A8gent
All articles

Industry

Gym Attendance Drop-Off Automation: Catch the Fade

By the time a member cancels, the decision is weeks old. The behaviour that preceded it is sitting in your check-in data, unread, because nobody is paid to notice an absence. Here is how to build the system that notices - and the much harder question of what it should do next.

A8gentAutomation practiceAugust 5, 202613 min read
Contents
  1. The Cancellation Is the Last Event, Not the First
  2. Defining the Signal Without Building a Model
  3. Getting the Attendance Data Out
  4. The Hard Part: What the Message Should Actually Do
  5. A Worked Example, With the Arithmetic Shown
  6. Measuring It Honestly
  7. FAQ

The Cancellation Is the Last Event, Not the First

When a member cancels, a gym experiences it as a sudden event. It almost never is one.

The sequence usually runs: something changes in the member's life - a new job, an injury, a schedule shift, a training partner who stopped coming. Attendance thins. A week is missed. The habit, which was the actual product, breaks. Then a stretch of not going, during which the membership becomes a monthly reminder of a failure. Then, weeks or months later, an administrative act: a cancellation form, an email, a chargeback.

Everything a gym typically does about churn happens at the last step. Win-back campaigns fire on cancellation. Retention offers arrive at the moment of exit. Both are attempts to reverse a decision that was made a long time ago and has already been rationalised - and by then the member has usually told other people they quit, which makes reversing it a social cost as well as a financial one.

The data that would have flagged it is sitting in the check-in table the whole time. It is not hidden or hard to obtain. It simply has the same structural property as every other high-value automation problem in this series: the signal is an absence, and absences do not generate events. Nobody's screen lights up when a member does not come in. The same blindness produces lapsed patients in a vet practice - see our recall build - and unscheduled referrals in a clinic.

So the build is, again, a scheduled query that manufactures an event. That part is genuinely easy. The part this article spends most of its length on is the part that gets skipped: what you do with the flag. Because the wrong intervention is worse than none, and this is one of the few automations where that is literally true.

For where this sits among a studio's other candidates, see AI automation for gyms and fitness studios and AI agents for fitness studios and gyms. The other half of churn - members whose payment failed rather than whose motivation did - is a completely separate build covered in failed payment recovery.

Defining the Signal Without Building a Model

The temptation is to build a churn prediction model. Resist it for the first version, and possibly forever. A simple relative-change rule captures most of the available value, is explicable to the staff who have to act on it, and does not require you to have the volume of data that makes a model meaningful. A single studio does not have enough members to train anything worth training.

Use the member as their own baseline. This is the single most important design decision. An absolute rule - "flag anyone who has not visited in 14 days" - is wrong in both directions. It flags the member who has trained once a week for three years and is perfectly happy, generating a false alarm and an unnecessary message. And it takes two weeks to notice that a five-times-a-week member has stopped, by which point the habit is already broken.

A workable definition: compare each member's visits in a recent window against their own established rate over a longer prior window, and flag a material proportional decline. For a member averaging four visits a week, one visit in the last fortnight is a dramatic signal after seven days. For a member averaging three a month, the same absence means nothing.

Details that decide whether the list is usable:

  • Exclude new members from the relative rule. Someone six weeks in has no stable baseline, and their pattern is still forming. New members need a different, simpler onboarding sequence - and they are also the highest-risk population, so give them their own workflow rather than distorting this one.
  • Handle known absences. Holds, freezes, holidays, notified injuries. If a member told the front desk they are travelling for a month and the system texts them about missing sessions, you have demonstrated that nobody listens. Whatever field records that, read it, and if there is no such field, create one - it is cheaper than the alternative.
  • Seasonality is real and it is not churn. Christmas, summer, school holidays. A rule tuned in October will flag half your membership in late December. Either compare against the same member's behaviour in previous equivalent periods, or suppress the whole workflow during known lulls and accept the blind spot honestly.
  • Set the sensitivity by what you can act on. This is the constraint nobody plans for. If the flag produces 90 members a week and your team can have 15 meaningful conversations a week, you have built a list that gets ignored by the second week. Tune the threshold to your actual intervention capacity, not to catching everybody. A short list that gets worked beats a complete list that does not.

Only once that is running, and only if you have the volume, is there a case for something more sophisticated. Most studios never need it.

Getting the Attendance Data Out

You need check-ins or class registrations per member over time. What that takes depends on your platform.

Mindbody publishes a public developer portal. Its documented endpoint listing groups the API into Appointment, Class, Client, Sale, Site and Staff. Directly relevant to this build, the Client group includes "View Visit History" and the Class group includes "Class Registration," "Update Clients Visit," "Enrollment Registration" and "View Class Schedule." That is precisely the surface an attendance detector needs, and it is unusually good news compared with the gated APIs common in other industries covered in this series.

On authentication, Mindbody's consumer documentation states that "All calls to the Consumer API must use HTTPS connections, using TLS v1.2 or higher," that you must "pass an API-Key header in your request to authenticate your integration," and describes OAuth 2.0 with access tokens "in JSON Web Token (JWT) format," a Bearer token type, an expires_in lifetime in seconds, and a refresh_token returned "when offline_access scope included." It also advises using different API keys per integration and storing them "only in intermediary server-side applications, not in client-facing applications" - sensible advice that is worth following literally.

Two honest gaps. We could not find published rate limits on the Mindbody documentation pages we were able to read - the consumer documentation page contains no rate limit information at all. Third-party blogs quote a figure; we are not going to repeat an unverified number, because a rate limit is exactly the kind of fact that changes and that you should not design a nightly bulk sweep against on hearsay. Confirm it with Mindbody before building anything that pulls visit history for your whole membership in one pass, and build in backoff regardless. See also rate limit.

And we found no documented endpoints relating to contracts in that endpoint listing. If your intervention logic wants to know a member's contract term, minimum period or notice requirements - which it should, because those change what you can sensibly offer - do not assume the API will tell you. Check.

If your platform has no usable API, a scheduled export of check-in data into a spreadsheet is entirely sufficient for this workflow. The analysis is a comparison of two date windows. You do not need real-time anything - the signal you are detecting took two weeks to form, so a weekly batch is appropriate and a nightly one is generous. Do not let integration ambition delay a build whose natural cadence is weekly. See polling vs webhook for why webhooks are the wrong instinct here.

The Hard Part: What the Message Should Actually Do

Here is the trap, and it is the reason most attendance automations quietly get switched off.

A member who has stopped coming knows they have stopped coming. They feel some combination of guilt, embarrassment and irritation about it. An automated message that says "we've missed you at the gym!" does three things simultaneously: it confirms that their absence was noticed, it reminds them they are paying for something they are not using, and it reveals that the noticing was done by software. For a member on the fence, that message can be the prompt to cancel. You have built a churn accelerator with a friendly tone.

This is not an argument against acting. It is an argument for these constraints:

1. Do not mention the absence as a fact about them. The difference between "we noticed you haven't been in for two weeks" and an invitation to something specific is the difference between an accusation and an offer. The second one works and costs the same.

2. Offer something concrete, not encouragement. Encouragement is free and therefore worthless. A specific class at a specific time that fits the pattern they used to train in, a slot with a coach, a re-assessment, a new block starting Monday - these are actionable. "Come back soon!" is not.

3. Prefer a human for high-value or long-tenured members. A three-year member deserves a conversation, and the conversation will tell you things no dataset contains - an injury, a shift change, a bereavement, a coach they did not get on with. Route these to a named staff member as a task with the attendance context attached, and let the automation handle only the detection. This is the single highest-return version of the whole build and it involves almost no automation at all.

4. Ask a question rather than making a statement. "Is there a time that would work better for you at the moment?" invites a reply that gives you something to work with. It also gracefully absorbs the case where your detection was simply wrong.

5. Fix the cause, not the symptom, where the data shows one. If flagged members disproportionately used to attend a class that got moved or a coach who left, the automation has found an operational problem, and messaging individuals is treating a symptom. This is the most valuable output of the whole system and it only appears if you look at the flags in aggregate. Review the list monthly as a group, not just as individuals.

6. Cap it hard. One touch, then a gap of weeks. A member being chased about their attendance repeatedly is being nagged, and nagging converts ambivalence into decision. Whatever your capacity, the answer is fewer, better contacts.

On channel: if you send SMS over a US 10-digit long code, A2P 10DLC applies. Twilio describes it as "the standard that United States telecom carriers have put in place to ensure that SMS traffic to US end users through long code phone numbers is verified and consensual," requiring a registered Brand and Campaign, and notes unregistered senders "will receive additional carrier fees for sending unregistered traffic." See our A2P 10DLC guide, the glossary entry, opt-in and STOP handling - and note that an opt-out here should suppress the member from every campaign, not just this one.

A Worked Example, With the Arithmetic Shown

Illustrative. We report no client results, and this is the build in this series where the honest answer depends most heavily on your own numbers.

Baseline. A studio has 540 members at $110/month. Monthly cancellations run at 4 percent: 540 x 0.04 = 21.6 members a month, of which suppose 70 percent are voluntary rather than payment failures: 21.6 x 0.7 = 15.1 voluntary cancellations a month.

How many are detectable in advance? Not all. Some cancel for reasons no attendance pattern predicts - a house move, a job relocation, a pregnancy, a redundancy. Suppose 60 percent show a meaningful attendance decline first: 15.1 x 0.6 = 9.1 a month are theoretically catchable. This ratio is measurable retrospectively and you should measure it before building. Pull last quarter's cancellations, look at each one's attendance in the eight weeks prior, and count how many showed a clear decline. That single afternoon of work is worth more than any vendor's projection.

How many can you actually save? This is the number nobody can give you honestly. A member whose life circumstances changed is not going to be saved by a text. Suppose an intervention that reaches them early - and for long-tenured members, a phone call - saves one in five of the detectable group: 9.1 x 0.2 = 1.8 members a month.

Value. If the average remaining lifetime of a saved member is 8 months, each save is worth 8 x $110 = $880. Monthly: 1.8 x $880 = $1,584. Annually: about $19,000.

Against cost. Build at $3,000 plus $50/month gives a first-year cost of $3,600 - your quotes; see automation build cost. Break-even is $3,600 / $880 = 4.1 saved memberships over the year, against a modelled 21.6. Comfortable, with room for the estimate to be substantially wrong.

But note where the sensitivity is. The whole case rests on that "one in five" save rate, which is the one input we cannot verify and you cannot know in advance. If it is one in twenty, the annual value drops to about $4,750 and the case is marginal. This is precisely why the next section insists on a control group: you are not measuring whether members came back, you are measuring whether more of them came back than would have anyway, and lapsed gym members have a well-known tendency to drift back on their own in January regardless of what you send them.

Two additional cautions. Do not count a save that was going to happen anyway - that is what the control is for. And do not discount to save a member unless you have decided that policy deliberately: a retention discount offered by automation will spread by word of mouth, and you will end up offering it to people who were never leaving. Our ROI of AI agents piece covers this class of error; for budget ranges see AI automation cost.

Measuring It Honestly

Hold back a control group. Randomly exclude 20 percent of flagged members from any intervention, and compare cancellation rates over the following 90 days. It is the only way to know whether the system works. It costs you a fifth of the upside for a quarter, and without it you will be arguing from anecdote in six months - and quite possibly defending a system that does nothing.

Unlike the clinical settings elsewhere in this series, there is no ethical obstacle here: not sending a marketing message to a gym member is not withholding care. Take the free experiment.

Report four numbers. Cancellation rate in the flagged-and-contacted group versus the flagged-and-held-back group. The proportion of cancellations that were flagged in advance - your detector's recall, and the number that tells you whether the rule is tuned right. Reply rate to the intervention, as a proxy for whether the message is landing as an offer or an accusation. And opt-out rate, which is your early warning that you are nagging.

The failure modes, in order of how much damage they do:

  • Messaging members on a documented hold or holiday. Demonstrates that the studio does not listen, which is the exact opposite of the intended message. Check the hold field before every send, and treat a missing field as a reason to build one.
  • Accidentally accelerating churn by reminding ambivalent members of unused spend. Detected only by the control group. If the contacted group cancels at a higher rate than the held-back group, switch it off and rewrite the copy - and be willing to believe that result rather than explaining it away.
  • A list too long to work, which decays into a list nobody opens. Tune to capacity.
  • Seasonal false positives flooding the list every December. Suppress or seasonally adjust.
  • The silent stop. The scheduled job fails, the list is empty, and an empty list looks like a healthy membership. Alert on zero output, not only on errors - the same trap as every absence-driven automation. See observability and error handling.
  • Double-messaging from a re-run. Key each contact on member ID plus flag date - see idempotency - and suppress anyone contacted by any other campaign that week.

One last piece of advice that is not about software. The most effective version of this system in a small studio is a weekly list of fifteen names handed to a coach who knows them, with a note on what changed. That is a report, not an automation, and it will outperform any messaging sequence. Build the report first. Add the messaging only for the segment your team cannot reach personally, and be honest that it is the weaker instrument.

If you want a second opinion on whether this or payment recovery is your higher-value build, the free automation audit works through it, and automation services explains how we scope it. General retention thinking is in AI agents for customer retention.

FAQ

Why not just trigger a win-back campaign when someone cancels?
Because by then the decision is weeks old, has been rationalised, and has often been told to other people - which makes reversing it a social cost as well as a financial one. The behaviour that preceded the cancellation was visible in your check-in data for a month beforehand. Acting at the point of cancellation means competing with a settled decision; acting on the attendance decline means intervening while the member is still ambivalent.
Should I flag anyone who has not attended in 14 days?
No - use each member as their own baseline. An absolute threshold is wrong in both directions: it raises a false alarm on the member who has trained once a week happily for three years, and it takes a fortnight to notice that a four-times-a-week member has stopped, by which point the habit is already broken. Compare recent visits against that member's own established rate and flag a material proportional decline instead.
Can I get attendance data out of Mindbody?
Yes. Mindbody's public developer portal documents endpoint groups for Appointment, Class, Client, Sale, Site and Staff, including "View Visit History" under Client and "Class Registration," "Update Clients Visit" and "View Class Schedule" under Class. Authentication uses an API-Key header over HTTPS with "TLS v1.2 or higher," plus OAuth 2.0 with JWT access tokens and a refresh token when the offline_access scope is included. We found no published rate limits on the pages we could read, so confirm those before building a bulk nightly sweep.
Will messaging lapsing members make them cancel?
It can, and this is the single biggest risk in the build. A message that names the absence reminds an ambivalent member that they are paying for something unused, and reveals that the noticing was automated. Avoid stating the absence as a fact about them, offer something specific rather than encouragement, and hold back a control group so you can actually detect it if you are making things worse. If the contacted group cancels at a higher rate than the control, believe the data and rewrite.
Do I need a churn prediction model?
Almost certainly not, and a single studio rarely has enough data for one to be meaningful. A relative-change rule against each member's own baseline captures most of the available value, can be explained to the staff who have to act on it, and can be tuned in an afternoon. A model that no one on the front desk understands produces a list no one on the front desk trusts, which is worse than a cruder list they will actually work.
How many members should the system flag each week?
As many as your team can have real conversations with, and no more. This is a capacity constraint, not a detection one. A list of ninety names handed to a team that can manage fifteen conversations gets ignored by the second week, and then the whole system is dead. Set the sensitivity from your intervention capacity and accept that you will miss some - a short list that gets worked beats a complete list that does not.
How do I prove it worked?
Randomly hold back around 20 percent of flagged members from any contact and compare 90-day cancellation rates against the contacted group. Lapsed members drift back on their own, particularly in January, so without a control you will credit the automation for behaviour that would have happened anyway. There is no ethical barrier here - declining to send a marketing message is not withholding anything - so take the free experiment.
What is the most valuable output of this system?
Often not the individual messages at all, but the aggregate pattern. If flagged members disproportionately used to attend a class that moved time or a coach who left, you have found an operational cause and messaging individuals is treating a symptom. Review the flagged list monthly as a group as well as person by person - that review costs nothing and regularly turns up a fix worth more than the entire campaign.

If this resonated

Module 02 is the long version of this post.

Eval set, refusal pattern, kill switch. Build it in a weekend.

All postsLast verified 2026-08-05

Done for you

Want this built for your business?

Tell us the workflow. We scope, build, and ship the agent with guardrails and the numbers to prove it worked. The scoping call is free.

Free scoping call. You own the code.

Request a scoping call