An n8n flow that computes a composite customer health score combining product usage from Gainsight, CRM activity from HubSpot, and conversation sentiment from Gong. Writes the score back to Gainsight nightly, with three sub-scores and a one-line “why this changed” note. Replaces the static, gameable health score most CS teams rely on.

What you’ll need

n8n self-hosted or Cloud
Gainsight with PX or NXT for usage data
HubSpot for support and CSM activity
Gong for conversation transcripts
Claude API key for sentiment scoring
Definitions of high/medium/low for each sub-score

Setup

Define the model. Decide your sub-score weights — for example, 50% product usage, 30% CSM activity, 20% conversation sentiment. Document this in health-model.md so the rationale is auditable.
Import the flow. Pull health-score.n8n.json. The flow has four parallel branches feeding a final aggregation node.
Wire data sources. Gainsight branch pulls daily usage deltas. HubSpot branch pulls case volume and CSM-touch recency. Gong branch pulls call transcripts from the last 30 days and runs Claude sentiment classification per call.
Aggregate and write back. The aggregation node computes the composite, generates a one-line “why” using Claude, and writes the score, sub-scores, and reason to a Gainsight custom field. Run nightly at 2am.
Validate against churn. Backtest the model against the last 12 months of churn. If high-health accounts churned, your model is wrong.

How it works

Each branch normalizes its sub-score to a 0-100 scale. Usage uses a rolling 28-day window comparing current activity to the account’s own baseline, not absolute thresholds — this catches drops that absolute thresholds miss. Activity is recency plus frequency of CSM-customer touches. Sentiment runs Claude on each call and produces a rolling sentiment average, weighted by recency.

The aggregation is a weighted sum, but the output isn’t just a number. Claude generates a one-sentence explanation: “Health dropped from 78 to 64 because product usage is down 35% week over week despite three CSM touches.” That sentence is what makes the score actionable for CSMs.

The flow is idempotent. Re-runs produce the same output for the same inputs.

Watch-outs

Garbage in, garbage out. If your usage events aren’t tagged correctly in Gainsight, no model fixes it. Audit data quality first.
Backtest before trust. Run the model retroactively for a quarter and check whether it predicted the churn you saw. If it didn’t, re-weight.
Account-level vs. user-level usage. Some products have one power user and 99 dormant users; that account is healthy or unhealthy depending on context. Build segment-specific weights if needed.

Stack

n8n — orchestration and aggregation
Gainsight — usage data and destination for the score
HubSpot — support and activity signal
Gong — conversation transcripts for sentiment
Claude — sentiment classification and reason generation