ChatGPT, launched by OpenAI in November, quickly went viral for its often elegant, information-packed responses to various questions, gripping the imaginations of regular people, business leaders and investors including

Microsoft Corp.

, which began backing OpenAI in 2019 and said Monday that it would make a multibillion-dollar investment in the startup. 

OpenAI last week said it would soon add ChatGPT, which stands for chat generative pre-trained transformer, to its application programming interface, or API, which lets developers embed OpenAI technology into their own products.

But customer-experience executives said overreliance on such AI models could lead to companies dishing out incorrect information to customers online without knowing they are doing so. 

While many chatbots are trained to deliver a version of “I don’t know” to requests they cannot compute, ChatGPT, for example, is more likely to spout off a response with complete confidence—even if the information is wrong.

“We don’t want to be in the bad answer business,” said

John Willcutts,

vice president and general manager of digital at Nice Ltd., a customer-experience software company. “A really bad answer in a very critical situation would be a very real problem.”

Sam Altman,

chief executive of OpenAI, has warned against relying on ChatGPT “for anything important right now.”

“Fun creative inspiration; great! Reliance for factual queries; not such a good idea,” Mr. Altman wrote in a tweet last month. 

Using AI to write chat responses in sensitive situations has backfired. Koko, a chat app used for emotional support, this month was criticized for an experiment in which human volunteers crafted their responses to the app’s users with the help of GPT-3. 

Koko’s co-founder said in a tweet that the startup pulled the AI from its system: “Once people learned the messages were co-created by a machine, it didn’t work. Simulated empathy feels weird, empty.”

But for a more typical customer-service interaction, such as querying the status of an online order or editing account details, the technology could prove useful.

Fanatics Inc., a seller of sports memorabilia, digital collectibles and trading cards, said it plans to use a customer-service chatbot fueled in part by GPT-3 when it launches an online sports-gambling division this year.

The company hopes a fast, reliable chatbot will be a differentiator for customers, said

Hollis Donaldson,

vice president of operations for the new division. “Speed equates to great customer experience in the betting and gaming industry,” he said. 

Fanatics’ customer-experience team is testing the chatbot before making it live, conscious of the risks using AI carries if not properly managed, Mr. Donaldson said.

Chasing the dream

Companies for decades have searched for automated solutions that can resolve customer requests as well as humans, or even better. But chatbots are often seen as clunky and not very helpful.

“There was a lot of hype around chatbots, probably five, six years ago, and a lot of vendors wanted to make people believe that it was magical, that it worked out of the box, that it was easy,” said

Yves Normandin,

vice president of AI technologies and products at Waterfield Technologies, a contact-center solutions provider owned by WTI Holdings LLC. “But the reality is that it wasn’t.“

ChatGPT stands out for its ability to provide reasonable-sounding answers to most prompts, regardless of users’ spelling, grammar and phrasing, and to respond in full, natural-sounding sentences that don’t require scripting, said

David Truog,

a principal analyst specializing in technology and design at

Forrester Research Inc.

It is also trained to admit to mistakes, challenge incorrect premises and reject inappropriate requests, according to OpenAI.

But companies should exercise care when dealing with the new AI, Mr. Truog said. “It’s appropriate to be doing some experimentation,” he said, “but it’s too early to deploy mission-critical systems based on this.”

Putting it into practice

Fanatics said its sportsbook’s chatbot will run on technology from Ada Support Inc., a customer-service automation platform. Ada has integrated GPT-3 and other such AI systems known as large language models into its chatbot offering, according to the company’s co-founder and chief executive,

Mike Murchison.

Mr. Murchison said Ada allows clients to customize these large language models by adding company-specific information or anonymized customer data, and deleting irrelevant material. Ada encourages clients to continually update their customized bots’ information, for instance when prices or company policies change, he said.

“Most brands are going to underestimate the importance of continuously improving this over time,” Mr. Murchison said. 

Some Ada clients are restructuring their customer-service organizations to put some contact-center workers in charge of monitoring chatbot conversations, reviewing where the technology gets things wrong or can’t answer, and feeding it new or updated information, he said.

Fanatics plans to follow that approach, as well as ensure that its chatbot interface lets customers reach a human right away, Mr. Donaldson said.

Nice is also working on building OpenAI’s language models into chatbots, Mr. Willcutts said, adding that the company plans to run more tests and fine-tune more models before selling its own take on ChatGPT to clients.

“We don’t get a chance to make a second impression on this one,” he said. “You do this badly once and it’s in the newspaper, and that’s not the kind of reputational risk we’re prepared to take.”

Write to Katie Deighton at katie.deighton@wsj.com