Customer Support

How to watch for churn risk accounts with AI

Customer Support3 AI tools6 steps6 friction points

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

Watching for churn risk accounts means identifying customers who are quietly heading toward cancellation before they actually leave. It typically involves reading signals across multiple sources — support ticket frequency, login patterns, payment failures, NPS scores, email engagement drops, and unanswered outreach — and synthesizing them into a judgment call about who needs attention now. For most operators, this lives nowhere, spread across a CRM, a support tool, a billing system, and their inbox.

The reason this feels like an AI-shaped problem is that the pattern recognition is genuinely complex but also repetitive. A human can learn to spot a customer who's gone quiet, filed three support tickets in two weeks, and skipped their last check-in call. The question is whether you can systematize that judgment so you're catching it at account #47, not just account #3 where it's obvious. AI is good at reading a set of signals and returning a structured risk assessment — which is exactly what this workflow requires.

ChatGPT, Claude, and Gemini can help meaningfully here — if you bring the data to them. You can paste in a CSV of customer activity, write a prompt that defines your churn risk criteria, and get a prioritized list with reasoning. Claude in particular handles long, structured data pastes well and can score accounts against multi-factor rubrics you define. The output is often genuinely useful. The friction is in the setup, the freshness of the data, and the fact that you're doing it manually every time.

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 Export your customer account data manually — pull a CSV from your CRM (HubSpot, Salesforce, Capsule), your billing tool (Stripe), and your support platform (Intercom, Zendesk). You'll want columns like last login date, open ticket count, days since last reply, MRR, and subscription status.

2 Open Claude or ChatGPT and define your churn risk rubric in the system prompt or first message. Specify what signals you weight most heavily — e.g., no login in 21+ days, 2+ unresolved support tickets, payment failure in the last 30 days, or no response to the last two outreach emails.

3 Paste your exported data into the chat. If the dataset is large, you may need to trim it or use ChatGPT's file upload feature (Code Interpreter) to handle tabular data without hitting context limits.

4 Prompt the model to score each account against your rubric and return a prioritized list: high risk, medium risk, low risk — with one-line reasoning per account explaining which signals triggered the flag.

5 Copy the output into a spreadsheet or Notion doc. Manually assign owners for each high-risk account, draft follow-up messages, and decide on next actions for each flagged customer.

6 Repeat this entire process next week, next month, and every time you want a fresh read — re-exporting data, re-pasting, and re-running the prompt from scratch.

Prompts you can copy

Here is a CSV of our customer accounts. Score each one for churn risk (high / medium / low) based on these signals: no login in 21+ days, 2 or more open support tickets, any payment failure in the last 30 days, and no reply to outreach in 14+ days. Return a ranked list with one sentence of reasoning per account.

We define a high-risk account as any customer who has done at least two of the following in the last 30 days: logged in fewer than 3 times, filed a complaint, missed a payment, or gone silent after previously active email engagement. Apply this rubric to the accounts below and flag the top 10 for immediate follow-up.

Read this list of customer support tickets and tell me which company names appear most frequently, which issues are repeated, and whether any single account has filed 3 or more tickets in the past two weeks. Format as a table.

Based on the account data below, draft three short follow-up email templates I can send to high-risk customers: one for an account that's gone quiet, one for an account with repeated support issues, and one for an account that just had a payment fail.

Which of these accounts shows the combination of declining login frequency AND recent support escalation? List them in order of urgency and explain what makes each one a priority.

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 Stripe, HubSpot, or Intercom data — every run requires a manual export, copy, and paste before the model can do anything.

Context window limits mean large account lists get truncated. If you have 200+ accounts with multi-column activity data, you're either splitting the job across multiple sessions or summarizing in ways that lose signal.

The rubric you carefully defined in last week's prompt isn't saved anywhere. Next run, you're reconstructing it from memory or hunting through chat history to find the version that actually worked.

Nothing triggers automatically. There's no alert when a high-value account hits three risk signals at once — you only find out when you remember to run the prompt again.

Output structure drifts between runs. The scoring format, column names, and reasoning style the model returns this week won't exactly match what you got last month, making it hard to track changes in account risk over time.

Acting on the output is entirely manual — copying flagged accounts into a doc, assigning owners, and drafting follow-ups all happen outside the AI tool, in separate tabs, with no connection back to your CRM or email.

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 — you describe the churn monitoring system you want, and an agent builds it as a persistent app running continuously against your live customer data, not a prompt you re-run from scratch every week.

Connect Stripe and HubSpot or Capsule CRM once — Starch syncs your subscription data, customer records, and deal activity on a schedule, so your churn risk app always reflects current account status, not last Tuesday's export.

Start from the CRM starter app, then describe your churn risk layer in plain English: 'Flag any account that hasn't logged in for 21 days, has 2+ open support tickets, or had a payment failure this month.' Starch builds the scoring logic and keeps it running.

Connect Intercom or Zendesk from Starch's integration catalog — the agent queries live ticket data when your churn risk app runs, so support escalation signals feed directly into account scoring without a separate export step.

Get a live dashboard that shows your at-risk accounts ranked by severity, updated automatically — so instead of a weekly prompt session, you open one view and see exactly who needs a call today.

Customer Support Agent (coming soon) will let you close the loop entirely — flagged accounts can trigger automated check-in emails and escalation workflows without you manually drafting and sending each one.

Automations can push churn risk alerts to Slack on a schedule — every Monday morning, Starch surfaces the top five accounts that crossed your risk threshold that week, with one-line context on why each one is flagged.

Get closed-beta access →

Toolkit