AI for Supplement Stores: What the Agent Can Say and When to Stop

An AI agent in a supplement store can answer with confidence about declared composition, the usage directions printed on the label, flavor, packaging, stock, delivery time and exchanges. It should not recommend a product for a health problem, suggest a dose outside what is approved, or claim that a supplement treats, cures or prevents any disease. The working rule is simple: the agent repeats only what is in the approved product sheet, and it stops the moment the conversation stops being about the product and starts being about the customer's body.

Key takeaways

• The agent states only what is written in the approved product sheet, with a source the staff can see, and does not fill gaps with general knowledge.
• Questions about symptoms, diseases, medication, pregnancy, age and health conditions are stop triggers, not sales opportunities.
• Health data that a customer shares must be handled carefully under Brazil's LGPD, so the agent should ask only for the minimum needed to answer the question.
• A handoff to a staff member must carry the conversation history, the product mentioned and the reason for stopping, so nobody has to ask the customer everything again.
• Before the agent goes live, the store must test its refusals with difficult questions, because that is where customer service usually fails.

Where information ends and health advice begins

In supplement retail, the line between informing and recommending is thinner than it looks. Saying that a product contains a certain amount of a nutrient is label information. Saying that it "helps you sleep better" or "is good for people with anxiety" is a claim about an effect on health. Language models do not know on their own what your product's regulation allows you to say. They complete patterns, and the most common pattern in supplement marketing copy is exactly the promise.

For that reason, the boundary has to be written down by the store, with examples, rather than left to the model's judgment. The most useful reference for your team is the product label and its registration: whatever the label states as purpose and directions for use is the limit of what the agent may repeat. Anything beyond that should be treated as out of scope. For specific regulatory questions, such as which claims are permitted for dietary supplements, check with Anvisa and a qualified professional, because these rules change and should not be summarized from memory in a prompt.

What the agent can safely tell customers

A well-configured agent can handle most of the question volume without real risk. This includes the ingredient list and nutrition table as printed on the package, the usage directions given by the manufacturer, the difference between two presentations of the same product, flavor, container size and expiry date, stock availability, delivery time and the exchange policy. It also makes sense to let the agent explain label terms, such as what a unit of measure means or what a serving is, as long as the explanation does not become a recommendation.

Beyond that, the agent can help customers find their way around the store: locate a product by name, compare two options using what the product sheets say, and route the customer to a staff member when the question is of a different kind. This part of the service usually creates fewer problems and saves the team a good amount of time.

The triggers that force the agent to stop

The most important part of the design is the trigger list. It should be short, concrete, and reviewed by whoever is responsible for operations and by the store's compliance function. When a trigger appears in the conversation, the agent does not try to solve it. It acknowledges the question, makes clear that it cannot give health guidance, and hands the conversation to a person within the service window you have agreed on.

• Mention of a symptom, disease, pain, test result or diagnosis, whether from the customer or from someone else.
• Use of a prescription medication or an ongoing medical treatment, especially with a question about interactions.
• Pregnancy, breastfeeding, very young or advanced age, when the customer asks for a recommendation.
• A request for a dose, frequency or combination of products for a health goal.
• A question about replacing medicine, a prescribed diet, or professional follow-up.
• A request to compare products by their effect on the body rather than their declared content.
• Any statement the customer uses to describe their own condition, even without asking for a product.
• A report of an adverse reaction, which must be handled as an incident and passed to a person with priority.

The last item deserves special attention. A bad reaction to a product is not an ordinary sales inquiry. The store needs a clear path for this case, with a record and immediate human follow-up, and the agent should not try to investigate what happened on its own.

How to write the agent's rules in five steps

The person who runs the agent needs a process that someone without a programming background can review. The sequence below keeps the rules from being scattered across a long prompt that is hard to audit.

• List the real questions your customers sent over recent months on WhatsApp and social media, without filtering, and group them by type: product, stock, delivery, exchange, health and other.
• For each product that the agent will discuss, write a sheet that covers what the label declares, the manufacturer's directions for use, and what the store is not allowed to say.
• Define in writing what the agent answers directly, what it answers with a standard notice, and what it only forwards to a person.
• Write a short, polite, non-judgmental standard message for stopping the conversation, and agree with the team on who responds and within what timeframe.
• Test with difficult questions before publishing, including questions designed to push the agent into a recommendation, and fix each failure before opening the channel.

The product sheet is the step that suffers most under time pressure. If a product has no sheet, the agent should not answer about it. That refusal may look like a limitation, but it is exactly what stops the model from filling the gap with a promise.

Illustrative scenario: the anxiety question

Illustrative scenario: a customer writes on WhatsApp asking which tub of a certain supplement "fixes anxiety" and whether it can be taken together with the medication their doctor prescribed. The agent identifies two triggers, one for a health condition and one for a possible medication interaction. It does not compare products, does not suggest anything, and does not comment on the medication leaflet. It explains that it cannot advise on health, that the question about the medication should be directed to the doctor or pharmacist, and offers to transfer the customer to a staff member. When the staff member takes over, they receive a summary of the conversation with only the necessary details, so the customer does not have to repeat everything. If the same customer returns later just to ask about delivery time for that product, the agent answers normally, because that question is within scope.

The example shows that stopping does not end the service. It changes who responds. The customer is still helped, and the store stops giving advice that it is not in a position to give.

How Juan Carlo's method applies to a supplement store

The work follows four stages: map, limit, pilot and hand over. In the case of a supplement store, mapping means gathering the real questions and measuring how many are about products, orders and health. Limiting means writing down the data the agent may use and the actions it may take, including the trigger list. Piloting means running the agent on a real flow for a set period, with a record of every stop and a metric agreed in advance, such as the share of handoffs that the team reviews and confirms were correct. Handing over means giving the operation documented rules and deciding whether the next step is to expand, adjust or pause. This path is slower than putting a chatbot online, but it leaves the store with something it can explain and audit.

Frequently asked questions

Can the agent recommend a supplement to someone who trains? It can present the products and state what is declared on the sheet, but it should not recommend a product for a performance or health goal. When the customer describes a condition, the conversation should be passed to a staff member. The same rule applies to any recommendation based on an effect on the body.

What data should the agent ask the customer for? Only what it needs to complete the task, such as name, order number and delivery address. Health information, such as conditions, medication or pregnancy, is sensitive data under the LGPD and does not need to be collected by the agent to answer a stock or exchange question. If the customer shares it voluntarily, the team should have a clear rule for deleting it or restricting its use.

What should we do if the agent gives a wrong answer? Treat it as an incident. Record the conversation, identify which rule was missing or not followed, correct the product sheet or the trigger, and test again before turning the flow back on. If the wrong answer touched on health, a staff member should contact the customer. The goal is that the correction is recorded and does not depend on the memory of whoever was on shift.

Should we start with the whole channel or with a single product? Start with a small set of products and well-built sheets, because it is easier to review the answers and learn from the stops. Once the flow is stable and the team knows how to read the records, the store can widen the catalog. Expanding without reviewing usually just moves the problem to the customer.

Sources

• Anthropic: Building effective agents
• OWASP Top 10 for LLM Applications
• ANPD: Autoridade Nacional de Proteção de Dados