Customer Support

How to score customer health with AI

Customer Support3 AI tools6 steps6 friction points

Want an agent to handle this on your live data? See the Starch version →

Customer health scoring means assigning each account a structured signal — a number, a grade, a traffic light — that tells you which customers are thriving, which are drifting, and which are about to churn. It pulls from product usage, support tickets, billing history, communication cadence, and any other signal that correlates with retention in your business. Most operators know they should be doing this. Few have a system that actually stays current.

The workflow feels like an AI problem because it's fundamentally a pattern-recognition and summarization task. You have data in several places — a CRM, a support inbox, Stripe, maybe a product analytics tool — and you need someone to read across all of it and render a verdict on each account. That's exactly what large language models are good at: given the right inputs, they synthesize fast and surface the customers you'd otherwise miss until it's too late.

ChatGPT, Claude, and Gemini can contribute meaningfully here today. You can paste in a customer's recent activity, support history, and billing events and get a structured health assessment back in seconds. Claude in particular handles nuanced context well and can apply a scoring rubric you define. The models are good enough that the bottleneck isn't AI quality — it's data assembly. Getting the right inputs into the prompt, consistently, for every account, is where most operators get stuck.

Skip the manual work → See the Starch version

Customer Support3 AI tools6 steps6 friction points

AI walkthrough

How to do it with AI today

A practical walkthrough using ChatGPT, Claude, and other off-the-shelf LLMs — what they're good at, what you'll have to do by hand.

Tools that work for this

ChatGPTClaudeGemini

Step-by-step

1 Define your health scoring rubric by pasting your customer segments and retention data into Claude and asking it to propose 4-6 weighted signals — e.g., last login date, open ticket count, days since last payment, NPS score. Refine the output until it matches how you actually think about account health.

2 Export your customer data manually: pull a CSV from your CRM for contact and deal activity, download a Stripe or billing export for payment history, and export your support inbox history from Intercom or Zendesk.

3 For each account (or a batch of 10-20 at a time), paste the combined activity data into Claude with a prompt that includes your scoring rubric. Ask it to return a structured JSON or table with a health score, a grade, and a 1-2 sentence rationale per account.

4 Copy the scores and rationales back into a spreadsheet or your CRM manually. Tag accounts as Healthy, At Risk, or Critical based on the output, and add a 'last scored' date column.

5 For accounts flagged At Risk or Critical, paste the full account context into ChatGPT and ask it to draft a personalized outreach message or a list of suggested next actions for the account owner.

6 Repeat this entire process next month by re-exporting all your source data and re-running the same prompts. Set a calendar reminder so it actually happens.

Prompts you can copy

Here is my customer health rubric: [paste rubric]. Here is the account activity for [Company X]: last login 45 days ago, 3 open support tickets, paid on time last 6 months, NPS 6. Score this account on a 0-100 scale and explain the rating in 2 sentences.

I'm going to paste data for 15 accounts. For each one, return: account name, health score (0-100), grade (A/B/C/D/F), and a one-sentence reason. Use this rubric: [paste rubric]. Here is the data: [paste CSV rows].

Based on this at-risk account's history — [paste support tickets, last login date, billing notes] — what are the top 3 actions I should take in the next 7 days to improve retention odds?

Suggest a customer health scoring model for a B2B SaaS company with 50-200 accounts. Include 5-6 signals, suggested weights, and thresholds for Healthy, At Risk, and Critical categories.

Here are the last 90 days of support emails from [Company X]: [paste emails]. Summarize the sentiment trend, flag any escalating issues, and tell me whether this account's health is improving or deteriorating.

Reality check

Where this gets hard

The walkthrough above works — until your numbers change, the LLM hallucinates, or you have to re-paste everything next month.

No live connection to your CRM, Stripe, or support tools — every scoring run starts with a fresh export, which means you're scoring against data that's already days old before you start.

Context limits cap how many accounts you can score in a single session; with rich per-account data, you're processing 10-20 at a time, not your full book of business.

Scores drift between runs — your rubric prompt and the model's interpretation both shift slightly each month, so a 72 last quarter isn't the same as a 72 this quarter.

Nothing persists. The health scores you generated last month live in a spreadsheet tab you'll have to dig up, not in a system that tracks trajectory over time.

Synthesizing across data sources is manual. Correlating a support ticket spike in Zendesk with a payment failure in Stripe and silence in email requires you to pull three exports and paste them together before the AI sees any of it.

Writing personalized outreach for at-risk accounts requires a separate prompt chain per account — there's no automated handoff from 'flagged at risk' to 'draft the save email and queue it.'

Tired of the friction?

Starch runs the whole workflow on live data — no copy-paste, no hallucinated numbers, no re-prompting next month.

See the Starch version →

Starch alternative

The same workflow on Starch

Starch is an agentic operating system. For customer health scoring, that means an agent builds a persistent app connected to your live CRM, support, and billing data — one that scores accounts on a schedule and surfaces the at-risk list without you re-running prompts each month.

Connect HubSpot or your CRM from Starch's integration catalog and Stripe as a scheduled-sync provider. Starch pulls live deal activity, contact history, and billing events automatically — no monthly export required before you can score anything.

Connect Intercom, Zendesk, or Freshdesk from Starch's integration catalog so support ticket volume and sentiment flow into the health model alongside CRM and billing data, giving each score a fuller picture.

Describe your scoring model in plain English — 'score each account 0-100 based on last contact date, open ticket count, payment history, and product engagement' — and the agent builds the health dashboard that recalculates on a schedule.

Use the CRM starter app as your base, then extend it: tell Starch to add a health score column, a trend line showing last 90 days, and a filtered view that surfaces only At Risk and Critical accounts each Monday morning.

Set up an automation — 'every Friday, identify any account that dropped more than 10 points this week and draft a save email for the account owner' — and Starch runs it on live data, not a spreadsheet you remembered to update.

Customer Support Agent (coming soon) will close the loop further — handling inbound tickets automatically and feeding resolution data back into the health score so accounts don't stay flagged after their issue is resolved.

Get closed-beta access →

Toolkit