Who Is Liable When Company AI Makes a Mistake? Brazil CDC and LGPD

Who Is Liable When Company AI Makes a Mistake? Brazil CDC and LGPD

When your company's AI makes a mistake with a customer, the one who usually answers to that customer is your company, not the tool vendor. Brazil's Consumer Defense Code (CDC) treats customer service as part of the service you provide, and the LGPD (Brazil's data protection law) assigns responsibility to whoever decides how personal data is used. A vendor contract and internal processes do not remove that responsibility toward the customer. They determine how much of the cost you can share later and how likely the mistake is in the first place.

This article is informational and is not legal advice. It separates what is in the law, what tends to happen in real implementations, and what is a hypothesis to validate with your legal counsel.

Key takeaways

• Toward the consumer, the company offering the service answers for what its customer service says or does, including when the answer came from an AI.
• The LGPD distinguishes controller and processor, but both can be held liable for harm to data subjects, and a right of recourse exists between them.
• A contract with the AI vendor divides the bill between the parties, but it does not remove the company's obligation to make the customer whole.
• The biggest reduction in exposure comes from process: clear limits, human review on sensitive topics, and a record of what the AI answered and why.
• If you cannot show what the AI did, you will struggle to defend yourself, so logging and traceability are part of the legal defense.

What does Brazil's Consumer Defense Code say about AI errors?

The CDC does not mention artificial intelligence, and it does not need to. The starting point is that the supplier is liable, regardless of fault, for repairing damage caused by defects in the service and by insufficient or inadequate information (art. 14). If your WhatsApp agent states the wrong deadline, promises something the company does not offer, or gives bad instructions for a procedure, the customer sees a single supplier: you.

Other provisions tend to come up in disputes of this kind:

• A sufficiently precise offer binds the supplier (art. 30). If the AI presents a condition clearly, the customer can demand that it be honored.
• The supplier answers for the acts of its agents and representatives (art. 34). In practice, an automated assistant tends to be treated as an extension of the company.
• When more than one party is responsible for the harm, liability is joint (art. 7, sole paragraph, and art. 25). That lets the customer pursue whoever is closest at hand, normally the company that sold to them.
• A judge may shift the burden of proof when the claim is plausible or the consumer is at a disadvantage (art. 6, VIII). If the customer shows a screenshot of the conversation and the company has no record of its own, the defense is weakened.
• Contract clauses that rule out, waive or reduce the duty to compensate in a consumer relationship are prohibited (art. 25). A notice such as "our AI can make mistakes" helps with transparency, but it rarely works as an exemption.

The CDC provides defenses, such as the absence of a defect or the exclusive fault of the consumer or a third party (art. 14, par. 3). Saying "it was the AI" or "it was the model vendor" is not usually read as the exclusive fault of a third party when the company chose, configured and deployed the tool to serve the public. That reading is an interpretive tendency, not a guarantee. Each case depends on its facts and on the court that hears it.

What about the LGPD: when does the mistake involve personal data?

The LGPD applies when the mistake involves personal data: the AI exposes another customer's information, uses data for a purpose different from the one disclosed, makes a decision about someone based on incorrect data, or leaks conversation content. The text of the law is available on the Brazilian government's official legislation portal (Planalto), and the national data protection authority (ANPD) publishes official guidance.

The basic design is this:

• The controller is whoever decides how data is processed. In most cases, that is your company.
• The processor handles data on behalf of the controller, following the controller's instructions. A customer service platform or model vendor can fill this role, depending on how the service works.
• Anyone who causes property, moral, individual or collective damage through the processing of personal data must repair it (art. 42).
• The processor is jointly liable when it breaches the law's obligations or acts against the controller's lawful instructions. Controllers directly involved in the processing are also jointly liable (art. 42, par. 1).
• Whoever repairs the damage has a right of recourse against the other responsible parties, in proportion to each one's participation (art. 42, par. 4).

The law also provides situations in which an agent is not liable, such as proving that it did not carry out the processing or that no violation occurred (art. 43). At the same time, it requires security measures and good practices (art. 46) and notification of relevant incidents to the ANPD and to the data subject (art. 48). If the AI decides something that affects a data subject's interests, the right to request review is covered in LGPD automated decision-making and the right to review.

To see how each role shows up in a contract, also read controller vs processor in an AI vendor contract.

Is the AI vendor liable instead of me?

Toward the end customer, as a rule, the vendor is not the one you put in front. Between you and the vendor, the relationship is contractual, and that is where risk gets divided. Two moves are common after a judgment or settlement: the company that paid seeks recourse against the vendor, and the discussion becomes what the contract says.

That is why AI contracts deserve attention on points many teams leave for later. In business-to-business relationships, liability limits are usually negotiable, and what is written matters a great deal. What to check:

• The parties' roles under the LGPD: who is controller, who is processor, and for which purposes.
• The vendor's liability cap and its exceptions (for example, security incidents and breach of confidentiality).
• Use of your company's data to train models, plus retention and deletion.
• Subcontractors and where processing takes place, including international transfer.
• Deadline and format for incident notification.
• Audit rights or, at minimum, access to security reports.
• Support and correction when the service's behavior changes without notice.
• An obligation to keep logs that allow the events to be reconstructed.

A practical hypothesis: language model vendors generally offer standardized terms with little room for negotiation by small companies. In that case, what is left to you is choosing well, limiting what the AI can do, and documenting your controls. Confirm current conditions directly in each vendor's terms.

Illustrative scenario: a mistake on WhatsApp

Imagine, as an illustrative scenario, a store that uses an AI agent on WhatsApp to answer questions about exchanges. A customer asks whether they can exchange a product after the deadline, and the agent, without the updated policy in its context, answers yes, at no cost. The customer comes to the store with the conversation on their phone.

What tends to happen in a situation like this:

• The store can hardly defend itself by saying it was an AI error, because the answer went out through the official channel.
• If the conversation is recorded, the company can see exactly what was said and decide knowingly whether to honor the offer.
• If the policy was correct in the agent's knowledge base and the agent ignored it, the problem is configuration and testing, and the conversation with the vendor becomes a different one.
• If the agent had a rule to never promise exceptions and to hand off to a person, the mistake probably would not have happened.

The same scenario involving personal data would be more serious: if the agent had shown another person's order by confusing phone numbers, the LGPD would come into play, with a possible incident notification. The difference between the two cases shows that severity depends on what the AI can access and do, not only on what it says.

How do you reduce exposure through contracts?

The contract is the layer that defines how risk is divided between companies. Use this list as a conversation guide with your legal counsel:

• Map every vendor involved: customer service platform, model provider, WhatsApp API, integration tools.
• Document who is controller and who is processor for each processing activity.
• Require a data processing clause with limited purpose, documented instructions, confidentiality and deletion at the end of the contract.
• Negotiate incident notification with a deadline that lets your company meet its obligations to the ANPD and to data subjects.
• Ask for information about subcontractors and international transfer.
• Agree on what the vendor logs and for how long, and how you access those records.
• Review your customer-facing terms: AI use notices, a channel to reach a person, and a privacy policy consistent with what the agent actually does.

In the consumer relationship, remember that waiver clauses tend to be invalid. The role of customer-facing terms is to inform clearly, not to exempt.

How do you reduce exposure through process?

Process is where real risk drops. A step-by-step plan:

• Define the scope: which questions the AI answers on its own and which it must hand off. Topics such as refunds, commercial terms, health, legal matters and collections deserve extra attention. The article on what not to delegate to AI in business helps draw that line.
• Limit the data: the AI should access only what it needs for that conversation, ideally only the data of the identified customer.
• Limit the actions: start with read-only access and drafting, and require human approval for anything that creates an obligation, such as a discount, a refund or a paid appointment.
• Put controls on the output: checks before sending, as in the guide to guardrails for AI agents, and refusal rules for anything outside the scope.
• Guarantee handoff to a person: the customer must be able to ask for a human, and the conversation must arrive with its history.
• Record decisions: the message received, the context used, the response sent, the action taken and who approved it. These records serve both the defense and the improvement of the system.
• Test before and after changes: hard questions, attempts to manipulate the agent, and edge cases. OWASP maintains a list of risks specific to applications built on language models, the OWASP Top 10 for LLM Applications.
• Have an incident plan: who decides, who tells the customer, who assesses whether the ANPD must be notified, and how to pause the agent quickly.
• Review periodically: samples of conversations, complaints, and changes to company policy.

Risk management as an ongoing discipline is described in a structured way in the NIST AI Risk Management Framework. It is a useful reference for organizing mapping, measurement and management, even for small companies that use only part of it.

How our method applies

The method we use for AI projects follows four moves that connect directly to this legal discussion. Map: understand the current process, who decides what, and which data moves through it. Limit: restrict the agent's data, tools and actions, with handoff to a person on sensitive topics. Pilot: run a real flow with logs and an agreed success metric before expanding. Transfer: hand over the operation with documentation and decide whether to expand, adjust or stop.

None of this replaces legal advice, but it lowers both the chance of the mistake and the difficulty of explaining it afterward. To follow the regulation under discussion in the country, also see Brazil AI regulation: what companies should watch and do now.

Frequently asked questions

Can the company claim the mistake was the AI vendor's fault? Toward the consumer, that claim on its own usually carries little weight, because the company chose the tool and put it into customer service. The discussion with the vendor happens afterward, through the contract and the right of recourse, when there is a basis for it.

Does a notice saying that service is handled by AI and may make mistakes protect the company? The notice is good transparency practice and helps the customer understand who they are talking to. It does not usually work as an exemption in a consumer relationship, because the CDC prohibits clauses that waive or reduce the duty to compensate.

When does an AI mistake become an LGPD problem? When it involves personal data, for example exposing information to someone who should not see it, using data for a purpose different from the one disclosed, or making an automated decision based on incorrect data. In those cases, the company should assess the incident, record the analysis, and check whether the ANPD and data subjects must be notified, with legal support.

Where should you start to reduce risk without stalling the project? Choose a low-risk flow, define what the AI cannot promise, turn on conversation logging, and guarantee handoff to a person. Then review the contract and terms with your legal team before widening the scope, treating the pilot as the stage where you learn where the agent makes mistakes.

Sources

• Planalto: Código de Defesa do Consumidor (Lei 8.078)
• Planalto: Lei Geral de Proteção de Dados (Lei 13.709)
• ANPD: Autoridade Nacional de Proteção de Dados
• NIST AI Risk Management Framework