Churn-Risk Early-Warning Monitor

The problem this kills

By the time an account actually cancels, the warning signs have usually been flashing for months - logins quietly trailed off, a couple of angry support tickets, an invoice that went 60 days late, the champion who hired you left for a new job, a 3-out-of-10 on the last survey. The trouble is that no single CSM is watching all of those signals across all of their accounts at once. They live in different tools - the product analytics dashboard, the helpdesk, the billing system, a survey platform, somebody's notes - and nobody has time to manually cross-reference them every week. So the first time the team hears "we're not renewing" is the day the renewal is already lost.

When teams do try to track this, it's usually a heroic spreadsheet that one person maintains, full of hand-typed RAG (red/amber/green) statuses that mean slightly different things to different people, with no record of why an account was marked red last month or whether the scoring rules have quietly drifted since then. Two CSMs look at the same account and disagree. A manager asks "how did we decide this is a high-risk account?" and the honest answer is "gut feel."

This tool turns that scattered, gut-feel process into one consistent, explainable early-warning system - so every account gets scored the same way, every at-risk flag comes with the specific signals that fired, upcoming renewals float to the top, and a human still decides what to do about it before it becomes the official call.

What you'll build

A small internal web app, just for your customer success team, that:

• Imports your accounts and their signal data from CSVs - usage metrics, open/recent support tickets, payment timeliness, login/engagement, NPS or CSAT scores, exec/champion changes, whatever you actually track.
• Loads a health rubric you define - the list of signals, the weight each one carries, and the thresholds that turn a raw number into "good / watch / bad." The rubric is versioned, so you can change it deliberately without losing the history.
• Computes a health / churn-risk score for every account and explains it - showing exactly which signals fired and how much each one pushed the score, so nobody has to guess why an account is red.
• Sorts accounts into risk tiers (e.g. low / medium / high / critical) and weights accounts with an upcoming renewal higher, so the ones you can still save get attention first.
• Puts every at-risk account in front of a CSM to review - confirm the risk is real (or dismiss it as a false alarm with a reason) and assign a save play (the intervention they'll run).
• Stamps the rubric version on every score so results are reproducible and you can prove how any number was produced.
• Ships a risk dashboard, a clean CSV export in the columns your CRM or QBR deck expects, and an at-risk alert digest emailed to the team via Resend.

What's inside the Implementation Plan

The plan is a single markdown file you paste into Claude Code (a free AI coding agent). It walks the agent through building the whole tool, step by step, each step ending with a ready-to-paste prompt.

The most important part: the plan opens by interviewing you about your business. Before it writes a single line, the agent asks how you spot at-risk accounts today, which systems your signals live in, the exact column names and units in your exports, which signals actually predict churn for your product and how you'd weight them, what your risk-tier cutoffs are, how much extra weight an upcoming renewal should get, and your messiest edge cases - brand-new accounts with no usage history yet, seasonal customers, multi-product accounts, planned downgrades that aren't really churn. It reads a short tailored spec back to you, you confirm it, and only then does it build - so you get a scoring tool shaped to your book of business, not a generic health-score template you'd have to fight.

Inside you'll find:

• The discovery interview and how the agent turns your answers into the data model, the signal definitions, and your scoring rubric.
• The full build: database, login, CSV import for accounts and signals with duplicate guards, the rubric loader with versioning, the scoring engine with per-signal explanations, the renewal-weighting and tiering logic, the CSM review-and-assign-a-play screen, the dashboard, the alert digest, and the CSV export.
• The manager approval gate on each rubric version, the human review gate on each at-risk account, and the rubric-version stamp on every score.
• Verification steps so you can prove it works, and the CSV-export fallback so the tool is fully usable today even before you connect it to your product analytics, helpdesk, or billing APIs.

The governance it includes (this is the point)

This isn't a toy. The plan builds in the controls a real retention program needs:

• Login so only your team can see or touch anything.
• Row-level security so people only ever see the accounts, signals, and scores that belong to your organization.
• A complete audit trail - every import, rubric version, score run, CSM confirmation, dismissal, and assigned play logged with who and when.
• A hard human-in-the-loop gate - the tool computes the risk and drafts the at-risk list, but a CSM confirms each account and assigns a play, and a manager must approve a rubric version before it can drive the official risk list. The AI proposes; people decide.
• Versioned, reproducible scoring - every score carries the rubric version that produced it, so you can always answer "why is this account red, and under which rules?"
• Duplicate guards - dedupe on account ID so the same account can't be scored twice in a run and re-imports update rather than multiply.

Who it's for

Customer success managers, account managers, and CS/RevOps leaders who are responsible for retention and are currently spotting churn risk by gut feel, scattered dashboards, or a fragile spreadsheet - and want a consistent, explainable early-warning system without hiring a developer or buying a heavyweight customer-success platform. You don't need to write code. You need your signal exports, a first draft of how you'd score health, and an afternoon-to-a-weekend.

You've got this - paste the first prompt and let the agent interview you.