A client who quits rarely announces it. They drift first: a check-in a day late, then a few days late, answers getting shorter, a session or two quietly skipped. By the time they cancel, the decision was made weeks earlier. The signs were in their data before they reached your inbox, which means you can catch them while there is still something to do about it.
What drifting looks like in the numbers
Four things move before someone leaves. The gap between check-ins widens. Their written answers shrink from a paragraph to a line. Their logged sessions fall below what you programmed. And their own ratings slip below what is normal for them. None of these trips a standard overdue flag on its own, which is exactly why they get missed until it is too late.
Baselines beat thresholds
A fixed rule like “flag anyone under 6 out of 10” is the wrong tool, because a 6 is a bad week for one client and a good one for another. Compare each client to their own recent history instead. A chatty client dropping to one-line answers matters; a reserved client who always writes one line does not. PTToolBoxscores drift against a client's own six weeks and shows the reason next to the name, so the list is the people who have actually changed.
Act while they are still training
The window that matters is before the overdue flag, while they are still showing up. A message that lands then reads as attention. The same message after they have gone quiet for a fortnight reads as chasing. Same words, different week, opposite effect.
What to actually send
Keep it specific and small. Name the thing you noticed, ask an open question, and offer to change one thing. “I saw the last two sessions got cut short, is the current split still fitting your week?” does more than a generic “how's it going”, because it shows you are actually watching. That is the whole point of catching the drift early: you get to coach the problem instead of processing the cancellation.