I Spent a Year Putting AI Into My Gym. Here's What Happened

Evan Vaughan

I Spent a Year Putting AI Into My Gym. Here's What Happened.

It's 2023. 7 PM on a Saturday. I'm toast.

I just scheduled next week's email blast. Another 60-hour week running my F45 studio. And it still didn't feel like enough.

Go home. Eat. Sleep. Wake up. Do it again.

But hey - this is the dream, right?

My background is in software and AI. I understood the tech. I just couldn't see where it fit in a business built on human connection.

So I made myself a deal: start small, and never let a robot ruin the thing that makes my studio mine.

One year later, here's the report.

Before I changed anything, I counted.

For two weeks I wrote down every task and every hour. It was ugly.

A normal week:

  • 12 hours on sales messages - texts, emails, calls, lead and trial nurture (ouch)

  • 2 hours pulling MindBody reports

  • 2 hours building offers and follow-up sequences

  • 2 hours talking to my team

  • 2 hours on retention outreach

  • 1 hour posting on social

  • 1 hour on terminations and account cleanup

  • 1 hour fixing schedule conflicts

  • 1 hour organizing events

A huge chunk of my week. Gone. On stuff that drained me.

So I asked one question for every task: Does this actually need ME?

Most of it didn't.

The goal: claw back 25% of my week. No new full-time hires. That time goes back to my coaches and my community.

My one rule

Before I built anything, I set a rule I never broke.

AI never presses "send" by itself.

I've been pitched by too many companies promising AI that runs your whole back office. It can. Just... badly. A machine doesn't care about my community the way I do.

So in every system I built, AI does the heavy lifting - watching, drafting, summarizing. Then it stops and waits for me to click approve.

AI drafts. I decide.

That one rule is why none of this ever felt fake.

Oh - and the tools. I live in two: ChatGPT and Claude. Both are free to start. Both have a paid plan around $20 a month. The third tool, RepFlow, is a custom one that delivers everything you read here into one system.

ChatGPT does more flashy stuff - images, video, voice. Claude writes more like me and chews through long docs and code. I bounce between them. Pick one and get good.

I started where it was easy

First target: social media.

Same message, four places - email, Instagram, Facebook, printed flyers.

So I made my email the master copy. Attached an image. Then let AI spin out the rest - captions and images shaped for each platform. I plugged into the Instagram and Facebook APIs so I wasn't copy-pasting all day.

An hour became thirty minutes. And the captions got better.

One thing clicked here that mattered everywhere else: I fed the AI my old emails and texts until it sounded like me. After that, it stopped sounding like a robot.

The same idea fixed my reports.

I hated pulling MindBody numbers. Attendance, late cancels, rosters. Two hours, every week.

First I just had AI summarize the export down to the few numbers I cared about.

Then I leveled up with something called MCP. Think of it as a plug that lets your AI talk straight to another app's data. Instead of copy-pasting a report into a chat, the AI reaches in and grabs it.

I found one online, checked it was safe, and connected it. (Here's how to do it in Claude and in ChatGPT.)

Now I just ask: "How did attendance trend last week? Which classes are slipping?" And it answers - from live numbers.

That's the jump from reading reports to talking to them.

Then I went after the tiny stuff. Terminations. Reconciliations. Account cleanup.

I killed most of it with simple triggers in Google Apps Script. Someone submits a form, a little workflow fires on its own - a text confirmation here, a summary to my inbox there.

Nothing fancy. But each trigger was one less interruption. Forever.

The surprise was on the floor

Here's the win I didn't see coming.

My sales team and my coaches lived in two different worlds.

A coach would have a great talk with a trialer after class. Then it vanished. It never made it back to whoever ran follow-up.

So I closed the loop.

Before class, coaches get a snapshot: who's coming, who's close, what to say. After class, they send a short report - conversations, issues, notes, what they finished.

Back then it was two Google Forms. I called them the pre-game and the post-game. (Today it's grown into an app we call RepFlow iOS. The duct-tape version already worked.)

On-floor sales jumped 4-6x.

And after a week of reports, AI rolled it all into one list - restock this, follow up with her, call him back. Patterns I would've missed.

Best part? I closed that loop without texting my coaches all day.

That win made me braver.

So I recorded and transcribed our sales calls, then had AI score them. What worked. Where we lost them. What to do next time.

Hiring got easier too. I taught the AI what a good call sounds like, and new hires practiced against it before touching a real lead.

They walked into call one having already done ten.

Scheduling came next. Not really an AI problem, but it's a universal headache, so here's my two cents.

I started with a spreadsheet of coach availability. Then I had AI build next month's schedule around it.

It worked so well that we turned it into a real web app - coaches got it on their phones with alerts.

That's the part that surprised me most: I could build real software now. A landing page, a form, a small app - for almost nothing. I used Vercel, which takes a little tech comfort. No-code tools like Lovable need zero.

Goodbye "I never saw the schedule."

Then I went after the monster

The biggest line on my audit was sales messages. Twelve hours a week. That's the one I really wanted back.

So I rebuilt it.

Static automations drove me nuts. A lead tells me they just got hurt. Next morning, a robot text fires: "So excited to meet you - come in tomorrow?!"

Slimy. Tone-deaf. And people are numb to that stuff anyway.

So I built dynamic campaigns. I tell the AI what to say and how to say it - with one rule: follow the actual conversation. If someone's hurt, the next message meets them there.

Still nurturing. Just human.

Then I turned to my old client lists - a goldmine I was sitting on.

I wired campaigns to an AI agent. Now I just say:

"Build an audience of every lead and expired trialer from March. Make a win-back campaign offering a free session."

Seconds later, they're all in a custom drip - friendly texts and emails, in my voice. Nothing goes live until I say go.

And then the big one. Maybe my favorite.

Using Twilio, I built an AI assistant that answers the phone when I can't.

New lead comes in? It calls ME first. If I pick up, it connects me while they're hot. If I miss it, it reaches out to the lead for me and tries to book them on the live offer (like 10 days for $29).

Inbound call? It rings me and my team first. No one grabs it in 40 seconds? It rolls to the assistant, who takes the call and sends me a transcript.

We don't miss calls anymore.

(If you try this, learn the rules on consent and opt-outs. When a machine dials people, that matters.)

Last piece: when I knew I'd be off my desk, I flipped on an auto-responder. In my tone.

The secret was context. I fed it everything - who the person is, their attendance, what they bought, where they're at. Then it drafts a reply that actually fits.

But it always asks me to approve first. I get a ping, glance, send in seconds.

If you ever find a tool that skips that step? Walk away.

Where I wasted time

Not everything was a win. Some tools looked amazing and did nothing. Here's where I got burned.

"Our AI does it ALL."

Be careful with anyone promising AI that handles 100% of your calls and texts. That's how you lose your personal touch.

AI is easy to spot. And even easier to break. "Just test it, it solves everything" is a trap. Flashy demos are easy. Real growth is hard.

I fell for it. I tried an app called Beside - Y-Combinator backed, great reviews, cool idea.

But it missed the point. The transcripts were cringey. It booked people on Google Calendar, which never turned into MindBody trials. And I could never tell what the AI did versus what I did.

No time saved. No extra trials. I'd skip it.

Tools that can't learn you.

If a tool can't learn from you, it's built around someone else's idea of how to run a gym. But you opened yours to build YOUR community. Not theirs.

I tried an AI sales helpdesk. I'm picky - cadence, wording, context. It kept adding weird punctuation and phrases that screamed "a robot wrote this." I tweaked prompts for hours. Still fake.

Gone.

Automating what I couldn't do by hand.

This is the big one. There's no bigger time-sink than automating a process you've never run yourself.

Early on, I tried to automate a referral program. Problem: I'd never run one by hand. I didn't know what to ask members, when, or what made them say yes.

So the AI built a referral flow based on... nothing. It flopped.

Later I ran a few referral asks myself. I saw what worked. Then I automated that. And it worked.

Do it by hand first. Then hand it off.

What a year taught me

AI is powerful. And it should never run alone.

Every system I built keeps me in the chair. AI does the boring, heavy lifting. I make the calls. That's what protected the human touch this whole business runs on.

A year in, I got my time back - about the chunk I set out to reclaim. I poured it straight into my coaches, my events, and the parts of my studio that actually need a person.

That's the whole game.

I'm going way deeper in my upcoming webinar - the exact setups, the flops, and where I think this is all headed. [Add your webinar date and signup link here.]