Churn Risk, Scored on Evidence: A Weekly Claude Agent for Renewals 90 Days Out

Write the scorecard: signals, weights, and what counts as evidence

Start with the churn you already had. Pull the last ten lost renewals and, for each, write down what was visible ninety days out. The list is usually short and repetitive: usage fell, a champion left or went quiet, tickets shifted from 'how do I' to 'why does it', the sponsor stopped attending, the customer asked about data export or a competitor's integration, the invoice was disputed. Those are your signals, and your own churn tells you how to weight them.

Group them into four families, usage, relationship, support, commercial, and give each signal a weight and a definition of the evidence that triggers it. 'Usage down' is a feeling; 'weekly active users down more than 30% over 8 weeks' is a signal the agent can compute and cite. The scorecard below is the one I start from, and the weights are the part every team tunes against its own losses.

Set the threshold and the tiers. A score above 60 is at risk and gets a brief; 40 to 60 is watch, listed on the board with its top signal; below 40 is healthy and appears only as a count. The threshold decides how many briefs a CSM gets a week, and a CSM who gets twelve briefs reads none of them.

Connect the sources and pull the renewal window

Connect HubSpot or Salesforce over MCP and confirm the agent can read renewal dates, ARR, the account team, contacts with roles and titles, and the activity log. The window is 90 to 120 days out, which is early enough to change an outcome and late enough that the signals are real. Accounts inside 90 days stay on the board with their last score; the save plan should already be running.

Connect Gong or Fireflies so it can read the last 90 days of calls per account: who attended, who used to attend and stopped, and the tone. Export usage and tickets as CSVs by account and week, the way the QBR prep play does, and drop them in the run's folder.

For champion changes, give the agent the contacts and titles from the CRM and let it compare to the call recorder's attendee lists; a sponsor who was on every call in Q1 and none in Q3 is a relationship signal with a source. Where you have an enrichment source that tracks job changes, add it; where you do not, the agent should say 'not verified' rather than guess.

• Reads: accounts renewing in 90-120 days: ARR, renewal date, owner, contacts (role, title), activity log
• Reads: usage export by account and week; ticket export by account (subject, priority, opened, resolved)
• Reads: last 90 days of calls per account: attendees, dates, transcript
• Writes: one brief per at-risk account, one renewal board, Slack posts. Never a CRM record

Write the brief so the score explains itself

The brief is the scorecard turned into a run plus one rule that makes it trustworthy: every scored signal cites the evidence that fired it, and a signal without evidence scores zero. The output for an at-risk account is one page: the score and its four components, the signals in order of weight with the source for each, the customer's own words where a call or ticket has them, and a save plan of three to five actions with an owner and a date. The CSM should be able to check every line in one click and disagree with the weight, never with the fact.

Make the agent conservative about interpretation. Falling usage over the summer at a school district is a calendar, not a churn signal, and the brief should surface that the agent noticed the pattern. One angry ticket is a ticket. The scorecard should require a trend, or a corroborating signal, before a single family pushes an account over the threshold, and the brief should show its reasoning so the CSM can see when the agent is wrong.

The save plan is a draft. The agent knows the signals and can propose the obvious responses: an exec-to-exec call when the sponsor went quiet, an enablement session when usage fell in one team, a call with the admin who asked about export. The CSM chooses, edits, and owns it. Tag the plan 'CSM to confirm' the way the QBR play does, so the boundary between assembly and judgment is visible on the page.

Every Monday at 7:00am {{TIMEZONE}}, run the renewal risk review.

Read from {{CRM}}: every account with a renewal date between 90 and 120 days from today. For each: ARR, renewal date, owner, contacts (name, title, role), activity log for the last 90 days. Also re-score accounts inside 90 days that were on last week's board.

For each account read: files/usage.csv (weekly active users and feature use by week), files/tickets.csv (subject, priority, opened, resolved, requester), and the last 90 days of calls via {{CALL_RECORDER}} (attendees, date, transcript).

Score each account against scorecard/churn-risk.md. Rules:
- A signal scores only if its evidence threshold is met; cite the evidence (usage weeks and numbers, ticket #, call date and speaker, CRM field). A signal without evidence scores 0.
- Note seasonal or structural explanations you can see (e.g. summer, a reorg the customer mentioned) beside the signal.
- No single family may push an account over the threshold alone; require a corroborating signal from a second family.

For each account scoring >= {{RISK_THRESHOLD}}, write briefs/{{DATE}}/{{ACCOUNT}}.md, one page:
1. Header: account, ARR, renewal date, days remaining, owner, score (and last week's).
2. Score breakdown: usage / relationship / support / commercial, with points.
3. Signals, highest weight first: the signal, the evidence with its source, the customer's words if any.
4. What might explain it: any benign explanation the data supports.
5. Save plan [CSM to confirm]: 3-5 actions, each with an owner and a date, matched to the signals.

Write board/{{DATE}}.md: every account in the window, sorted by score: account, ARR, days to renewal, score, change vs last week, top signal in six words.

Rules:
- Read only. Write nothing to the CRM, the call recorder, or the ticket system.
- Fields and quotes, never adjectives. 'WAU 38 -> 11 over 10 weeks' beats 'usage is declining'.
- Post each brief to its owner in Slack (DM or the account channel). Post the board to #{{CS_CHANNEL}}.

Run it Monday, brief the CSM, board the manager

Schedule the job for Monday 7am as a scheduled Claude Code job, so the CSM opens Slack to the briefs and the manager to the board. Each brief goes to the owning CSM directly, with the score, the top signal, and days to renewal in the first line. The board goes to the CS channel: every account in the window, sorted by score, with the week-over-week change, so the manager sees movement without opening thirty briefs.

Run the first one by hand with the CS lead and read every brief against their gut. Where the agent flagged an account the CSM knows is fine, look at the evidence: usually a seasonal pattern the scorecard needs to allow for, or a weight that is too high. Where the CSM names an account they are worried about that scored low, find the signal the scorecard lacks. Both fixes are lines in a file, and the second one is the more valuable, because that account is the one the old health score would have missed too.

Keep the play read-only. The agent reads renewal data and writes documents; the save plan is executed by people, logged in the CRM by people, and checked by next week's run against the activity log. An agent that starts 'reaching out' to a wobbling customer on its own is a churn signal in itself.

Check the save plan against what actually happened

The re-score each week is where the play earns its place. An account that was at 68 with a silent sponsor should, two weeks after the exec-to-exec call in the save plan, show the sponsor back on a call and the relationship score down. If it does not, the brief says so, and the CSM knows the plan is not working while there is still time to change it. The agent reads the activity log for the plan's actions and reports which happened and which did not.

The board's week-over-week column is the manager's view of the whole book. Accounts moving up in score get attention; accounts moving down confirm that the plan worked; accounts that sit high for four weeks with no movement are the escalation list. The manager spends ten minutes on the board instead of an hour asking each CSM how their renewals feel.

Post the outcomes back. When a renewal in the window closes, won or lost, the run notes it on the board with the score it carried eight weeks out. That is the calibration data for the next step.

Calibrate the scorecard against real outcomes every quarter

Every quarter, pull the renewals that closed and lay their scores at 90 days beside the outcome. A good scorecard puts most of the lost renewals above the threshold at 90 days and most of the retained ones below it. Where it missed, find the signal that was visible and not scored, and add it. Where it cried wolf, find the weight to lower or the benign explanation to encode. The scorecard is a file, and the calibration is a one-hour meeting with the data on the table.

Watch for the signal that predicts your churn and nobody expected. In one book it was the 'export our data' ticket; in another it was the invoice dispute; in another it was the champion's calendar going quiet six months before the title changed. Your own losses are the training data, and the agent reads all of them the same way every week, which is more than any human on the team can promise.

Then share the calibration with the CS team. A scorecard that has been tuned against ten quarters of real renewals is a credibility asset: the CSM trusts the brief because they watched it get right the account they had a bad feeling about, and that trust is what makes them open the brief on the Monday when it matters.