By Cesar Augusto Taborda Lima
A company that puts a chatbot or an AI feature in front of its customers answers for what it says, on the same terms as it would answer for an employee or a page of its website. The warning that “AI may make mistakes”, now found in almost every set of terms, discharges the duty to inform; towards consumers it does nothing more. From 2 November 2026, in Brazil’s regulated sectors, a new customer service decree sets out how automated service must coexist with human service. Between businesses, liability caps hold, provided they were written for a reader who could have negotiated them.
Art. 30. Any information or advertising that is sufficiently precise, conveyed by any form or means of communication in relation to products and services offered or presented, binds the supplier who conveys it or makes use of it and becomes part of the contract that comes to be concluded.
Law 8,078/1990 (Brazilian Consumer Protection Code), art. 30 (free translation).
What the bot promises, the company delivers
In November 2022 a Canadian passenger asked Air Canada’s website chatbot how bereavement fares worked. The bot told him he could buy at the normal price and claim the discount within ninety days. The airline’s actual policy said otherwise: no claims after travel. Air Canada refused the refund and argued before the tribunal that it was not liable for information given by its agents or representatives, a chatbot included. The British Columbia Civil Resolution Tribunal called that a “remarkable submission” and observed that the chatbot is simply part of the airline’s website (Moffatt v. Air Canada, 2024 BCCRT 149, 14 February 2024, paragraph 27). The award was 812 Canadian dollars. The precedent is worth a great deal more.
In Brazil the outcome would be the same by a shorter route. A chatbot’s answer about price, deadline, coverage or cancellation terms is sufficiently precise information, conveyed by the company, about the service it offers. Under art. 30 of the Consumer Code that information binds the supplier and enters the contract. Art. 14 completes the picture: the supplier is liable, regardless of fault, for damage caused by insufficient or inadequate information about the use and risks of the service.
There is a valve, and it is narrow. In January 2026 the Fourth Panel of the Superior Court of Justice held that the duty to honour an online offer admits an exception for justified reasons, such as a gross pricing error or signs of fraud, assessed case by case under objective good faith (REsp 2,167,766, Justice Isabel Gallotti reporting). Carried over to AI, the test protects a company whose bot advertises a car for ten reais. It protects little when the bot invents a plausible refund policy, because an error the consumer had no way of noticing will rarely be called gross. The believable hallucination is the worst case: it convinces the customer and binds the company.
Telling people they are talking to a machine
The Consumer Code requires adequate and clear information about the service (art. 6, III). From that alone it follows that a customer should know when an answer comes from an automated system, because that changes how much trust the answer deserves. Bill 2,338/2023, Brazil’s AI bill, would turn that reading into express text: art. 5, I, of the version passed by the Senate guarantees the right to information about one’s interactions with AI systems, in an accessible, free and easily understood form. The bill has been awaiting the rapporteur’s report in a special committee of the Chamber of Deputies since May 2025, with no movement of its own since.
Abroad the rule is already law. In the European Union, art. 50(1) of the AI Act has required since 2 August 2026 that providers of systems interacting with people inform them that they are dealing with AI, unless that is obvious from the context. Regulation (EU) 2026/1744, which postponed much of the high-risk regime, left that duty untouched and granted only a transitional period, until 2 December 2026, for generative systems placed on the market before that date to add machine-readable marking to their outputs. A Brazilian company whose chatbot serves customers in Europe is inside that rule.
In the United States, the Federal Trade Commission applied existing law to AI products and summed up its stance in a sentence from its then chair, in September 2024: “there is no AI exemption from the laws on the books”. In the same announcement, DoNotPay, which marketed itself as “the world’s first robot lawyer”, agreed to pay 193,000 dollars and to tell its subscribers that the service did not replace a lawyer, because it had never tested whether its answers matched one (FTC, Operation AI Comply). The states went their own ways. Utah requires generative AI to identify itself when a consumer clearly asks and, in regulated professions such as law, health and finance, to do so from the outset of any high-risk interaction. California regulated companion chatbots from 1 January 2026 (SB 243), and Oregon, Washington, Nebraska and Idaho passed laws that take effect in 2027.
Good practice converges: the notice appears in the interface itself, at the start of the conversation, and the terms of use repeat it. A notice buried in clause 14.3 informs little and protects less.
The customer service decree and the button for a human
Decree 13,135 of 9 October 2026, published on the 13th and in force since 2 November, replaced Brazil’s 2022 customer service (SAC) decree and is the first Brazilian rule to address artificial intelligence in consumer service by name. It covers suppliers of services regulated by federal agencies, such as telecommunications, energy, health plans, aviation and transport, and now also those who provide financial services through digital channels under Central Bank supervision (art. 1, sole paragraph).
The central rule sits in art. 3, § 3. The supplier may adopt automated procedures, including artificial intelligence systems, to triage, route and resolve requests, provided the consumer is assured the option of real-time service by a human being. That option has hours: eight in the morning to eight at night, every day, by telephone and on at least one electronic channel (arts. 12, § 1, and 14, § 1). It has a position: the choice of a human and immediate cancellation must be available at every level, the first included, and no channel may use more than five levels to resolve a request (arts. 13 and 16). And it has an entry condition: neither the start of service nor the transfer to a person may depend on the consumer handing over data first (art. 8, sole paragraph). The answer to the request, when it comes, is due within seven calendar days and must be clear, objective and conclusive, addressing every point raised (art. 21), which rules out the generic reply some bots produce so readily.
Anyone designing the product should note a few details. The decree excludes from the concept of SAC any service aimed at offering and contracting products (art. 3, § 4), so the sales chatbot falls outside it; it remains, however, inside art. 30 of the Consumer Code, which is exactly where the Air Canada risk lives. And the decree does not require telling consumers they are talking to a machine; it requires telling them, at the very first level, that they may talk to a person (art. 16, § 1, I). In practice the two notices travel together, and the first follows from the general duty to inform.
What the terms of use can do
Towards consumers, little beyond informing. Art. 51, I, of the Consumer Code voids clauses that prevent, exonerate or attenuate the supplier’s liability for defects in the service, and art. 25 forbids any stipulation to that effect. The warning that AI may err has value, and not a small one: it describes the service, calibrates the consumer’s reasonable expectation (the test for defect under art. 14, § 1) and makes it harder to claim the company promised accuracy. It does not, however, shift onto the customer the loss caused by wrong information the company itself conveyed. Where the consumer is a legal entity, compensation may be limited in justifiable situations (art. 51, I, final part), and that is the only room for limitation the consumer statute leaves open.
The terms also lose to the advertising. If the product is marketed as a substitute for a professional, a clause reading “suggestions only, consult a specialist” sits badly with art. 30, which makes advertising part of the contract. That very mismatch is what cost DoNotPay. The description of the product in the terms must match the website, the ads and what the bot says about itself when asked.
Between businesses the picture changes. The Civil Code presumes parity and symmetry in commercial contracts and directs courts to respect the allocation of risk the parties agreed (art. 421-A, II). In 2023 the Third Panel of the Superior Court of Justice upheld, by majority, a one-million-dollar cap on damages in a contract between a multinational technology company and its Brazilian representative, holding that a medium or large company is not presumed vulnerable and that the cap yields only to fraud or to an express provision for supplementary compensation (REsp 1,989,291/SP). An AI vendor selling to businesses may therefore cap liability for the model’s outputs, exclude lost profits and make compensation conditional on use in line with the documentation. For the clause to survive in court, the customer cannot be so small as to be treated as a weaker party, and the limitation must be written clearly, in a clause of its own, easy to find. What the vendor actually did about testing and controls also weighs, because gross negligence tends to be equated with fraud and brings the cap down.
Bill 2,338, if enacted as it left the Senate, would change little here. Art. 35 keeps consumer relations under the Consumer Code, art. 36 sends the remaining cases to the Civil Code, and art. 37 lets the judge reverse the burden of proof when the victim is the weaker party. The bill confirms what already applies and adds a documentation duty that, in litigation, will be the evidence that the company tested the system before placing it in front of customers.
It seems to me that the most valuable clause in these terms turns out to be the most modest one: the clause that describes frankly what the product does, where it tends to fail and what the user should check before acting. When the system presents itself on screen as an aid to decision and keeps a record that the user saw that notice, the company arrives in court with something to show, when it arrives at all.