Building a Customer Health Score Model with AI

Customer health scores are supposed to be early warning systems, but most implementations are lagging indicators dressed up as predictions. The typical health score adds up login frequency, support ticket volume, and NPS scores — all things that tell you what already happened, not what is about to happen. At GTM11, we build AI-powered health scoring models that genuinely predict churn before it manifests, giving customer success teams the lead time they need to intervene.

Why Traditional Health Scores Fail

The standard health score model assigns weights to 5-10 metrics and produces a green/yellow/red status. The problems are structural:

• Linear scoring misses complex patterns: A customer might have high login frequency (green) but declining feature adoption (red) and increasing support tickets (yellow). The linear average says "yellow" when the pattern actually screams "imminent churn."
• Static weights ignore customer segments: Login frequency matters differently for a 10-person startup versus a 10,000-person enterprise. Static weights cannot accommodate this variation.
• No temporal analysis: A customer who logged in 50 times last month is healthy, right? Not if they logged in 200 times the month before. Trajectory matters more than absolute values.
• Missing qualitative signals: Sentiment in support tickets, tone in emails, and engagement quality in meetings all carry predictive signal that numerical scores miss.

The AI Health Score Architecture

Our model processes four data dimensions through Claude AI to produce a composite health score with explanatory reasoning:

Dimension 1: Product Usage Patterns

We pull product usage data from your analytics platform (Amplitude, Mixpanel, Pendo, or custom events) and analyze it for trends rather than snapshots:

• Feature adoption breadth: What percentage of purchased features are being used?
• Usage trajectory: Is usage increasing, flat, or declining over 30/60/90-day windows?
• Power user concentration: Is usage concentrated in 1-2 users or distributed across the team?
• Session depth: Are users performing meaningful workflows or just logging in and bouncing?

Dimension 2: Support and Engagement

We ingest support ticket data from Zendesk, Intercom, or your help desk and process it with Claude AI for sentiment analysis:

• Ticket volume trend (not just count, but direction)
• Ticket severity distribution — an increase in P1 tickets is more concerning than P3 increases
• Sentiment analysis of ticket content — frustrated language patterns are early churn indicators
• Resolution satisfaction — are tickets being resolved to the customer's satisfaction?

Dimension 3: Relationship Signals

These are the qualitative signals that most scoring models ignore:

• Executive sponsor engagement: Has the executive sponsor attended recent QBRs?
• Champion risk: Has your primary champion changed roles or left the company?
• Meeting attendance: Are meetings being rescheduled, shortened, or cancelled?
• Renewal conversation timing: Are they engaging proactively or avoiding renewal discussions?

Dimension 4: Business Context

External signals that indicate organizational changes:

• Company layoffs or restructuring (monitored via LinkedIn and news alerts)
• Leadership changes in the department that owns your product
• Competitor product announcements that might attract your customer
• Funding or acquisition events that could change priorities

How Claude AI Processes the Score

Every week, an N8N workflow collects data from all four dimensions for each customer and sends it to Claude AI with a structured prompt. The prompt includes:

• Historical data for the past 90 days across all dimensions
• The customer's segment (SMB, Mid-Market, Enterprise) for context-appropriate weighting
• Historical examples of customers who churned and the patterns they exhibited

Claude returns a JSON response with:

• Overall health score (0-100)
• Individual dimension scores
• Risk level (Low, Medium, High, Critical)
• Top 3 risk factors with explanations
• Recommended CSM actions
• Confidence level in the assessment

The explanatory reasoning is what makes this model actionable. Instead of just seeing a "62" health score, the CSM sees: "Health declining due to 40% drop in feature adoption over 30 days, combined with two escalated support tickets about the reporting module. Executive sponsor has not attended last two QBRs. Recommended action: Schedule a focused session on reporting module with hands-on training, and reach out to executive sponsor directly."

Implementation with Your Tech Stack

The data pipeline connects your product analytics, support platform, CRM, and communication tools through N8N:

1. Weekly cron trigger in N8N initiates the scoring cycle
2. Parallel API calls fetch data from Amplitude/Mixpanel, Zendesk/Intercom, Salesforce, and LinkedIn
3. Data aggregation node combines all inputs into a structured customer profile
4. Claude AI node processes the profile and returns the scored assessment
5. Results are written to Salesforce custom fields and a Google Sheet dashboard
6. High-risk alerts are posted to Slack with full context

Calibrating the Model

Accuracy improves over time as you feed outcomes back into the model. Every time a customer churns or renews, we log which signals were present 30, 60, and 90 days before the event. After 6 months of data, the model becomes highly predictive because Claude AI has real examples of what churn looks like in your specific customer base.

We have seen our AI health score models predict churn with 78% accuracy at 60 days out, compared to 45% accuracy for traditional weighted scoring. That early warning is the difference between saving an account and writing a post-mortem.