AI in customer service: applications and risks

You open a company’s chat, type your question and get a fast, polite and confident answer. Only it is wrong. Who pays for that mistake?

This question is no longer theoretical. Companies use AI models to answer customers by chat, WhatsApp and phone. This article shows what research says about the gains, what consumers report fearing, a decided case in which a chatbot got it wrong and what Brazilian rules require from those who automate customer service.

What AI does in customer service today

In practice, AI shows up in customer service in three main ways.

• Self-service. A chatbot, a program that converses by text, answers frequent questions, looks up orders and opens tickets without an agent.
• Agent support. AI suggests replies, finds the right policy and summarizes the history, but a person is the one talking to the customer.
• Triage. AI classifies the request and routes it to the appropriate queue or department.

The form with the most robust evidence so far is the second. A study by Erik Brynjolfsson, Danielle Li and Lindsey Raymond followed the gradual rollout of a generative AI assistant to 5,172 support agents at one company (Brynjolfsson et al., revised version from November 2024). According to the authors, productivity, measured as issues resolved per hour, rose by about 15% on average.

The gain was not the same for everyone. Less experienced agents improved in both speed and quality. The most experienced gained little in speed and saw small declines in quality. The authors also observed that customers became more polite and asked less often to speak to a manager.

Two caveats when reading this result. First, it is a study at one company, with one specific tool, not a market average. Second, the AI there supported people. It does not measure what happens when the customer talks only to the machine.

What customers say they fear

On the consumer side, the available data comes from opinion surveys. A Gartner survey of 5,728 customers, conducted in December 2023 and released in July 2024, found that 64% would prefer companies not to use AI in customer service (Gartner, July 9, 2024). In the same survey, 53% said they would consider switching to a competitor if they found out a company was going to use AI for customer service.

The main concern cited was the difficulty of reaching a person. Next came job displacement and wrong answers from AI. The press release does not detail which countries were in the sample, so these numbers cannot be taken as a portrait of the Brazilian consumer. They do indicate where the distrust lies: in blocked access to humans and in errors.

When the bot gets it wrong, the company answers for it: the Air Canada case

The most cited case on the subject was decided on February 14, 2024, by the Civil Resolution Tribunal, a small claims tribunal in the Canadian province of British Columbia (Moffatt v. Air Canada, 2024 BCCRT 149).

The facts, according to the decision:

• In November 2022, after his grandmother died, Jake Moffatt asked the chatbot on Air Canada’s website about bereavement fares, a discount for people traveling due to a death in the family.
• The chatbot replied that it was possible to request the discount after the trip, within 90 days of the ticket being issued.
• The airline’s own page on the subject said the opposite: the policy did not apply to requests made after travel.
• Moffatt bought the tickets at full price and requested the refund within the period indicated by the bot. The company refused. In an email, a representative admitted that the chatbot had used misleading words.

In the proceedings, Air Canada argued that it could not be held liable for information provided by the chatbot. The tribunal called the argument “remarkable” and wrote that, although it has an interactive component, the chatbot “is still just a part of Air Canada’s website.” For the tribunal, it makes no difference whether the information comes from a static page or a chatbot.

The decision also rejected the idea that the customer should check the bot’s answer on another page. The tribunal found that the company did not take reasonable care to ensure its chatbot was accurate. Air Canada was ordered to pay CA$812.02: CA$650.88 in damages, plus interest and tribunal fees.

The amount is small. The principle is what matters: the bot’s answer was treated as information from the company. It is a decision by a Canadian small claims tribunal. It does not serve as precedent in Brazil, but it shows the kind of risk a chatbot creates. Errors of this kind have a name: AI hallucinations, false answers given with confidence.

What Brazilian rules say

As of the date of this article, Brazil has no specific law on chatbots in customer service. But general consumer rules already apply to any channel. Below is what they say. The reading of how they apply to AI is the editorial team’s, not a court’s, and does not replace legal advice.

Consumer Protection Code (CDC). Law No. 8,078/1990, Brazil’s Consumer Protection Code, contains at least four relevant points:

• Article 6, item III, guarantees adequate and clear information about products and services;
• Article 14 states that service providers are liable, regardless of fault, for damages caused by defects in the service and by insufficient or inadequate information;
• Article 30 states that sufficiently precise information, conveyed by any means, binds the supplier that publishes it;
• Article 34 makes the supplier jointly liable for the acts of its agents or independent representatives.

In the editorial team’s assessment, the text of the CDC makes no exception for answers generated by software. Wrong information given by the chatbot on the company’s behalf is likely to be treated as information from the company.

SAC decree. Decree No. 11,034 of April 5, 2022 regulates the SAC (Serviço de Atendimento ao Consumidor, Brazil’s mandatory customer service channel). Note its scope: it applies to providers of services regulated by the federal Executive Branch, not to all businesses. For these providers, the decree provides, among other points:

• access to the SAC 24 hours a day, seven days a week, through at least one channel (Art. 4);
• mandatory phone support, with at least eight hours a day of human service and complaint and cancellation options in the first menu (Arts. 4 and 5);
• a ban on asking consumers to repeat their request after it has been recorded in the first contact (Art. 10);
• the right to track the request by a reference number and to receive its history (Art. 12);
• a response within seven calendar days, in a clear, objective and conclusive manner (Art. 13).

The decree does not mention AI, but several of these rules directly affect automated service. A bot that does not let the customer reach a person on the phone, that forces them to tell everything again to the agent or that does not generate a reference number conflicts with the text.

General Data Protection Law (LGPD). The SAC decree itself refers to Law No. 13,709/2018, the LGPD (Brazil’s data protection law), for the processing of consumer data (Art. 9). The LGPD, in turn, gives data subjects the right to request review of decisions made solely on the basis of automated processing that affect their interests (Art. 20). If the customer service AI decides on its own, for example, to deny a request or classify a credit profile, this right comes into play.

Human handoff: the point that decides the experience

In customer service, handoff is the transfer of the conversation from the bot to a person. This is where the distrust measured by Gartner is concentrated.

In the same press release, Gartner recommends that the chatbot tell customers it will connect them to an agent when it cannot solve the problem. The conversation should then pass to the human, picking up where the bot left off.

Based on the sources above, the editorial team suggests some practical criteria for those designing this flow:

1. Visible exit. The customer should know from the start how to ask for a person.
2. Clear triggers. Complaints, cancellations, disputed charges, requests involving money and customers who repeat the same question are good reasons to hand off.
3. Context goes along. The agent receives the conversation history. The customer repeats nothing.
4. Reference number from the bot onward. The request is logged from the first automated contact.
5. Limits on what the bot promises. Deadlines, refunds, discounts and policy exceptions should come only from the company’s official sources, preferably with the answer grounded in the right document (see RAG and AI with documents).

Other risks beyond the wrong answer

Outdated information. In the Air Canada case, the chatbot and the official page said different things. When policy changes, the knowledge base used by the bot needs to change with it.

Actions, not just answers. An assistant that only answers makes mistakes in text. One that cancels, refunds or changes account details makes mistakes in actions. That is the leap from chatbots to AI agents, and it calls for limits and human approval for sensitive operations.

Personal data. Customer service conversations include CPF (Brazil’s individual taxpayer number), addresses, payment and health data. Processing follows the LGPD, including when an AI vendor processes these conversations.

Lack of measurement. Without tests using real questions and periodic review of conversations, the company only discovers the error through complaints. We show how to set up this check in how to evaluate the quality of AI responses.

Many of these problems are not technical but design problems: poorly defined goals, bad data and little governance. We cover this in why enterprise AI projects fail.

An example from the channel

The short Token & Tech channel video attached to this article (in Portuguese) covers the use of AI in the real estate market: listings, property valuation, photos, projects and customer service automation. According to the video description, the team tested AI for customer service on WhatsApp and found that human contact still makes a difference. It is an account of experience, not a study, but it points in the same direction as the Gartner survey.

Key takeaways

• According to the available evidence, AI in customer service works best as support for the agent. The study of 5,172 agents measured an average gain of about 15% in issues resolved per hour.
• Customers distrust mainly the difficulty of reaching a person, according to a Gartner survey from December 2023.
• In the 2024 Air Canada case, the tribunal treated the chatbot as part of the website and held the company liable for the wrong information.
• In Brazil, the CDC binds the supplier to the information it publishes, and the SAC decree requires, in federally regulated services, human phone support, a reference number and a response within seven days.
• Design the handoff before the bot: a visible exit, clear triggers, preserved context and limits on what the AI can promise.