Spotting members at risk of leaving with AI

Members rarely announce that they are about to cancel. They drift. Logins thin out, a course stalls halfway, emails go unopened, and one day the cancellation notice arrives. By then the decision is made. The signs were there all along, but they were buried in exports and dashboards you did not have time to study.

This is a job AI is genuinely good at: looking through a table of activity data and describing the patterns that separate members who stayed from members who left. It will not predict the future with certainty, and it should never decide who gets contacted. But it can turn a spreadsheet you would never read into a short, testable list of warning signs.

Why patterns are hard to see by eye

You probably have hunches already. “People who skip the first live call tend to leave.” Some hunches are right, and some are based on the two or three members you happen to remember. Churn, which simply means the members who cancel or lapse over a period, is usually driven by combinations of small signals: fewer logins plus a stalled course plus a billing page visit. Those combinations are exactly what people miss and what AI analysis can surface quickly.

Prepare an anonymized export

Start with whatever your membership platform can export. You want one row per member and a column for each signal, plus a column showing whether they are still a member. Useful columns include:

• Months since joining and membership level
• Logins in their first month and in their most recent month
• Days since last login
• Lessons completed, events attended and community posts
• Support tickets opened
• Status: active or cancelled, with the month they cancelled

Before this file goes anywhere near an AI tool, remove names, email addresses, phone numbers and anything else that identifies a person. Replace each member with a code such as M001, and keep the key that links codes to people in a separate file that stays on your own computer. Even with anonymized data, use a paid business plan and check that the data and training settings keep your uploads out of model training. There is more on this in protecting member privacy when you use AI tools.

If you are not confident cleaning a spreadsheet, the techniques in AI for spreadsheets will help, and you can ask an assistant to write the formulas without ever showing it real data.

Ask AI to compare leavers and stayers

Assistants such as ChatGPT, Claude, Gemini and Microsoft Copilot can read an uploaded spreadsheet and run simple analysis on it. Upload your anonymized file and give a precise brief:

This file contains anonymized activity data for members of [describe your membership]. Each row is one member. The Status column shows whether they are active or cancelled. Compare the cancelled members with the active members. Tell me which signals differ most between the two groups, especially in the 30 days before cancellation. Show the numbers behind each finding and how many members each finding is based on. Flag any finding based on fewer than [20] members as weak. Then suggest three simple warning rules I could check each week, written in plain English. Do not guess at reasons members left; describe only what the data shows.

Imagine a hypothetical online chess academy called Knight School. Its owner, Farah, ran this analysis and got back something like: “Cancelled members attended an average of 0.4 group lessons in their final month, compared with 2.1 for active members. Members who completed no puzzles for 21 days were much more likely to appear in the cancelled group.” That is useful because it is specific, it shows the counts and it is something she can check.

Check the pattern before you trust it

AI analysis can be confidently wrong. It may miscount rows, misread a column or present a coincidence as a pattern. So before you act, verify:

1. Spot check the numbers. Filter the spreadsheet yourself for one or two findings and confirm the counts match.
2. Test on a different period. If the warning rule was found using members from the first half of your data, check whether it also picks out the people who left in the second half.
3. Ask whether it makes sense. A pattern that fits your knowledge of the membership is more trustworthy than a strange one. Treat surprising findings as questions, not answers.

Farah found that the group lesson signal held up, but a suggested link between profile photos and cancellation vanished when she looked at the next few months. She dropped it.

Turn the rules into a weekly flag

Once you have two or three rules you trust, make them routine. Ask the AI to write spreadsheet formulas that mark any member who meets a rule, for example no login in 21 days and fewer than two lessons in the past month. Each week, paste in a fresh anonymized export and let the formulas produce a short list of member codes. Then match the codes to real people using the key file you kept locally.

If you want to automate this later with a tool like Zapier, Make or n8n, keep the same principle: the automation produces a list; a person reviews it.

A person decides what happens next

A flag is a reason to look, not a verdict. Before anyone is contacted, read each flagged member's situation. Someone on a planned break, a member who prefers the email digest over logging in, or an annual member who is simply between projects may all trip a rule without being at risk. For the members who genuinely seem to be drifting, choose a response yourself: a helpful personal note, a pointer to the lesson they stalled on, or an invitation to a small group call. Keep any AI-drafted outreach under your own review before it is sent.

For the non-AI side of reading warning signs and responding with care, see spotting members at risk of cancelling before they leave.

Your first steps

• Export one row per member with activity signals and an active or cancelled status.
• Strip identifying details and replace members with codes; keep the key locally.
• Upload to an AI tool on a business plan with training turned off, and use the prompt above.
• Spot check the numbers and test each rule on a different period.
• Build weekly flags from the two or three rules that hold up.
• Review every flagged member yourself before any outreach.
• Rerun the analysis every few months as your membership changes.