Churn Prediction
Know which accounts may leave, and why.
Churn Prediction
Customer Support & Success
- RuleDone:
Every Monday, or when a strong warning sign arrives.
- RuleDone:
It loads your approved signals: payments, usage, tickets, engagement, champion changes.
- RuleDone:
It scores each account and marks "act now" cases.
- RuleDone:
It picks the top accounts that fit your CSMs' capacity.
- AIWaiting for approval:Waiting for you
AI writes plain reasons; rules check them; tasks and the digest go out.
- RuleNot started:
After 90 days it records who renewed or left, vs a control group.
For example: A 900-account HR-software company gets its Monday list.
Waiting for you. Nothing has changed yet.
Review for important actions. Customer-facing and record-changing actions are reviewed first.
What is Churn Prediction?
In shortChurn Prediction is an AI automation for Customer Support & Success teams that works with Gmail, HubSpot, Outlook, Slack, Stripe and Zendesk. Every Monday, your CS team gets a short list of accounts most at risk of cancelling, with plain reasons and next steps from your playbook. It is built for head of Customer Success or founder, CS lead and CSMs.
Last reviewed
Key facts
- Department
- Customer Support & Success
- Starts when
- Every Monday at 06:00 (your time), or right away on a failed payment, a cancel-at-period-end, or a HubSpot deal moving to "Renewal at risk"
- Works with
- Gmail, HubSpot, Outlook, Slack, Stripe, Zendesk
- Run modes offered
- Shadow, Draft, Assisted and Auto
- Human control
- Review for important actions
- Status
- Available now
The problem today
Churn is noticed when the cancel email or failed payment arrives. Today: CSMs check HubSpot, Stripe and Zendesk separately for 2–4 hours a week and still miss quiet accounts with no logins, tickets or replies.
What it does
Each week it scores every paying account from billing, CRM, support tickets and (optional) product usage. It uses a clear points model, or a simple model fitted on your own history once it passes a back-test. It picks the riskiest accounts within your team's capacity, writes a HubSpot task and risk score for each, and posts a digest to the CS lead. Strong signals such as a failed payment trigger an "act now" task within minutes, and optional check-in drafts are saved for a CSM to send.
How does Churn Prediction work?
How it works- Rule you set
- AI step
- A person decides
- Step 1Rule
Every Monday, or when a strong warning sign arrives.
- Step 2Rule
It loads your approved signals: payments, usage, tickets, engagement, champion changes.
- Step 3Rule
It scores each account and marks "act now" cases.
- Step 4Rule
It picks the top accounts that fit your CSMs' capacity.
- Step 5AI
AI writes plain reasons; rules check them; tasks and the digest go out.
- Step 6Rule
After 90 days it records who renewed or left, vs a control group.
What you get
Key features- Default points: payment failed 25, usage down 30% 20, no engagement in 45 days 15, 3+ urgent tickets 15, CSAT below 3 10, champion left 15.
- "Act now" signals: cancel at period end, 2+ failed payments, renewal within 30 days with no engagement in 45 days, champion left.
- Capacity: 15 accounts per CSM per week by default, ranked by risk × MRR.
- Monthly back-test, quarterly re-fit and a drift alert.
- Bias check across company segments (size, region, plan) before go-live.
- 10% control group for 6 months to measure real retention lift.
How much does it do on its own?
Run modesPick a mode to see what Churn Prediction would do at each step.
Every automation starts in Shadow. Move it up when you are ready, and back down at any time. How run modes work
Churn Prediction in Shadow
Runs on real data and shows what it would have done. No app is changed and no message is sent.
- Rule
Every Monday, or when a strong warning sign arrives.
Watches for it
- Rule
It loads your approved signals: payments, usage, tickets, engagement, champion changes.
Runs on real data, read only
- Rule
It scores each account and marks "act now" cases.
Runs on real data, read only
- Rule
It picks the top accounts that fit your CSMs' capacity.
Runs on real data, read only
- AI
AI writes plain reasons; rules check them; tasks and the digest go out.
Shown to you as a suggestion
- Rule
After 90 days it records who renewed or left, vs a control group.
Shown to you as a suggestion
Shadow vs Auto for Churn Prediction
Auto is the most it can do on its own. Here is each step side by side.
| Step | Shadow | Auto |
|---|---|---|
| In short | Runs on real data and shows what it would have done. No app is changed and no message is sent. | Runs without approval, with limits, alerts and a kill switch. Unlocks after a trial period in Assisted. |
| Step 1Every Monday, or when a strong warning sign arrives. | Watches for it | Runs on its own within your limits |
| Step 2It loads your approved signals: payments, usage, tickets, engagement, champion changes. | Runs on real data, read only | Runs on its own within your limits |
| Step 3It scores each account and marks "act now" cases. | Runs on real data, read only | Runs on its own within your limits |
| Step 4It picks the top accounts that fit your CSMs' capacity. | Runs on real data, read only | Runs on its own within your limits |
| Step 5AI writes plain reasons; rules check them; tasks and the digest go out. | Shown to you as a suggestion | Runs on its own within your limits |
| Step 6After 90 days it records who renewed or left, vs a control group. | Shown to you as a suggestion | Runs on its own within your limits |
| True in every mode | Never sends outreach, offers discounts, credits or refunds, or changes subscriptions; Stripe is read only. | |
What it will never do on its own
Safety & human control- Never sends outreach, offers discounts, credits or refunds, or changes subscriptions; Stripe is read only.
- Never uses protected traits (age, gender, race, religion, nationality and similar) or their proxies; they are refused at setup.
- A model that has not passed the back-test is labelled "unvalidated, use as checklist" everywhere.
- Outreach drafts always need CSM approval and never contain offers or "you were scored"; ticket text is redacted and never stored.
- It escalates on high-value accounts (MRR above 2,000 by default), low data quality, drift and bias flags.
- You approve before it acts
- Your data stays yours
- It knows its limits
- Stop anytime
What will you see on your dashboard?
Results you can trackTargets are quality goals the automation is built to reach. Estimates are calculated from your own volumes.
- Tracked
Saved MRR (listed accounts that renewed, vs control group)
Measured on your own data
- Tracked
Churn rate trend
Measured on your own data
- Tracked
CSM action rate on tasks
Measured on your own data
- Tracked
Drafts approved vs discarded
Measured on your own data
- Target
Recall at capacity, rolling (target: ≥ 50% of churned accounts listed 30+ days before)
A goal it is built to reach
- Target
Back-test top-decile lift (target: ≥ 2.0)
A goal it is built to reach
- Target
Back-test recall at capacity (target: ≥ 0.40)
A goal it is built to reach
- Target
Reason fidelity (target: 100%)
A goal it is built to reach
- Target
Offer-free drafts (target: 100%)
A goal it is built to reach
- Target
Segment false-positive ratio (target: ≤ 1.5, setup gate)
A goal it is built to reach
A 900-account HR-software company gets its Monday list. An account with falling usage and no meeting in 50 days ranks third, with that reason in plain words and "book a check-in call" from the playbook. When Stripe reports a second failed payment on another account on Tuesday, an "act now" task appears within minutes.
Who is Churn Prediction for?
- Head of Customer Success or founder
- CS lead
- CSMs
Industries: SaaS & tech, Professional services, Agencies, Education & training
Also useful for: Analytics & Reporting
Connects to the apps you already use
Works withApp names are trademarks of their owners. Charviam connects to them; it is not a partner or reseller.
- Gmail
- HubSpot
- Outlook(optional)
- Slack
- Stripe
- Zendesk
Better together
Automations that hand work to Churn Prediction, or pick up where it stops.
What do you need to set it up?
Setup- Connect HubSpot, Stripe (restricted read-only key), Zendesk and Slack; optional product usage through the metric layer and Outlook for drafts.
- Approve the list of signals used (protected traits are refused).
- Set points, CSM capacity, ranking, the high-value MRR line and playbook steps.
- Choose the holdout % and whether to turn on outreach drafts (suggested after week 4).
- Run the back-test, then Shadow for 8 weeks or until 20 churn events.
Typical setup
- 01Connect your apps
- 02Configure the rules
- 03Run a watch-only trial
- 04Go live
Not included
- Automatic discounts, refunds or plan changes, and sending outreach.
- Consumer (B2C) scoring of individuals.
- Black-box machine-learning models.
Questions about Churn Prediction
FAQSee Churn Prediction on your own data
WalkthroughIt starts in Shadow, where it only watches and suggests, so nothing changes in your apps until you say so.