Respond to Consumer Complaints in Brazil With AI: Consumidor.gov.br

Respond to Consumer Complaints in Brazil With AI: Consumidor.gov.br

To respond to a complaint on Consumidor.gov.br with AI, first gather the full case history (order, conversations, payments, deliveries) and ask the AI for a draft that cites only documented facts. A person then reviews it, decides the solution and publishes the reply within the deadline shown on the platform. The AI organizes and drafts. The commitment to the consumer is always made by the company, through a human owner.

Key takeaways

• AI helps more with preparation (history, timeline, draft) than with deciding what the company will offer.
• Every draft should cite only facts that exist in some record: an order, a message, an invoice, a protocol number.
• The deadline is a process control, not an AI feature: record the due date of each complaint somewhere visible, with a named owner.
• Consumer data is personal data. It calls for LGPD care, restricted access and a deliberate choice of AI tool.
• Start with one type of complaint, measure with simple criteria, and only then expand.

How does each official channel work, and what changes in the reply?

Consumidor.gov.br is a public service where the consumer files a complaint and the registered company answers directly on the platform. The consumer sees the reply, can comment, and rates whether the problem was solved. That is why the text must be clear and complete, and written for the person who complained, not only for an internal analyst.

Procon works differently. Each agency (state or municipal) has its own procedures, and the company usually receives a notification or invitation to provide clarifications, with a deadline and format set in the document itself. The reply tends to be more formal and may require attached documents. The tone changes. The discipline around facts and deadlines does not.

There are also other public and private channels, such as Reclame Aqui, each with its own logic. If you handle complaints there too, read Handling Consumer Complaint Sites With AI: Draft, Review, Send, because much of the preparation flow is the same.

A note of honesty: the rules, deadlines and screens of these platforms can change. Before building any process, confirm the current conditions on the platform itself and in the document you received from Procon, and keep that check as part of your internal procedure.

What can AI do, and what should stay with a person?

The safest split separates preparation tasks from decision tasks. AI is good at organizing text and information. It should not be the one granting a refund, admitting fault or negotiating terms.

What can go to the AI:

• Summarize the customer's history in a short timeline.
• Identify what the consumer actually asked for (refund, exchange, cancellation, explanation).
• Compare the request with company policy and point out gaps.
• Write a draft in clear, polite language without jargon.
• List what is missing before you can reply (a receipt, an answer from another department).

What should stay with a person:

• Decide the solution and which exceptions the company accepts.
• Approve any promise of a deadline, amount or action.
• Acknowledge responsibility in sensitive cases (health, safety, fraud, data leaks).
• Post the reply on the platform and answer any follow-up from the consumer.

This line matches what we cover in What Not to Delegate to AI in Business: 3 Tests to Apply: when the reply creates an obligation for the company, a human signs off.

How do you gather the history before asking for a draft?

The quality of the draft depends almost entirely on the input. An AI model without context fills gaps with plausible sentences, and in an official complaint a plausible but false sentence can become a problem. The rule is simple: what is not in the case file does not go into the reply.

Build a case file for each complaint with these items:

• The full text of the complaint and the date it was received.
• The reply deadline shown on the platform or in the Procon document.
• Order or contract data: date, product or service, payment method.
• Earlier conversations on WhatsApp, email, phone or chat, with dates and protocol numbers.
• Operational records: delivery, technician visit, service order, invoice.
• The applicable policy (exchange, cancellation, warranty) in its current version.
• Earlier complaints from the same customer about the same issue, if any.
• A note from the responsible department on what really happened.

The last item is the one most often skipped. Before drafting, someone in operations needs to say in a few lines what occurred. If the history contradicts itself, the draft should flag the contradiction instead of picking a version.

If part of the history is in WhatsApp voice messages, a transcript helps but needs checking. See Transcribe WhatsApp Voice Messages With AI: Accuracy, Cost, Privacy before adding that material to the file.

What is the step-by-step process for preparing a draft with human review?

A lean flow that fits a small team can follow these stages:

• Log the complaint: note the source (Consumidor.gov.br, Procon), the date received, the deadline and the owner.
• Build the case file: gather the items from the previous section in one place, without unnecessary data.
• Ask for the timeline: the AI produces a timeline with dates and sources. The reviewer checks each item against the original record.
• Confirm the facts with operations: the responsible department validates what happened and what the company can offer.
• Decide the solution: a person with the authority to do so defines a refund, exchange, repair, explanation or a reasoned refusal.
• Generate the draft: the AI receives the case file, the decision and the desired tone, and writes the reply. The instruction must forbid inventing data and require that any missing information be flagged.
• Review: check dates, amounts, names, promises and tone. Remove defensive or generic sentences.
• Publish: the authorized person posts the reply on the platform or sends it to Procon, within the deadline.
• Record: save the final version, the reviewer, the sending date and the outcome.

A good drafting instruction asks for a predictable structure: acknowledge the reported problem, explain what was found, state the solution with a realistic timeframe, and say how to continue the conversation. If you write instructions for support work, you will find more detail in How to Write Instructions for an AI Support Agent: 5 Blocks.

Tone matters. Replies that sound like a canned template irritate consumers, who read the platform rating as part of the relationship. Prefer direct sentences, the person's name when possible, and a concrete solution. Avoid excessive apologies that do not say what will be done.

How do you control deadlines without relying on memory?

A missed deadline is the costliest error and the most avoidable. AI can help with reminders, but the control must live outside it, in a spreadsheet, CRM or ticketing system, with a named owner.

A simple routine:

• Record the due date when the complaint arrives, copied from the platform or document and never calculated from memory.
• Set an internal deadline ahead of the official one, to leave time for review and for hearing from operations.
• Create alerts for each stage: case file ready, decision made, draft reviewed, reply sent.
• Name a backup for each owner, to cover vacations and absences.
• Review open cases every week, starting with those closest to the limit.

If the company uses automation, workflows in tools like n8n can create the record and notify the owner. In that case, handle failures carefully: an alert that did not fire cannot go unnoticed. The article on n8n Error Handling and Retries: AI Workflows That Don't Break shows how to keep a workflow from failing silently.

What LGPD and security precautions are needed?

A complaint contains personal data: name, contact details, purchase history and sometimes sensitive information. Before pasting that content into any AI tool, the company needs to know where the data goes, who can access it and whether the provider uses the content for other purposes. The ANPD, Brazil's data protection authority, is the official reference for guidance on data protection in the country.

Practical measures:

• Send the AI only what the case needs, and a relevant excerpt instead of a whole document when an excerpt is enough.
• Restrict who can open the case files and keep a record of who consulted them.
• Check whether the tool involves processing outside Brazil. International Data Transfer Under LGPD When Using AI Providers details what to verify.
• Define how long case files and drafts are kept, in line with company policy.
• Consider a model running in a controlled environment when the data is highly sensitive, as described in Running Ollama Local Models in a Company: A Practical Guide.

There is also a technical risk: the complaint text is external content and may contain disguised instructions, such as asking the assistant to ignore its rules. Treat the complaint as data, not as a command, and keep the AI without the power to send anything on its own. The OWASP list of top risks for LLM applications describes this kind of threat, and the guide on Prompt Injection Protection for AI Agents: A Practical Guide offers practical measures. To structure risk management for AI use as a whole, the NIST AI Risk Management Framework is a useful reference.

What does a case look like in practice? (illustrative scenario)

Illustrative scenario: an online store receives a consumer complaint saying the customer asked to cancel a purchase and never got an answer. The analyst opens the case, records the due date shown on the platform and sets an earlier internal deadline.

The case file gathers the complaint, the order, several WhatsApp messages in which the customer asked to cancel, and the logistics status showing the package had already left for delivery. The AI generates the timeline, and the analyst checks each date against the records. One of the customer's messages arrived outside support hours and went unanswered, something the draft needs to acknowledge.

The analyst checks the cancellation policy, and the supervisor decides the solution within her authority. With the decision in hand, the AI writes the draft: it acknowledges the delay, explains what was found, states the solution and says how to follow up. The supervisor reads it, cuts a sentence that promised more than the company could deliver, and authorizes sending. The analyst posts the reply, attaches the record in the system and schedules a check for any follow-up from the consumer.

The value of the process is not the speed of the text. It lies in three guarantees: nothing was stated without a record, a human made the decision, and the deadline was met with room to spare.

How do you apply the method: map, limit, pilot, transfer?

The method we use for AI projects fits this case well.

• Map: describe how complaints arrive today, who answers them, which systems hold the history and where deadlines slip. Record the starting point before changing anything.
• Limit: define which data the AI sees, which actions it never takes (publishing, promising, refunding) and which types of case stay out, such as health or safety situations.
• Pilot: pick one recurring type of complaint and run it with logs, mandatory human review and an agreed metric, for example the time to prepare a draft and the number of corrections the reviewer makes.
• Transfer: hand the documented flow to the team, with owners, a review routine and a clear criterion to expand, adjust or stop.

If the company already uses AI in customer service more broadly, connect this flow to what was defined in AI Customer Support: What to Automate and What to Keep Human, so that tone, escalation and record-keeping rules are the same across all channels.

Frequently asked questions

Can I use AI to respond to a Consumidor.gov.br complaint without review?

It is not recommended. The reply creates a commitment from the company to the consumer and stays visible on the platform. Use the AI to organize the history and write the draft, and let an authorized person decide the solution, review and send.

What should I do if the AI includes information that is not in the history?

Remove the information and investigate the cause. Usually the case file lacks context or the instruction does not explicitly forbid inventing data. Adjust the instruction to require that gaps be marked as pending, and keep fact-checking as a mandatory step.

Does the same reply work for Procon and Consumidor.gov.br?

Rarely without adaptation. The channels have different formats, audiences and requirements, and Procon may ask for documents and formal clarifications. The facts in the case file are the same, but the text should follow what each channel asks for, according to the current rules they provide.

How do I start without building a big project?

Pick one frequent type of complaint, set up the case file template and use the AI only for the timeline and the draft, with human review. Record how long preparation took, how many corrections were needed and whether any deadline got tight. With that data, the team decides whether to expand, adjust or stop.

Sources

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