A demonstration on sample data
Nobody cancels. They just stop.
A customer who leaves a small business rarely says so. The orders get smaller, then they get further apart, then there are none. By the time it shows up in the monthly numbers it has been true for a season.
Run it on
Most people guess low.
The calls worth making this week.
| # | Who | What happened | Why we say so | Revenue at risk / yr | Margin at risk / yr | Orders | Done this before? |
|---|
Each bar is one order, its height the amount. The grey block is the window their next one was due in, taken from their own spacing.
The arithmetic
What it read
| Date | Amount |
|---|
Run it on your own book.
Export your invoices or orders from whatever you already use. It needs three columns: who the customer is, the date, and the amount. Any spelling of those headings is fine.
A QuickBooks report comes with the company name in the first row, subtotals between every customer, a grand total at the bottom, and the customer's name written once above a block of transactions that all have that cell empty. This reads that shape and says what it set aside.
The file never leaves this page. There is no upload and no server call. The reading happens in your browser and disappears when you close the tab. You can check that by turning off your network connection and doing it anyway.
or drop it here
How often it is wrong.
This page can tell you how often it is wrong, because the sample book knows something the engine does not. Every customer in it was built with an intention, and the engine never sees those labels. Scoring one against the other is the whole of the table beside this.
A sales manager who works a list and finds the first two names are fine does not call again. So the number that matters is not how many it catches. It is what share of the calls it asks you to make are a waste of your morning.
On your own book there are no labels and no way to compute this. What carries across is the shape: lateness is close to certain, and a shrinking spend measured over a handful of orders is a lead, not a finding.
How it decides.
Four steps. All of them are arithmetic you could check by hand, which is the point.
- 1Measure each customer against themselves. Take the gaps between their own orders and use the median. A customer who orders every ninety days is not late at forty; one who orders weekly is in trouble at three. A single company-wide rule gets both wrong, and most reports use one.
- 2Score the silence. Past a customer's own normal gap, the chance they have stopped grows exponentially, with a scale of one and a half gaps. At their normal interval it is zero. At twice, 49 percent. At four times, 86 percent. The formula is 1 − e^(−excess ÷ 1.5 × gap), and it is printed here because a score nobody can check is a score nobody should act on.
- 3Catch the fade separately. Some customers never go quiet. They order the same week they always did, for less, and no gap ever appears. A trailing window is compared with the window before it, both annualised, so a change in order size and a change in frequency land in the same number. The window is scaled to the customer's own spacing, and a fade is not called unless the earlier window held at least four orders, the recent one at least two, and the drop is larger than that customer's own order-to-order variation. A flat 25 percent cut, which is what this used to use, sits inside the noise on most books and mostly catches noise. Below the floor there is not enough to see, and those customers are listed as not judged rather than told they are on pace.
- 4Rank on three things you can check, and no probability. Anyone still inside their own rhythm goes to the bottom. Above them, the customers who have never been this quiet before come first, because a gap a customer has set twice already is a habit and a gap they have never set is news. Within each of those groups the order is the margin behind the account, largest first. An earlier version of this page multiplied the money by a probability that the customer would still answer — e−(overdue − 1.5) ÷ 1.7, and a flat 0.82 for one still ordering. Those constants were invented. Nothing in a sales ledger measures them, and a number nobody can check should not be the number that decides the order of the list, so they are gone rather than disclosed. What replaced them is worse at sounding precise and better at being true.
Where this does not work.
A demonstration that hides its limits is selling something. These are real and worth knowing before anybody builds on it.
A remodeler has no churn. Customers who buy once and are done have no rhythm to break, so there is nothing here to measure.
No rhythm on file. Those customers are listed as unjudged rather than given a score that would be invented.
A pool company's customers all go quiet in October. Seasonality has to be modelled before any of this means anything, and this page does not model it.
Lateness is close to certain: a customer who has not ordered in four of their own gaps has stopped. A shrinking spend is a much weaker signal, measured over a handful of orders, and a real share of those calls will find a customer who is fine. The measured rate on this book is in the section above.
An equipment rental yard's customers come back when the next job starts. A long gap there can mean finished, not lost, and nothing in the order record distinguishes the two.
The larger end of distribution and residential home services are covered by tools that already ship this, and they reach further down-market than is comfortable. The gap is everyone below and beside them.
This took a day.
It is a demonstration, not a product. The version that matters is the one fitted to how one business actually works: its own definition of a customer, its own seasons, the export it already has, and the place its staff would actually look.
If you want to know what this would find in your numbers, send me the question.
Caspase LLC · Virginia · caspase.ai