Knowing who is leaving before they go.

Anything sold on a renewal has the same structural flaw: the renewal conversation happens at the renewal. By that point the decision was made weeks or months earlier, and the call is a formality that everyone involved can feel is a formality.

The signals were all there. They stopped opening the emails in the spring. They have not attended anything since February. The person who championed you internally changed jobs in March and the new one has never spoken to you. None of that was hidden. It was simply spread across four systems and nobody has a spare morning a week to cross-reference four hundred accounts.

Churn also has a visibility problem that acquisition does not. Nobody circulates an email celebrating the customer who did not leave, so nobody is ever staffed against it, and it stays everyone’s responsibility and therefore nobody’s.

The data exists and is already paid for. Email engagement sits in the marketing platform. Attendance, logins or usage sit in whatever runs the service. Support history sits in the ticketing system. The last real conversation sits in a CRM note somebody typed after a call.

The free text is where the strongest signals live, and it is exactly the part no rules engine can read. A ticket that mentions in passing that they are reviewing vendors this quarter outranks every behavioural metric on the list, and it will sit there unread forever.

It is also worth being honest about volume. A small organization does not have enough renewals a year to train a churn model anybody should believe. Anyone selling you one is selling you a number with a confidence interval wider than the decision it is meant to inform.

So this is deliberately not a prediction model. It is a signals system: a small set of explicit, explainable indicators, weighted, with the underlying evidence carried alongside every score. A number you cannot interrogate is a number nobody acts on, and a retention tool nobody acts on is a dashboard.

The model is used for the part it is genuinely better at than a rules engine, which is reading the free text, notes, tickets, replies, and drafting an opening message that reflects the specific reason this account is on the list.

The output is a short ranked list delivered on a Monday morning, not a dashboard. A dashboard is a thing you have to remember to open, and retention work loses that competition to everything urgent, every week, forever.

Every Monday the system pulls engagement, activity, support and CRM history for every account inside its renewal window, scores each one against the signal set, and delivers a ranked list of the fifteen most at risk with the evidence and a drafted first message attached to each.

Five decisions that make it get used

1. Every score is decomposed into the facts behind it. Not “risk 82” but “has not opened an email in ninety-four days, missed the last three events, primary contact’s title changed in March.” The person doing the outreach can then disagree with the ranking on the evidence, which is exactly what you want them doing.
2. The free text is read properly. A support ticket mentioning a vendor review, a reply that went unanswered, a note recording a complaint that was resolved but never followed up: these move an account up the list ahead of anything behavioural, and they are invisible to every rules-based approach.
3. The drafted outreach is written to the specific reason. What you say to somebody who has quietly stopped attending is not what you say to an account whose champion left, and a generic note asking how everything is going is worse than saying nothing, because it proves you have not been paying attention.
4. The list is capped at fifteen. This is a product decision, not a technical limit. A list of ninety accounts gets skimmed and abandoned; a list of fifteen gets worked through by Wednesday. Everything else waits for next week, and the cap is what makes the whole thing survive contact with a real week.
5. Outcomes are written back. Whether the account renewed, and what was actually said to them, is recorded against the original prediction. Within two renewal cycles the weights are grounded in what happens at this organization rather than what usually happens somewhere else.

The measure of success is not the accuracy of the ranking. It is whether somebody worked the list on Monday, three months in, without being asked.