Customer Health Score Operational Framework

Trigger

• CSM team lacks a data-driven way to prioritize at-risk accounts
• Churn is rising but aggregate retention metrics still look healthy → need a leading indicator
• Product usage data exists but isn’t tied to retention outcomes
• Want to automate QBR prep and proactive outreach
• Board/investor asks for churn prevention metrics beyond lagging churn rate

---

Step-by-Step Workflow

1. Map Outcomes to Signals

Pick signals that correlate with renewal, churn, or expansion in your historical data.

Signal categories:

• Product engagement: logins/week, active seats, feature adoption, API events, sessions per day
• Financial health: days to payment, invoice disputes, downgrade history, discount depth
• Support health: ticket volume, severity, CSAT/NPS/CSM score, response/resolution time
• Contractual timing: days to renewal, contract value, change in billing frequency

Rule: Cap at 8–10 signals. More than that collapses signal-to-noise and makes weighting impossible.

2. Define the Scoring Model

Two proven approaches:

• Weighted sum: score = Σ(weight_i × normalized_signal_i) — fast to launch, easy to explain.
• Churn probability (logistic): outputs 0–1 risk probability — more accurate once you have labeled churn data.

Start with weighted sum; migrate to probability once you have 12+ months of churn labels.

Normalization:

• Min-max: (x - x_min) / (x_max - x_min)
• Z-score if distribution is roughly normal
• Inverse ratio for "closer to bad" metrics (e.g., days to renewal: 1 - min_max(days))

Cap outliers at the 95th percentile to prevent a single abnormal week from tanking a score.

Weights: Derive from historical churn correlation, or initialize with equal weights and calibrate monthly. Example default weighting:

• Product engagement: 40%
• Support health: 25%
• Financial: 20%
• Contract timing: 15%

3. Set Thresholds and Buckets

Score range: 0–100.

• Red (0–39): High risk → escalate to CSM lead within 48h
• Yellow (40–69): At risk → CSM action plan within 7 days
• Green (70–100): Healthy → standard nurture, expansion outreach

Calibrate thresholds by overlaying historical churn rates. If the green bucket churns at >5%, raise the threshold. If red bucket churn is <2× baseline, lower it.

4. Build the Automation Pipeline

• Ingest: CRM (Salesforce/HubSpot), product analytics (Amplitude/Mixpanel), support (Zendesk), billing (Stripe/Chargebee)
• Compute: Weekly batch for most SaaS; daily for high-touch enterprise
• Outputs: per-account score, trend (Δ from prior period), top contributing risk factors, recommended CSM play

5. Tie to CSM Playbooks

Red account:

• Auto-create task: "Executive health review call within 48h"
• Highlight top 2–3 risk factors
• Suggest stack-ranked plays: onboarding refresher, usage training, executive business review, discount/payment review

Yellow account:

• Automated best-practice email or in-app guidance
• Flag for next QBR agenda

Green account:

• Expansion playbook: new feature trial, seat expansion offer, case-study invitation

6. Validate and Calibrate

Monthly:

• Precision = (# churned or downgraded in Red) / (# total in Red)
• Lift = (Churn rate in Red) / (Baseline churn rate)
• Target: precision ≥ 60%, lift ≥ 3×

Quarterly:

• Re-evaluate weights based on new correlation data
• Add or drop signals if predictive power drops
• Check for data drift (e.g., a feature launch changes engagement patterns)

---

Formulas

Weighted Sum Health Score

Score = [ Σ (w_i × norm(s_i)) ] × 100

Where:

• w_i = weight for signal i (sum of weights = 1)
• norm(s_i) = normalized value 0–1
• Result scaled to 0–100

Min-Max Normalization

norm(x) = (x - x_min) / (x_max - x_min)

Use capped percentiles: set x_min to 1st percentile and x_max to 99th percentile to reduce outlier influence.

Churn Probability (Logistic Regression)

p(churn) = 1 / (1 + e^( - (β₀ + β₁·x₁ + β₂·x₂ + ... ) ))
Health = (1 - p(churn)) × 100

Precision and Lift

Precision = (# churned in Red bucket) / (# total in Red bucket)
Lift = Churn rate in Red bucket / Baseline churn rate

---

Worked Example

Company: PLG SaaS, $50/mo, 5,000 customers.

Signals & weights:

Signal	Weight	Direction
logins_last_14d	0.25	Higher = better
core_feature_used_last_7d	0.20	Higher = better
support_tickets_last_30d	0.20	Higher = worse (invert)
nps_score	0.15	Higher = better
days_past_due_invoice	0.10	Higher = worse (invert)
days_to_renewal	0.10	Closer = worse (inverse)

Customer values:

• logins: 2 (min=0, max=20) → norm = 0.10
• core_feature: 0 → norm = 0.00
• tickets: 2 (min=0, max=5, cap) → invert: (5-2)/5 = 0.60 → norm = 0.60
• nps: 5 (min=0, max=10) → norm = 0.50
• days_past_due: 0 → norm = 0.00 (inverted = 1.00)
• days_to_renewal: 45 (min=0, max=90) → inverse norm = 0.50

Score = (0.25×0.10) + (0.20×0) + (0.20×0.60) + (0.15×0.50) + (0.10×1.00) + (0.10×0.50)
= 0.025 + 0 + 0.12 + 0.075 + 0.10 + 0.05 = 0.370 → 37.0 → Red

---

Pitfalls

Pitfall	Why It Hurts	Fix
Metric gaming	Support or CSMs manipulate signal data to avoid "red" flags	Keep score internal; tie to process, not individual performance only
Tenure blindness	New accounts look identical to mature ones; zero usage may be expected early	Add tenure as a factor or segment weights by account age
Seasonal usage mismatch	Monthly-reporting customers look dead in weekly metrics	Align signal windows to actual customer usage cadence
Ignoring nulls	Missing NPS treated as zero scores healthy accounts down	Impute missing values with segment median, or exclude signal per account
Static weights	Product expansion or pricing changes shift drivers of churn	Revalidate weights quarterly using latest churn labels
Threshold drift	Changing customer mix makes original 0–39 red band inaccurate	Recalibrate buckets monthly based on actual churn distribution

---

Verification Step

1. Backtest precision and lift: Run the current model against the last 6 months. If precision < 50% or lift < 2×, recalibrate before wide rollout.
2. Sanity check top-risk list: The top 10% highest-risk accounts should contain ≥ 80% of accounts that churned in the last 30 days.
3. Data reconciliation: Sum of accounts in score output must equal total active customer count. Verify no duplicate account IDs and no missing billing-status fields.