AI Monitor explained: how OLM predicts churn 14 days early

The problem AI Monitor solves

By the time a member emails to cancel, they've already mentally checked out. The decision was made two to three weeks ago when their attendance started dropping. The cancellation email is a formality.

If you wait for the cancellation report, you find out after the fact. Your retention strategy becomes 'send a save-attempt email to the people who already canceled' — which works on roughly 5-10% of cases. AI Monitor is built around the observation that the actual signal happens 14 days earlier and is visible if you know where to look.

What the model actually does

Every member has an attendance baseline — typically 'classes per week over the trailing 90 days'. AI Monitor watches the rolling 14-day window against that baseline. When a member's recent attendance drops by more than ~50% relative to their baseline, they move into the at-risk bucket.

Critically, the model accounts for org-level patterns separately. A 30% July dip across the entire roster is treated as a context shift (snowbird season, summer break, weather event), not 50 individual churn signals. Per-member at-risk only fires when an individual deviates from their own baseline relative to the cohort.

The 14-day lead time is the median time between the at-risk flag firing and the cancellation email. About 70% of flagged members do churn within that window if no intervention happens. The other 30% are catches — members you reached out to and re-engaged.

False positives are the point

The model is tuned for false-positive tolerance, not false-negative. The cost of catching someone who wasn't actually quitting is one slightly awkward conversation. The cost of missing a quitter is a permanent revenue loss. The math heavily favors over-flagging.

Concretely: if a member misses a week because they were on vacation, they'll get flagged. The right response is a quick 'hey, missed you, hope you're well' message — not a guilt-trip retention pitch. The flag is a heads-up, not a verdict.

What to do with the alert

Each flagged member comes with a recommended action: a direct outreach via the messaging tool, a free private session offer, a class-time suggestion based on when they used to attend, or a billing-pause flag if their cancellation seems financially driven.

The actions are advisory, not automated. Most owners take the suggestion in 80% of cases and override it in the other 20% based on context the model doesn't have ('Jordan's mom just had surgery, give them space'; 'Sam mentioned getting laid off last month, offer comp'). The model doesn't know that; you do.

When AI Monitor doesn't help

Brand new gyms with under 50 members and three months of attendance history don't have enough data for the model to pick up patterns reliably. The system still runs but at-risk flags should be treated as low-confidence until your roster grows.

Highly seasonal academies (ski-town gyms, university gyms) need a few months of full-cycle data before the model accurately separates seasonal patterns from individual churn. Until then, expect more false positives during the obvious seasonal transitions.