AI for Swim Schools: Classes, Enrollment and Health Notes

An AI agent for a swim school receives enrollment requests, checks open spots by class, offers make-up lessons when a student misses a session, and records health alerts such as allergies or medical restrictions so the instructors know what to watch for in the pool. It works well when every piece of information has an owner, a access limit and a record of who checked what. The main risk is not the conversation itself but a leak: the same WhatsApp window should never show another parent the name, class or health condition of a child.

Key takeaways

• Spots by class should come from the registry the front desk controls, not from an answer the model assembles on its own.
• Make-up lessons are a written business rule: notice period, limits, level and equivalent class. The agent applies the rule and does not invent one.
• Health data is sensitive personal data. It should stay out of the ordinary chat, with restricted access and only the minimum necessary.
• Each guardian sees only their own child, and the system must enforce that separation, not just a line in the model's instructions.
• A pilot with one class, conversation logs and a metric agreed in advance shows whether it is worth expanding.

What the agent does and what the team still decides

It helps to separate three roles. The agent answers frequent questions, collects enrollment details, checks availability and routes cases. The team decides the exceptions: a student without an assessment who wants an advanced class, a complimentary lesson, a discount outside the price list, or a change of instructor for personal reasons. An agent that decides exceptions on its own creates promises the school cannot keep, and those promises turn into complaints.

It is also useful to be clear with leadership that the word superintelligence, used in the title because of search demand, does not describe what is being operated here. This is a software system with a language model, written rules, integrations and logs. If the team uses these words differently, the superintelligence glossary for executives can help align the vocabulary before a decision is made.

How the agent handles spots by class

Checking availability sounds simple, but it depends on data the school must keep organized. Each class needs a recorded capacity, age range, level, instructor, days and time. Without that, the agent ends up offering a time slot that is already full or a class that does not match the child's age.

In practice, the flow usually follows this logic:

• Capacity and level for each class are entered by the front desk and read at the moment of the question, not remembered from earlier conversations.
• When there is no space, the agent offers a waiting list entry with the date of registration, and does not promise a place.
• When a spot opens, the offer goes to the registered guardian, with a deadline to respond and a note that another interested family may take the spot.
• When a request fits no rule, the case goes to the coordinator with a short summary of the history, so the guardian does not have to repeat everything.

On the waiting list, the agent never shares the names of other families. It only says that the class is full and what the next step is.

Make-up lessons without confusion

Make-up lessons are where confusing messages most often appear, because every school has its own rules: minimum notice, number of make-ups per month, lessons that can be made up in another class and lessons that cannot. Before automating, the rule must be written down and approved by the coordinator. If it exists only in the head of whoever answers the phone, the agent will repeat that inconsistency at scale.

Illustrative scenario: a guardian messages on WhatsApp that the child will miss Tuesday's class because of a fever. The agent confirms which student is linked to that number, checks the make-up rule in the registry, offers two dates in compatible classes and records the choice. It does not comment on symptoms or suggest any medical action. If the notice arrives after the deadline, it explains the rule clearly and routes the case to the coordinator, who decides whether to make an exception.

Health data: collect little and store it elsewhere

The LGPD treats health data as sensitive personal data, and information about children and teenagers requires extra care. For a swim school, this translates into practical decisions that should be made before the pilot, not after the first incident.

The agent should not ask for medical reports or exams in the open chat. The guardian can be directed to send the document through the school's official channel, using its own form. The conversation keeps only an operational note, for example "student with diving restriction, see file", with access granted to instructors and the coordinator. The agent never interprets a diagnosis, never recommends treatment and never decides whether a child can enter the water.

It is also worth confirming with the vendor where this data is stored, who can access it and whether it is used to train models. These answers should be in writing before signing, because they change the school's compliance analysis.

How to stop one parent from seeing another family's data

The most common mistake is treating privacy as a text instruction, such as "do not reveal data about other students". That helps, but it is not access control. A model can be steered into revealing information it received in its context, so the rule has to live in the system.

In a sound architecture, the WhatsApp session is linked to a registered guardian. The query searches only the records tied to that link, and the tool returns only the permitted fields. The model never receives the full list of students in a class. The OWASP authorization guidance treats this kind of check as a server responsibility, enforced on every request, not as something the interface or the prompt can be trusted to handle.

In practice, this leads to a few simple rules:

• An unregistered number receives only general information and a path to verify through the front desk.
• Answers about a class show only the time and open spots, never the list of who is enrolled.
• Conversation logs record who asked and what was queried, but do not copy health content into debugging logs.
• Any request to see another student's data triggers a standard refusal and a record for review.

If the school does not yet have this separation, the agent should not be put in contact with families. Human service remains the path until the permissions have been tested.

Plan for a one-class pilot

Our method is to map the current process and its baseline, limit data and actions, pilot a real flow with records and an agreed metric, and then hand over and decide whether the operation expands, adjusts or stops. Applied to a swim school, the pilot covers one class and one type of request, not the whole school.

• Map the current flow: who receives the spot request, who approves make-ups and where the health file is kept today.
• Define in writing what the agent may look up and what it never answers, approved by the coordinator.
• Turn the make-up rules and the waiting list into approved text, with examples of cases that count as exceptions.
• Test with simulated conversations, including attempts to get another student's data and attempts to force an exception, before any real contact.
• Run with one class for an agreed period and review the cases sent to humans, the spot errors and the complaints.
• Decide using the metric set at the start: expand, adjust or stop.

If the school does not yet have an organized enrollment process, the approach in AI for driving schools: enrollment, lessons and reminders follows similar logic, with the same questions about notices and scheduling. For the channel, it is worth understanding the WhatsApp 24 hour customer service window, because it defines when the agent can start messages and when it must wait for the family to write first.

Frequently asked questions

Can the agent confirm enrollment on its own? It can confirm that a request was received and explain the next steps, but enrollment depends on documents, acceptance of the rules and validation by the front desk. Ideally, the agent collects the information, organizes the request and hands it to the human step with everything recorded. That way the family knows what is missing, and the school does not make commitments it has not checked.

Can I put the student's health file in WhatsApp? We do not recommend it. The file should sit in a system with restricted access, and the chat should receive only the operational information needed, such as a restriction the instructor must know about. Medical documents should arrive through the official channel, and the agent should not interpret them.

How does the agent know the person writing is the right guardian? It cannot know from what the person says alone, which is why identification must come from the registry. A phone number linked to a guardian allows specific queries, and any unknown number receives only general information. When there is doubt, the front desk verifies before releasing any data.

Where should a school start if it does not have written rules yet? Start with the mapping, not the tool. Bringing the coordinator together to write the rules for spots, make-ups and waiting lists is often the most useful work in the whole project, because it serves both the agent and the team. Only after that does it make sense to test a pilot with one class and conversation logs.

Sources

• OWASP: Authorization Cheat Sheet
• ANPD: Autoridade Nacional de Proteção de Dados
• Anthropic: Building effective agents