Churn signal watch

Churn is visible in usage data months before it appears in a cancellation. The agent's job is not prediction magic, it is watching a handful of honest signals and putting a name in front of a human while there is still time.

The job on one screen

Runs when	Weekly across accounts with a renewal in the next 120 days.
Reads	Active users, core action counts, support ticket sentiment, admin logins, seats used against seats bought.
Decides	Which accounts have moved outside their own normal, and how urgent that is.
Produces	A ranked list with the specific signal that fired, ready for a success manager.
Stops when	List produced. It never emails the customer.
Tools it needs	Product analytics, CRM, ticketing (read).
Autonomy to start at	Approve. It does the work and stops before the irreversible step.

Optional: load a real model

Run it

What it watches, and how far is too far

Tolerances are the whole design. Too tight and everyone mutes it, too loose and it misses the thing you built it for.

Signal	Baseline	Tolerance	First thing to check
Weekly active users	48	25%	Check whether a champion left. Compare against the admin login list.
Core action per week	1200	30%	Usage collapse without a user collapse usually means a workflow moved to another tool.
Seats used of seats bought	42	25%	Unused seats are the single strongest renewal risk signal and the easiest to act on.
Support tickets this month	3	100%	A spike is not always bad, but a spike plus falling usage is the classic pre-churn shape.

Where this one goes wrong

Seasonality read as decline

Every business has quiet months. Baseline against the same account last year, not last month.

Scoring instead of explaining

A risk score of 72 tells a success manager nothing. Name the signal and the number behind it.

Acting automatically

An automated we noticed you are not using us email to a happy customer is worse than silence.

How you would know it is working

Measure	Why that one
Churn preceded by a flag	Coverage. If accounts leave unflagged, the signals are wrong.
False flag rate	Success managers stop reading a list that is wrong half the time.
Save rate after intervention	The business outcome, and the only argument for keeping it running.

Earning more rope

Assist	Weekly list for the success team with signals attached.
Approve	Auto-create a task on the owner with a suggested play. Right level.
Auto	Never auto-contact customers based on a churn model. That message is unrecoverable if it is wrong.

Explain, do not score. The success manager needs to open the conversation with something specific, and a probability is not something you can say to a customer.

Related: Follow up · Escalation watch · Anomaly explainer · all agent jobs · Agent Lab home

Free from AI School - no signup, everything runs in your browser.