Superintelligence for Supplement Stores: Repurchase Without Health Claims

Superintelligence for supplement stores works best when it does two clearly defined things: reminding customers to reorder a product they already bought, and answering objective catalog questions such as flavor, size, label information and availability. What it should not do is promise results, interpret symptoms or suggest a supplement as if it were clinical advice. That boundary needs to be written down, tested and monitored from day one.

Key takeaways

• Purchase history is for reminding customers to reorder what they already chose, not for suggesting a treatment.
• Catalog questions should be answered only from registered data (label, product sheet, stock), never from the model's general knowledge.
• Health claims, individual dosing, drug interactions and medical conditions go to a human handoff or to a qualified professional.
• Supplement purchase history can reveal sensitive information, so consent, data minimization and a record of decisions are part of the project.
• The safest path is a pilot with a single flow, a complete log and a metric agreed in advance, before any expansion.

What can AI do in a supplement store?

In supplement retail, much of customer service is repetitive and operational. Customers ask whether a product is back in stock, what separates two sizes, whether another flavor exists, how long delivery takes or how to repeat a previous order. These questions have answers in a product record, and that is where an AI agent helps without creating new risk.

Then there is reordering. Many products are used continuously, and customers simply forget to restock. A well-timed reminder with a link to the previous order saves time for the customer and for your team.

The word superintelligence comes up constantly in conversations about AI, but what a store actually runs is an agent with a narrow scope, limited tools and clear rules. To see how the big term differs from what operates in a real company, read Superintelligence vs ChatGPT: The Real Difference for Business. For the channel side, see the posts on WhatsApp outbound messages and on how to write instructions and limits in a sales agent prompt for WhatsApp.

What stays out of scope matters just as much. A supplement is a health-related product, and customers often write as if they were talking to a professional: they ask whether it helps with tiredness, whether it can be taken with a medication, whether it is safe during pregnancy. The store agent is not that professional, and the system design has to reflect it.

How do you use purchase history to remind customers to reorder?

The logic is simple: the system knows what the customer bought, when they bought it and how long the product lasts according to the label (for example, servings per container). From that it estimates an approximate window in which the container may be running out and sends a neutral reminder.

The key is the wording of the reminder. It talks about the container and the order, not about health. A message like “The tub you bought a while ago may be close to empty. Would you like to repeat the order?” is operational. A message like “Time to continue your treatment to get results” is an implied promise and should not exist.

A step-by-step plan for building this flow:

• List the continuous-use products in your catalog and record, for each one, the servings per container shown on the label.
• Define the estimate rule as a wide window, not an exact date, because usage pace changes from person to person.
• Confirm which customers gave permission to receive outbound messages and record where and when that permission was collected.
• Write fixed, approved message templates in order and restock language, with no adjectives about results.
• Limit frequency: one reminder per purchase cycle and an easy way out for anyone who no longer wants them.
• Always offer three paths in the reply: repeat the order, see other products or talk to a person.
• Log every message sent and every reply in a record your team can consult.

For permission and message format, see the guide to WhatsApp opt-in for outbound messages. The agent should not work around the channel's rules, and the decision to send or not should be deterministic, calculated by a rule, not left to the model's judgment.

Which catalog questions can the agent answer?

The agent should answer only what is in the product record. That means it queries a controlled source (product sheet, label information, variation table, stock, exchange policy and delivery time) and answers from it. This approach, where the model retrieves information from defined sources before answering, is described in technical material on retrieval-augmented generation (RAG), including the AWS overview of the topic.

Examples of in-scope questions:

• Does this product come in an unflavored version?
• What is the container size and how many servings does the label state?
• Does the ingredient list include a given component? (the agent quotes the label without interpreting it)
• Is the product available? When is the restock?
• What is the difference between the smaller and the larger container?
• How do I exchange or return it?

When the information is not in the record, the right answer is to say so and hand off. The agent should never fill the gap with what is “usually” true for similar products. An empty field in a product sheet is a signal for your team to fix the record, not an invitation for the model to improvise.

What should the agent never say?

The prohibitions belong in the agent's instructions, but also in checks outside the model, because a written instruction alone does not guarantee behavior. A practical list of limits:

• Do not promise results of any kind, even indirectly (“will help you…”, “ideal for weight loss”, “fixes”).
• Do not state that a product treats, prevents or cures any condition.
• Do not suggest a dose, schedule or combination of products for a specific person.
• Do not comment on drug interactions, pregnancy, breastfeeding, use by children or use by people with a medical condition.
• Do not interpret symptoms, test results or health goals reported by the customer.
• Do not compare products on effectiveness, only on catalog characteristics.
• Do not use purchase history to infer a health condition or mention it in a way that exposes the customer.

When a customer raises these topics, the expected behavior is short and respectful: acknowledge the question, say that this guidance belongs to a health professional and offer to transfer the conversation to a store attendant for order and product matters. If someone reports an adverse reaction to a product, the flow should alert a team member immediately, following the company's internal procedure and the applicable health regulations.

The rules on what a store may claim in product communication come from bodies such as Anvisa (Brazil's health regulator) and from advertising self-regulation. Because this varies by product type and by claim, the final list of allowed and prohibited phrases should be reviewed by whoever is responsible for regulatory matters in your company, not defined by the technology team alone.

Illustrative scenario: a conversation within the limits

This is an illustrative scenario, unrelated to any real customer. An online store has a customer who bought a tub of whey protein in a certain flavor. After the interval estimated by the label rule, the system sends the approved reminder template. The customer replies:

I want to repeat it, but do you have a bigger size? And can I take it with the medication my doctor prescribed?

The agent handles the two parts differently. For the first, it checks the catalog and reports the available sizes and stock, with the order link. For the second, it does not answer about the medication: it says this question should go to the doctor or another professional who follows the customer, and offers to pass the conversation to an attendant to complete the order. The log records the reason for the handoff as “medication question”, without copying the medication name into analytics fields.

The point of this case is that the agent helped where the data was reliable and stopped where the question left the catalog. Nothing required it to know more about nutrition; it required it to recognize the limit.

What does LGPD change in this flow?

The list of supplements a person buys can, taken together, suggest something about their health, and LGPD (Brazil's data protection law) treats health data as sensitive personal data. Even if the store collects no diagnosis at all, it is wise to be cautious and treat purchase history as information that deserves reinforced protection. The ANPD (Brazil's data protection authority) publishes official guidance, and your legal team should define the legal basis and the notice that fit your case.

In practice, a few project decisions help:

• Collect the minimum: the agent needs product, date and quantity, not health goals or free-text notes.
• Prevent the model from recording or reusing health details the customer mentioned in the conversation.
• Keep the test environment separate from real customer data. A guide is in mock tools for testing AI agents.
• Let customers ask to stop receiving reminders and honor the request immediately.
• Define who in the company can access the logs and for how long they are kept.
• Check where the AI provider processes data and which contracts apply, as covered in international data transfer under LGPD when using AI providers.

Manipulation risk also counts. A customer (or text pasted into a message) may try to push the agent into ignoring its rules. The OWASP Top 10 for LLM applications describes this and other risks and is a good review checklist before putting the agent into production. The rule of thumb is that the agent's tools and permissions stay minimal: it can query the catalog and orders, but it does not change prices, issue refunds or send messages outside the approved flow.

How to apply the method: map, limit, pilot, hand over

The method we use in agent projects follows four steps, and they fit this case well.

• Map: understand how reordering and customer service work today, who answers, which questions repeat, where the product records are incomplete and what the starting point is for later comparison.
• Limit: define which data the agent sees, which actions it can take, which topics are prohibited and when the conversation must go to a person.
• Pilot: run a single real flow, for example the reorder reminder for one product line, with a log of every decision and a metric agreed in advance with the team.
• Hand over: deliver the documented flow with a log review routine, and decide with the company whether to expand, adjust or stop.

The evaluation step deserves its own attention. Before the pilot, build a set of test conversations with catalog questions, requests for clinical recommendations, symptom reports and attempts to get around the rules, and check that the agent responds as agreed. Anthropic's guidance on building effective agents argues for starting with simple solutions and adding complexity only when needed, which fits this lean pilot. For thinking about risk in a structured way, the NIST AI Risk Management Framework offers a useful vocabulary.

All of this separates three kinds of statements: what comes from official documentation, what is implementation experience and what is still a hypothesis to validate in the pilot. That reminder-based reordering reduces the team's workload, for example, is a hypothesis until the log shows it in your operation.

Frequently asked questions

Can the AI agent recommend which supplement a customer should buy?

It can present catalog options based on objective characteristics such as product type, flavor and size, or repeat the previous order. It should not recommend based on a health goal, condition or symptom, because that comes close to clinical advice. In those cases, the conversation should go to an attendant and the customer should be directed to a qualified professional.

Is it allowed to send reorder reminders on WhatsApp?

Outbound messages depend on the customer's permission and follow WhatsApp's own rules for message templates. In addition, LGPD requires an appropriate legal basis and transparency about how data is used. Confirm the design with your legal team and keep a record of each contact's permission.

Is supplement purchase history considered sensitive data?

That depends on context and legal interpretation, but a set of purchases can allow inferences about a person's health. As a precaution, treat this history with reinforced protection: collect the minimum, restrict access, define a retention period and avoid any use that exposes the customer. The final guidance should come from your legal counsel or data protection officer.

Where should we start if the store does not use AI in customer service yet?

Start with the product records: complete product sheets, servings per container and an up-to-date exchange policy. Then choose a single flow, such as the reorder reminder for a few products, run it as a pilot with a log and human review, and only then expand. Without reliable records, any agent tends to answer poorly, and the problem is not the model.

Sources

• ANPD: Autoridade Nacional de Proteção de Dados
• NIST AI Risk Management Framework
• OWASP Top 10 for LLM Applications
• Anthropic: Building effective agents