Customer Health Scoring Workflow: Churn Risk You Can Act On
A customer health score is only useful if it moves before the customer leaves and tells the CSM what to do. The workflow computes one score per account every week from usage trend, support experience, billing behaviour, engagement and sentiment, alerts on drops with the cause, and has an AI agent draft the check-in — so the CSM's day starts with the accounts that need them.
Written by Max Zeshut
Founder at Agentmelt · Last updated Sep 11, 2026
The problem
Churn surprises. The renewal call reveals the champion left months ago, usage halved in the spring, and three support tickets went badly. All of it was visible in different systems and nobody assembled it.
What changes when it runs
Every account has a score with a trend and the components behind it. A drop of more than a set number of points in a week alerts the owner with the reason (“admin user inactive 21 days; two negative-sentiment tickets”) and a drafted message. Renewal forecasts stop being a guess because the score is calibrated against actual churn every quarter.
Trigger, then 8 steps
Trigger
Schedule Trigger (weekly) + event triggers
Weekly scoring; immediate re-score on high-severity events such as a champion's user being deactivated, a payment failure, or a very negative ticket.
Collect the components
HTTP RequestUsage (active users, sessions, key actions, trend vs previous 8 weeks), support (ticket volume, sentiment, resolution time), billing (late payments, downgrades), engagement (QBR attendance, email responses), and relationship (champion still active, number of contacts).
Score each component
CodeEach component becomes a 0–100 sub-score with segment-specific thresholds (an enterprise account's normal usage differs from an SMB's). Weights live in a sheet.
Analyse ticket sentiment
Sentiment AnalysisRecent tickets are scored for sentiment and frustration markers; a single furious ticket outweighs three neutral ones.
Compute score and trend
CodeWeighted score, 8-week trend, and the top two components that changed. Score history is stored per account.
Detect drops and thresholds
IFDrop above the threshold, or score crossing into red, creates an alert; accounts already in an active save play are not re-alerted.
Draft the intervention
AI AgentFrom the components and the account's history the agent drafts a check-in that names the observed change without sounding like surveillance, and suggests the play (re-onboard a new admin, executive check-in, training session).
Alert the owner and update the CRM
SlackOwner gets the alert with score, cause and draft; the CRM health field and a timeline note are updated so the score is visible in every view.
Calibrate quarterly
Schedule TriggerCompares scores from 90 days ago to actual churn and downgrades, reports precision/recall by score band, and proposes weight changes for RevOps to approve.
Data it touches
- Product analytics (usage and adoption)
- Help desk tickets and CSAT
- Billing (payment behaviour, plan changes)
- CRM activity and contacts (champion status)
- Meeting and email engagement
Guardrails
- Scores show their components; no black-box number.
- Alerts are suppressed for accounts already in a save play, so CSMs are not double-pinged.
- Outreach is drafted for a person to send; nothing goes to the customer automatically.
- The score is calibrated against real churn quarterly, and the calibration report is visible to leadership.
Health scores fail when they are static
Most health scores are a one-time formula that nobody revisits, so they drift: the score says green while the account churns, and CSMs stop trusting it. Two design choices fix that. The score shows its components, so a CSM can see whether 'red' means usage or sentiment and act accordingly. And the score is calibrated every quarter against what actually happened — how many accounts that were red churned, how many green ones did — with the result reported and the weights adjusted. A score that is measured stays useful.
The champion signal
The single most predictive signal in B2B SaaS is whether the person who bought and championed the product is still there and still active. It is also the one most CRMs cannot see. The workflow checks whether the champion's user account is active and whether their login recency has changed; a deactivated champion drops the relationship component immediately and triggers a play focused on finding and onboarding the replacement. Teams that add only this signal see their surprise churn fall noticeably.
From score to action
The alert is not the point; the play is. Each cause maps to a play: usage drop → training or re-onboarding; support frustration → executive check-in and a ticket review; champion loss → new-admin onboarding; billing trouble → finance-to-finance conversation. The agent's drafted message opens the play with specifics, and the CRM note records which play ran, so calibration can also tell you which plays work.
Tools in the stack
| Tool | Role in this workflow |
|---|---|
| n8n | Collection, scoring, alerting, calibration |
| Mixpanel / Amplitude / PostHog | Usage components |
| Zendesk / Intercom | Support components |
| Claude | Sentiment analysis and intervention drafts |
| HubSpot / Salesforce | Score field, alerts, timeline |
Want this running without building it?
Automation workflow
$297/month
We set up, host and maintain this workflow on n8n and connect it to your tools. Setup included, cancel monthly, you keep the JSON.
Custom build
$4,500–8,000 one-time
Your systems, your rules, your edge cases. A one-off build on Claude and n8n, delivered with documentation and a walkthrough.
Covers up to 2,000 accounts across one analytics, one help desk and one CRM. Predictive (ML) churn models trained on your history are a custom build once you have 12+ months of outcome data.
Frequently asked questions
How many data sources do we need?
Usage plus support plus CRM is the minimum that produces a meaningful score. Billing and engagement improve it. Start with what you have; the weights adapt.
Is this a predictive churn model?
It is a calibrated rule-based score, which is what most teams should run first: transparent, tunable, and accurate enough. A trained model needs a year of outcome data and is offered as a custom build once that exists.
Can the score show in HubSpot or Salesforce?
Yes — it writes to a custom property on the company/account with the trend and top causes, and posts a timeline note on each change.
Calculator
Put your own numbers in
Estimate the cost of the current process and what this workflow changes.
The pillar
AI Customer Success Agent
Automate onboarding, monitor customer health scores, identify expansion opportunities, and prevent churn.