How to Set Up an AI Chatbot for Your Gym: 7 Steps

Start with the 15 to 20 questions your front desk answers in a normal week. Sit with whoever works the desk for twenty minutes and write them down. The list is shorter than owners expect, because members ask the same things in the same order.

Sort what you get into six buckets:

• Price and membership: monthly cost, drop ins, trials, joining fee, student and off peak rates.
• Timetable and classes: what is on tonight, who teaches Saturday, when the next beginners block starts.
• Facilities: showers, parking, lockers, step free access, whether there is somewhere to leave a bike, and what to bring or wear to a first class.
• Policies: cancelling a class, cancelling a membership, freezing, bringing a guest, arriving late, refunds.
• Booking and access: how to book, how to get in, what to do when the app will not let you.
• Location and reaching a person: where you are and how to get there, reception hours, who handles billing, how to get a callback.

Only write down what stays true for at least three months. Anything that changes faster, like class times, is read live (step 3).

A short, well written set beats a folder of uploaded PDFs every time: a scoped answer set knows where it ends, while a document dump has no edges and fills the gaps by guessing.

Write each answer yourself, in the words your staff use at the desk, instead of accepting the defaults. Tone is not decoration. Research published in Frontiers in Psychology (2022) found, in two experiments in online retail, that a warm, socially oriented chatbot style raised customer satisfaction, mainly for customers who were more anxious about the exchange. A member cannot tell whether your knowledge base is well structured. They can tell in one line whether the reply sounds like your studio.

“Dear valued member” is wrong for a hard training CrossFit box, and “Hey, sick question!” is wrong for a calm yoga studio. Take one real question and answer it three ways to find your register:

• Reserved: “Thank you for your enquiry. Our membership options are set out below.”
• Warm and direct: “Good question. Here is how the memberships work.”
• Casual: “Sure thing, here is what we have got.”

Then write three lines your staff would never send, so the chatbot learns where your voice stops as well as where it starts.

Most boutique studios sit between the second and the third, and the choice is less important than sticking to it. Write five answers in that register and keep them as the reference the chatbot is calibrated against. Concrete examples teach a voice; instructions like “be friendly and professional” teach nothing. On WhatsApp keep replies shorter than on your website, but let them sound like the same person.

Two things to hold to in every answer: give the number or the time first and the explanation second, and never write a sentence your staff would not say out loud. What this reads like on the member’s side of the thread is worked through in why gym members hate chatbots.

• Read live from your booking system: the class schedule, remaining spots, opening hours over a bank holiday, current prices.
• Kept in the answers you wrote: anything stable, such as the freeze policy or where the door code lives.

You don’t build these connections yourself. A good AI receptionist is set up with you, connected to your booking system and to your own WhatsApp number, the one members already message. Your booking system stays the record for classes, memberships and payments; the chatbot reads from it and books into it. On WhatsApp it runs on WhatsApp’s official business connection, which needs a Meta business account.

A chatbot that can also book and move classes like this is what people call an AI agent. The AI agent for gyms page shows how it works.

Then set the rule that matters more than the connection: if the chatbot cannot look a number up, it does not produce one. An assistant that improvises a price is not a small bug. It is a member arriving on Saturday expecting a rate you do not offer, and a refund conversation you did not budget for.

Practical checks before you move on: open the chatbot and ask for tonight’s timetable, then move a class in your booking system and ask again. If the second answer still shows the old slot, the chatbot is reading a copy, and every price change you make from now on is a future complaint.

Hand the chat to a person when a member asks for one, after two failed answers, on a sensitive subject, or when the member is clearly frustrated. Members forgive not knowing. What they do not forgive is being circled back into the same menu with no way out. Those four triggers cover almost everything:

• Someone asks for a person. Any phrasing counts. The next message hands over, with no “are you sure” and no extra questions.
• Two failed attempts on the same question. The chatbot stops, says plainly that it cannot answer this one, and passes the thread on.
• A sensitive subject. Injuries, health conditions and complaints go straight to a person, whether or not the chatbot could have produced an answer. Cancellations and billing disputes follow your own rules: the chatbot handles them where your booking system and your policy allow it.
• Obvious frustration. Short, sharp or repeated messages are a signal to stop answering and start apologising.

Decide the other half too, because it is the half that gets skipped: where the handed over thread lands, who is responsible for it during opening hours and after, what the member is told about timing, and what travels with the conversation: the member’s name, their question and the chat so far. A member who has to repeat everything to a human has had a worse experience than if nothing had answered at all.

If you would rather your team send every message themselves, with the AI helping behind the scenes, that is a real setup too: AI agent vs AI assistant compares the two.

Say in the first message that the member is talking to an AI, and how to reach a person. Article 50 of the EU AI Act requires that people are told when they are interacting with an AI system, and it has applied since 2 August 2026. For a gym, that is a line in the first message: say the studio’s name, say the assistant is automated, say how to reach a person. Members read it as honesty rather than a disclaimer.

Three guardrails go alongside it:

• Topic boundaries. The chatbot answers within the set you wrote and declines everything else. No medical questions, no training advice, no opinions on other studios.
• A refusal that is allowed. “I do not have that one, let me get someone from the team” must be a permitted outcome, not a last resort after three guesses.
• A kill switch you have found before you need it. When a price is wrong or the timetable breaks, you want the chatbot off in one action, not a support ticket.

Personal data follows the same rules as everything else in your stack: keep the conversation where you can find it, keep it as long as you need it and no longer, and say so in your privacy notice.

Send the chatbot fifteen real member questions from your own phone before any member does. Use the channel members use. The first thirteen should be answered from the set you wrote; the last two should go to a person.

Score it honestly. Wrong answers are worse than missing ones, and a wrong price counts double. The last two are the real exam: both should reach a person, and the injury question gets a short, kind refusal, not advice. Run the same test again after any big change, such as new prices.

Half an hour a week keeps the answers up to date, because an AI chatbot drifts the moment the studio changes. Split it in two.

Twenty minutes reading the last week of conversations, looking only for four things: questions that got no answer, answers that were wrong, the same new question appearing more than twice, and any thread where a member got annoyed. Every one of those is either a missing entry or a missing handover trigger.

Ten minutes updating whatever moved. A price, a coach, a class time, a policy. Do it the day it changes, not the week after, because stale answers are the single most common way a chatbot puts wrong information in front of a member.

Once a quarter, take an hour to reread the whole set: remove what’s out of date and check your example replies still sound like you. Put all of it in one person’s calendar, or nobody does it.