Superintelligence for Cosmetics: Batches, Anvisa and AI Customer Support

Superintelligence for cosmetics, in practice, is an AI layer that organizes what the company already records: batches, product documentation, customer support questions and response deadlines. It helps the team find information, sort messages and prepare drafts, but it does not decide what counts as a claim, efficacy or adverse reaction, and it does not replace the technical manager. The value comes from limiting what the AI can read and defining where it must hand the conversation to a person.

Key takeaways

• The word 'superintelligence' helps little in a decision. What matters is the design: which data the AI reads, what it can answer and where it stops.
• Batch traceability is the foundation of any answer about a specific product, and the AI should only query records that already exist and are reliable.
• Regulated topics, such as claims, efficacy, safety and possible adverse events, need a fixed rule that sends them to a human.
• Customer support is the most useful entry point, as long as the AI classifies and logs messages instead of improvising technical explanations.
• A pilot with one type of question, a log of every conversation and an agreed metric before launch prevents uncontrolled expansion.

What the AI layer does and what it does not do

The first decision is to separate function from responsibility. The company remains responsible for registration, labeling, published information and its relationship with Anvisa, Brazil's health regulator. The AI enters as operational support, with narrow permissions, and not as a source of regulatory truth.

In practice, the layer can:

• Look up, from a batch code, the manufacturing date, the internally approved label version and the product sheet.
• Classify a customer message as a usage question, an exchange request, a complaint, a technical information request or a suspected reaction.
• Answer routine questions the company has already approved, such as storage, general usage instructions and delivery times.
• Draft replies for human review in cases that require care.

The same layer should not interpret Anvisa rules, compare efficacy with other products, advise on skin treatment or decide on refunds. These points belong to people with technical and legal responsibility for the subject.

Batch traceability: the data that supports the answer

A question like 'is this the same batch I bought?' can only be answered if the company reliably knows which batch went to which sales channel. In many operations, this data is scattered: part in the inventory system, part in invoices, part in distributor and marketplace spreadsheets. Before thinking about AI, it is worth mapping where each piece of information originates, who owns it and how often it is updated.

The logic is the same one used in other industries. When a company starts with quality and maintenance, with traceable data and clear processes, the AI has something solid to rely on. This reasoning appears in our article on AI in manufacturing, starting with maintenance and quality.

If the batch data is incomplete, the AI must not fill the gaps. It should state that it could not find the record and route the case to a person. An invented answer about traceability carries a bigger risk than a delay of a few hours.

Regulatory requirements: where the boundary sits

Anvisa sets rules for cosmetic products covering registration, labeling, claims, warnings and consumer information. For the AI layer, the practical consequence is simple: any text the customer reads must come from content the company has approved, not from free generation by the model.

A useful rule is this: if no approved text exists for a question, the AI routes it. It does not complete the answer with general knowledge. The following topics should always go to a person:

• A claim of efficacy, a comparison with another product, or a question about expected results.
• A report of irritation, burning, allergy, swelling or any symptom after use.
• A request for a recall, a formal complaint or a threat to contact a consumer protection agency.
• Use during pregnancy, by a baby or child, on damaged skin or in any other condition that requires professional guidance.
• A question about an ingredient that does not appear on the approved product sheet.

The boundary also applies to what the AI writes on social networks and third-party channels. When the company uses AI to reply on other channels, the same list of forbidden topics must apply.

AI customer support: triage, signals and handoff

Customer support holds most of the practical benefit. Repeated questions about use, delivery, exchanges and batches can be handled by an agent with clear limits, freeing the team for cases that require judgment. The breakdown of which tasks to automate and which to keep with people is covered in AI customer support: what to automate and what to keep human.

Illustrative scenario: a customer writes on WhatsApp saying her skin turned red after the second application and asks for the batch number of the product. The agent asks for the code printed on the packaging, finds the batch in the internal record and does not suggest any treatment or diagnose the cause. It tells the customer that a member of the team will continue the conversation, opens a record with the batch, the message, the time and the photo she sent, and routes it to the technical manager within a defined deadline. At no point does the AI state whether the product caused the reaction or not.

This example shows three elements that should exist in any flow like this: secure identification of the product, a block on diagnosis, and a handoff with enough context so the person does not have to ask everything again.

A step-by-step plan to start

For a cosmetics company that wants to test this layer without betting everything at once, the sequence usually looks like this:

• Map the real questions received by customer support, separating routine questions from regulated topics.
• List where batch, expiry and distribution data live, with the owner of each source identified.
• Define, together with the technical manager, the set of approved texts the AI may use in answers.
• Write the handoff rules: what goes to a person, who receives it and within what deadline.
• Test the agent with real questions and risk cases before any customer talks to it.
• Run a pilot with one type of question, record each conversation and measure one agreed metric, such as the time until the correct handoff.
• Decide at the end: expand, adjust or stop.

The step that slows this process down most is the approved text library. If the company lacks an organized repository of official answers, the AI will reproduce whatever it finds in inconsistent places. Organizing this material is human work, and it is usually the investment with the highest return at the start.

How we apply the method: map, limit, pilot and hand over

In our work with companies, the sequence is to map the current customer support and traceability process, limit data and actions, pilot a real flow with logs and an agreed metric, and then hand over the operation with written rules and decide whether to expand, adjust or close it. In cosmetics, limitation is the most important step: the AI reads what it needs, does not write to regulated systems and does not change batch status. When the operation requires ongoing monitoring, maintenance is handled by Be Human (be-human.ai).

Personal data, permissions and decision records

Customer support conversations contain personal data, such as names, phone numbers, addresses and skin photos. Under the LGPD, Brazil's data protection law, the company should collect only what is necessary, inform customers how it is used and define how long each record is kept. Whenever an automated decision affects the consumer, the matter deserves specific analysis, as we discuss in LGPD automated decisions and the right to review.

On the technical side, the rule is least privilege. The agent should have read access to the batch registry and nothing more, with no power to issue refunds, change inventory or publish content. Customer messages may also contain instructions that try to change the agent's behavior, and this must be treated as untrusted input. The records of each conversation should show which data was consulted, which text was used and why the conversation was or was not routed to a person. This trail is what allows a case to be reviewed months later.

Frequently asked questions

Can AI tell whether a cosmetic product is safe for pregnant women? No. That kind of guidance depends on information approved by the company and, in many cases, on professional advice. The agent should recognize the topic, avoid answering about safety and route the conversation to a person with the context recorded.

Do I need a superintelligence to set this up? Not necessarily. What supports the operation is organized batch data, approved texts, handoff rules and logs. Before discussing any technology, define these four points, because they determine what the AI can safely do.

How does the AI identify a product's batch? Usually through the code printed on the packaging, sent by the customer as text or photo and matched against the internal batch registry. If the code is not found or no matching record exists, the AI should not guess. It informs the customer that it could not locate the information and routes the case.

Where should I start without buying a new system? Start with customer support: classify the questions from recent months, organize the approved texts and define the handoff rules. Then choose one type of question for a pilot with logs and an agreed metric. If you want to see how this work is carried out, read how I work with companies as an AI consultant.

Sources

• NIST AI Risk Management Framework
• OWASP: Authorization Cheat Sheet
• ANPD: Autoridade Nacional de Proteção de Dados