AI's key brain trusts are already looking beyond today's popular chatbots to a world in which AI agents act on your behalf.

Why it matters: As AI races from just saying things to doing things for us, its potential benefits and harms will multiply fast.

Driving the news: ChatGPT maker OpenAI is deep into development of AI agent systems, as the Information recently reported.

• Former Salesforce co-CEO Bret Taylor and former Google executive Clay Bavor have launched Sierra, a company that aims to fully automate a huge range of online customer interactions using AI agents.

Between the lines: Such "set and forget" systems that take care of things for you are convenient and wonderful — when you can trust them.

• But generative AI is still very young — and even at its best is still prone to making things up and going off the rails.
• In recent days, ChatGPT started spitting out gibberish, while Google was forced to halt image generation of people after its Gemini engine was generating images with a range of races even in queries where such diversity didn't make sense, such as an image of Nazi-era German soldiers.

The big picture: Moving to a world in which AI systems take actions rather than just provide insights will requires a significant shift in both technology and human behavior.

• "Once we get agent-like systems working, AI will feel very different to current systems, which are basically passive Q&A systems, because they'll suddenly become active learners," Google DeepMind CEO Demis Hassabis told Wired.
• "Of course, they'll be more useful as well, because they'll be able to do tasks for you, actually accomplish them. But we will have to be a lot more careful."

Google researchers have been exploring the challenges presented by such a shift.

• "The 'agents' question really brings the philosophical question of value alignment to the fore," Iason Gabriel, a research scientist in the ethics research team at Google DeepMind, tells Axios' Alison Snyder.
• "It is also extremely important that the AI respects the needs and interests and well being of other people."

Zoom in: Customer service has emerged as ground zero for the debate over how much autonomy to give to AI-based systems.

• Sierra's Bavor and Taylor tells Axios that full automation will make the biggest difference to companies' bottom lines.
• That can boost productivity gains from simply making existing human agents 10-20% more efficient to independently handling 70% of cases, Bavor says.
• Taylor notes that Sierra's system doesn't just show an AI model 1,000 examples and have it infer the rules for customer service. Rather, he says, "Our agents are following the same instructions that a human agent would be given."
• He adds that the company uses multiple AI models — including one that acts as a supervisor to ensure other AI systems are performing as expected.

Sierra charges its customers only for each incident handled by its AI system. "We don't charge unless the AI fully resolves it," Taylor says.

• Bavor says early results are encouraging, with the system built for Weight Watchers able to resolve nearly 70% of customer inquiries while receiving 4.6 stars out of five for customer satisfaction.

The other side: Salesforce has largely focused on having "a human in the loop" with its early generative AI systems.

• "I feel like there's a reason why we're calling this generation of products copilots... and not autopilots," Paula Goldman, Salesforce's chief ethical and humane use officer, tells Axios.
• However, Goldman says the company is shifting its conceptualization to embrace a "human at the helm" approach — under which, for example, a person might not need to review all of the hundreds of similar e-mails crafted by an AI system.
• In many cases, though, Goldman says it still makes sense to have a human review each response. "When you're dealing with models where the results are unpredictable, where you could put in a prompt and get a bunch of different results, depending on the context, you need to have more controls in there," she says.

Flashback: Digital innovators have dreamed about autonomous agents for decades.

• Early visions like Apple's Knowledge Navigator in the 1980s and mobile-device pioneer General Magic's software in the 1990s failed because the networks and hardware of the time couldn't support them.
• But the world might be ready now.

What's next: Taylor says AI agents will expand beyond customer service and over time, become the dominant way that individuals interact with businesses. "It's going to be the digital version of the company," he says.

• "I have to imagine in five years it will be everything from commerce to even investor relations," Taylor tells Axios.

Yes, but: As businesses hand off more tasks to agents, their customers will, too — speeding us toward a world in which "my bot will talk to your bot."

• We could reap a bounty of saved time — and face a whole universe of new scams, cyberattacks and privacy violations.