Generative AI isn’t ubiquitous in the business world—at least not yet

Tobacco company Reynolds American, for instance, is taking a go-slow approach, testing gen AI in a limited capacity. It is experimenting with using AI to improve analysis of large data sets, but not in high-risk finance applications or in customer-facing roles, said Chief Information Officer Aaron Gwinner.

“There’s some inherent risks with gen AI," Gwinner said. “Before we run off and we just start doing AI projects, we need to get the foundations and the basics in place."

Microsoft’s Copilot, Google’s Gemini and OpenAI’s ChatGPT have sought to bring generative AI to the business world. Boosters hope the technology will be transformative. Swedish financial-technology company Klarna, for example, in February said a ChatGPT tool now does the work of 700 customer service agents.

But many companies aren’t leaping in, just yet.

Broad surveys of business sentiment show many, like Reynolds, are proceeding with caution. U.S. Census Bureau data released in March found only an estimated 5.4% of businesses use AI of any type to produce goods or services. Businesses in the information sector and large businesses use AI more than others, the bureau’s data show.

Companies most likely to say they intend to adopt generative AI are life sciences, oil and gas, and communication-service providers, with 42% to 49% in those industries reporting that intent, according to a report published in March by Gartner, a research company. Most companies that say they are going to adopt AI, though, ultimately don’t do it, many surveys have shown, a phenomenon Gartner calls a “gap between ambition and achievement."

Organizations are still trying to see how useful generative AI actually is, said Daniel Colson, the co-founder of the AI Policy Institute, a research and advocacy group.

“That’s been really unclear for quite a while. And I think that’s led to uncertainty about whether language models are actually that valuable for business applications," Colson said, referring to the AI systems behind ChatGPT and other programs.

Some have concerns that generative AI isn’t worth the cost. Many organizations simply can’t yet see a use case.

Law firm Luminos Law has thought about using generative AI but hasn’t moved forward, even though the firm specializes in advising clients on AI risk management.

“We’ve talked about it, but the truth is that the real value of generative AI is to get knowledge and insights, at scale, to a lot of people," said Andrew Burt, Luminos’s co-founder. “We are small and pretty elite, and it just doesn’t make sense."

Online mental-health service Koko drew controversy as an early ChatGPT adopter when it revealed last year it had let the popular tool author some empathetic messages sent to users.

Though the actual sending was approved by a human agent, and messages went out faster than they would have, users were put off when they learned a simulated confidant was behind the messages, according to Koko Chief Executive Rob Morris.

“Simulated empathy feels weird, empty," Morris said at the time. Morris said the use of ChatGPT had been cleared by an external board, but he said the company has stopped using the technology until a better use case emerges.

Mind Meld PR, a Vancouver-based public relations agency whose work involves sending pitches to scores of journalists, said a foray into generative AI didn’t turn out as well as it might have hoped since the whole process, including honing the output, took about as long as having staff perform the entire task.

In other instances, the AI would provide confident answers to research questions but later admit its responses weren’t actually real, Mind Meld Chief Executive Jonathon Narvey said.

“We tried. Total disaster," Narvey said.

So-called hallucinations, a phenomenon in which an generative AI gives a confident but made-up answer, have been well-documented in the ChatGPT era.

Air Canada was in February ordered to refund a customer after the airline’s chatbot gave the wrong details about its bereavement policy. Air Canada said the technology behind the bot was older and less sophisticated than current AI, but the episode underscored the risk of letting software make representations to customers. Air Canada has withdrawn chatbots from service, it said.

Though large businesses are more likely than smaller ones to have adopted generative AI, they also see its potential to pose security risks. Ninety-two percent of respondents to a recent Cisco Systems’ survey of privacy and security professionals said they believed generative AI was fundamentally different from other technologies and required new techniques to manage data and risks. More than a quarter had gone as far as banning its use.

Time might be the main factor limiting widespread adoption. Search engines and social media, two groundbreaking technologies of the internet era, took years to adopt. By 2002, four years after Google incorporated, only 29% of internet users used any search engine on a typical day, according to the Pew Research Center. ChatGPT amassed 100 million monthly users just two months after its launch in November 2022, making it the application with the fastest-growing user base in history at that point, according to analysts at UBS.

But generative AI could ultimately follow the same trajectory toward near ubiquitous adoption as those earlier technologies. Microsoft research has found that 77% of users who have tried its generative AI product Copilot don’t want to give it up, the company said.

Reynolds American is trying to balance the competitive gains that can come from moving quickly against the risk of moving too fast, Gwinner said.

“If you think about the revolutions in the last 100 years in technology, this is going to be one of the biggest," Gwinner said. “[But] it’s going to take us time to be ready for it."

Write to Richard Vanderford at Richard.Vanderford@wsj.com