predictive — assuming the green color means low churn risk when it has never actually been validated against churn.

A robust health score draws from multiple categories so that no single signal can dominate. Typical inputs:

• Product usage — DAU/MAU, license utilization, depth of feature adoption, breadth of teams using the product.
• Engagement — meeting cadence with the buying team, response times, QBR attendance, executive sponsor engagement.
• Support activity — ticket volume, severity, time to resolution, escalations, NPS or CSAT trends.
• Stakeholder strength — Champion confirmed and active, executive sponsor engaged, multi-threaded coverage of the buying team, no single-point-of-failure relationships.
• Commercial signals — payment timeliness, contract maturity, expansion activity, willingness to be referenced.
• Outcome realization — the customer can articulate ROI, has measured outcomes against the original business case, has shared results internally.

The trap: weighting too heavily toward easy-to-measure signals (usage, tickets) and underweighting the qualitative signals that actually predict (Champion strength, outcome realization, executive engagement). The hardest signals to measure are often the most predictive.

A score built only from quantitative data is structurally blind. Two examples:

• An account with 95% license utilization, low ticket volume, and on-time payments looks green. But the original Champion left three months ago, the new VP has never met your team, and there is no executive sponsor relationship. The quantitative model misses all of this. The account churns at renewal with 60 days notice and the scoring system records 'unexpected' loss.
• An account with 40% license utilization and several open escalations looks red. But the customer is in the middle of a planned migration, the executive sponsor is highly engaged, and a 3x expansion is being negotiated. The quantitative model triggers a churn intervention that confuses the customer and damages the relationship.

The operational answer is to overlay a qualitative scorecard maintained by the CSM/AM on top of the quantitative score. Senior orgs require the AM to weigh in monthly with three qualitative inputs: Champion status (active/at risk/missing), executive sponsor engagement (active/lapsed/none), and outcome narrative (defensible/in development/missing). When the quantitative and qualitative diverge, the qualitative usually wins — and the divergence itself is a signal worth investigating.

A practical build sequence:

1. Pick 5–8 inputs maximum. More inputs feel rigorous; in practice they dilute the signal and make the score impossible to debug. Five well-chosen inputs outperform fifteen weak ones.
2. Weight by predictive value, not by what is easy to measure. If Champion strength predicts renewal more than license utilization, it should get a higher weight even though it is harder to capture.
3. Define thresholds for color bands. Green/yellow/red is more actionable than a continuous score. The bands should map to specific intervention cadences (see next section).
4. Validate quarterly. Pull the last 90 days of renewals, downsells, and churns. Did the score correctly classify them 60 and 90 days out? If not, adjust the weights and inputs.
5. Hold the model accountable. A health score that does not improve in accuracy over time is being maintained as theater. Either fix it or decommission it.
6. Keep the inputs visible to the AM. A black-box score that the AM cannot explain to their manager destroys trust in the system. Every score should be decomposable into the underlying inputs.

The score is an input to action, not an output of analysis. Tie each color band to a specific operating cadence:

• Green — monthly check-in, expansion scan quarterly, light-touch QBR. The risk is under-investing time on accounts that are quietly on a slow decline; audit green accounts that have been green for 4+ quarters for hidden complacency.
• Yellow — biweekly review, formal action plan with named owner and date, escalation trigger if no improvement in 60 days. Yellow is the most operationally important band — it is where intervention is still cheap and the outcome is genuinely changeable.
• Red — weekly war-room with CSM, AE, AM, executive sponsor; written recovery plan reviewed at the leadership level; explicit decision at 90 days whether to invest further or accept the loss. Red accounts that linger as red for 6+ months without resolution are usually evidence of organizational unwillingness to make the hard call, not of recoverable risk.

The portfolio view matters as much as the individual account view. If your portfolio shifts 10% from green to yellow in a quarter, that is a structural signal — about a product issue, a market shift, or a coverage problem — that deserves leadership attention regardless of which specific accounts moved.

A SaaS company built a health score with 14 inputs covering usage, support, NPS, and commercial activity. After two years it correlated weakly with actual churn (~52% accuracy at 90 days out — barely better than chance). A new CRO simplified the model to six inputs and added two qualitative AM-maintained signals (Champion status and outcome narrative). After the redesign, accuracy at 90 days improved to 81%. More importantly, the operational behavior changed: yellow accounts now generated specific action plans within seven days of being flagged, and the proportion of red accounts that recovered to green within 90 days rose from 18% to 47%. NRR improved from 102% to 119% over the following four quarters, with the bulk of the improvement attributed to earlier intervention on accounts the prior model had been classifying as green until the renewal quarter.