OpenAI’s ‘Operator’ Agent Can Buy Groceries, File Expense Reports

Many tech companies besides OpenAI have announced the development of similar capabilities.

Operator is available in what OpenAI calls “research preview”—an indication the product has limitations and will make mistakes as it evolves—to ChatGPT Pro users in the United States. ChatGPT Pro costs $200 a month.

OpenAI Chief Operating Officer Brad Lightcap said in an interview that Operator can save time at home and work, where there is “huge opportunity” for automating common tasks. But to start, the company wanted Operator to work with its most active users, who are “more willing to recognize that the product is still very much a research preview,” according to Lightcap.

“It is a fundamental difference in the way that people interact with computers,” he said. “It’s a hard technical challenge, and it’s only as good as it is useful.”

OpenAI is also working with tech firms including Instacart, Uber, eBay, Priceline, OpenTable and Etsy to make their webpages more accessible to users on the Operator home page. The companies don’t have a financial relationship with OpenAI as part of the Operator collaboration, Lightcap said.

Wall Street Journal owner News Corp has a content-licensing partnership with OpenAI.

Today’s announcement marks OpenAI’s first official foray into the intensifying AI agent race. As agent technology has evolved, business software companies from Microsoft to Salesforce and Workday have released versions of agents, which can do things like summarize reports and contact sales prospects and job candidates.

Google and AI startup Anthropic also recently released agents, which are similar to Operator in that they can browse webpages and interact with menus and buttons.

One key difference between the companies is reach. ChatGPT has 300 million weekly active users, plus OpenAI said last fall it had 1 million paying business customers. That user base presents one of the most significant opportunities, compared with those of some of its competitors, for agents to reach a large number of users. OpenAI declined to say how many people pay for its Pro plan.

Operator uses a new AI model called “Computer-Using Agent,” or CUA, which OpenAI said combines its GPT-4o model’s vision capabilities with “advanced reasoning.” The company said it became more optimistic about its models’ image and reasoning improvements over the last year, and CUA was trained to interact with text, buttons and menus people typically see on webpages. 

Still, usability is a challenge for AI agents. Though they have promised to deliver time and efficiency savings by doing things for users, most people aren’t using them in their daily lives. Apple launched its AI assistant Apple Intelligence on its iPhone operating system last fall, but it’s not yet used to help with everyday tasks. Even for businesses, most AI agents are being tested or used in limited ways, where it is less likely they expose private company data or open up cybersecurity risk.

While Lightcap said OpenAI might consider adding specific controls or guardrails for corporate customers, it is currently focused on its first batch of users. He said OpenAI has built privacy, security and control features that help ensure the agent doesn’t veer away from its programming, and, most important, keeps the user in control of the AI.

Some of the harms or misuse of Operator, the company said, include websites designed to trick users, users trying to trick the bot, and “prompt injections” that direct users to send sensitive information or money to malicious sites.

Operator has a feature called “takeover mode” that asks people to take control to enter payment details or login information. OpenAI said Operator also asks for approval before completing higher-stakes tasks like sending email, and it won’t work for banking transactions or making a decision on a job application. Operator won’t use data users have previously shared with ChatGPT to take actions, the company added.

For Instacart, making its grocery delivery services more accessible on Operator means the company can tap the potential of AI agents—and OpenAI’s reach with users—without doing that work on its own. “We’re not trying to build an agent,” said Instacart Chief Product Officer Daniel Danker.

Uber’s collaboration with OpenAI on Operator “gives us the opportunity to shape the product’s development,” according to Sachin Kansal, the ride-sharing company’s chief product officer.

Despite its current limitations, OpenAI determined Operator was ready for a limited release after “taking the time to get it right,” Lightcap said.

Write to Belle Lin at belle.lin@wsj.com