Churn announces itself early

Almost nobody churns abruptly. Usage thins, orders stretch out, logins drift from weekly to monthly, the invoice gets paid three days later than usual, and then, weeks after the real decision was made, the cancellation makes it official. Which means churn is one of the few business problems where the data warns you in advance, in signals you already collect, and most companies read the warning only in the exit interview.

This piece is about building the early-warning view: which signals actually lead churn, how to combine them into a score simple enough to maintain, what to do when the score fires, and, because this is us, how to measure whether your saves are real or whether you are sending discounts to people who were staying anyway.

The quiet weeks: what decay looks like

Plot the activity of customers who eventually cancelled, backwards from their cancellation date, and a consistent picture appears across business models. There is a long stable period, then an inflection, then a slope that runs for weeks, and only at the very end, the event your dashboard records. The inflection is the moment worth owning. At the cancellation screen, you are negotiating with someone who already decided, and your tools are discounts and friction. At the inflection, weeks earlier, the customer is not leaving yet, they are drifting, and drift responds to much cheaper interventions: a check-in, a nudge back to the feature or product that hooked them, a fix for the silent problem that started the slide.

The reason most teams miss the bend is not inattention, it is framing. Their metrics compare customers to other customers, and against the population, a decaying customer can still look average for weeks. The signal lives in a different comparison: each customer against their own past. A customer who ordered every three weeks for a year and has now gone six weeks quiet is screaming, even if six weeks is normal for someone else. Baseline-relative measurement is the single design decision that makes everything below work.

The signal is not "this customer is below average." It is "this customer is below themselves."

The signals worth watching

Different businesses see different tells, but the reliable ones cluster into five families, and all five live in systems you already run. None of this requires new tracking; it requires reading what the billing system, the CRM, the order history, and the support inbox already know, together instead of separately.

Signal family	What to measure	Where it lives
Frequency decay	Gap since last order or login vs the customer's own typical gap	Orders, product analytics
Depth decay	Narrowing basket, fewer features touched, shorter sessions	Orders, product analytics
Support friction	Unresolved tickets, sentiment turn, silence after a complaint	Helpdesk, inbox
Payment friction	Late invoices, failed cards not retried, downgrade inquiries	Billing
Relationship decay	Champion left, emails unopened after months of engagement	CRM, email tool

Two of these deserve a note. Support friction is the most misread: a complaint is not by itself a churn signal, engaged customers complain because they care. The dangerous pattern is a complaint followed by silence, the customer who stopped arguing. And payment friction is the most underrated, because it is often the only signal for otherwise quiet customers: the card that fails and never gets updated is frequently a decision that has already been made, wearing an administrative costume.

A score you can build this quarter

You do not need machine learning to act on these signals, and for most small and mid-size businesses we would argue you should not start there. A points-based score, exactly the pattern we laid out for scoring leads with the data you already have, applied to the other end of the customer life, catches the large majority of the value, and everyone can see why a customer was flagged, which matters when a human is deciding whether to call them.

The build: pick one signal from each family you have clean data for, define a threshold relative to the customer's own baseline, assign points weighted by how strongly each signal predicted churn in your last year of history, and sum. Then, and this is the step that turns arithmetic into a tool, check the score against reality before trusting it: of the customers who churned last quarter, what share would this score have flagged four weeks ahead? Of the customers it would have flagged, how many actually churned? Those two numbers, coverage and precision, tell you whether to tighten or loosen the thresholds. Expect to iterate twice.

Pricing the save: your arithmetic, not folklore

Retention pitches lean on a famous statistic: acquiring a customer costs five to twenty-five times more than keeping one. We will be straight with you, that figure is folklore, endlessly cited and only loosely sourced, and we do not build cases on it. The good news is you do not need it. The comparison is computable from your own books in ten minutes, and your version is defensible in any meeting.

Notice the last line of that box, because it is the part the folklore misses: the value difference is not just cost. A saved customer sits on the flat part of the retention curve we drew in the cohort piece, while a new customer has the whole early cliff still ahead of them. Priced by what they contribute over the following year, per the method in LTV you can actually defend, the gap is usually wider than the cost gap.

Measure saves like a skeptic

Here is the failure mode that ruins most retention programs, and it is a measurement failure, not an execution one. The team builds the score, launches outreach to everyone flagged, and reports that 70 percent of contacted customers stayed. Success? Unknowable, as reported. Some of those customers were always going to stay; flags are warnings, not verdicts. Crediting the campaign with every retained customer it touched is the retention version of the ad platform grading its own homework, the exact pattern we take apart in incrementality over attribution.

The fix costs nothing but discipline: hold out a random slice of flagged customers, 10 to 20 percent, who receive no intervention, and compare retention between the groups after a quarter. If contacted customers stay at 70 percent and held-out ones at 62, your program is worth eight points, and you can put a dollar figure on that and decide if the outreach effort earns it. If the holdout also stays at 69, the program is theater, and the honest move is to redesign the intervention, not the report. Either answer is worth having. Only one of them is available without the holdout.

What to actually do when the score fires

A word on the intervention itself, because scoring without a playbook just produces better-documented churn. Three rules from the field. First, match the response to the signal: frequency decay wants a reason to return, an unresolved complaint wants resolution before any marketing touches the account, payment friction wants a human making the card update effortless. Second, lead with service, not discount; a concession offered to a drifting customer teaches them that drifting is how you get concessions, and it spends margin on people who wanted a fix, not a coupon. Third, sequence by value: work the list in order of customer worth, because attention is the scarce resource and persistence pays here just as it does with leads, most saves take more than one touch.

And keep the loop honest quarter over quarter: re-run the back-test, re-tune the weights, keep the holdout running permanently. Churn patterns drift as your mix and market drift. The companies that keep customers longest are not running cleverer models. They are the ones who noticed that leaving takes weeks, decided those weeks belonged to them, and then measured, with a straight face, whether their reclaiming of those weeks actually worked.