User trust isn't Meta Platforms' forte. Yet the company's AI future increasingly depends on it.

Last week, Meta rolled out Muse, Chief Executive Mark Zuckerberg's "personal AI agent built for everyone." Artificial-intelligence agents within Muse can book doctor's appointments, buy products, manage finances, send emails and do any number of other things on users' behalf.

In theory at least, this saves people time and organizes their lives with less effort. In Meta's telling, it also democratizes AI. Zuckerberg first described his vision for "personal superintelligence" last year and previewed Muse in an August essay where he spoke of "putting power in people's hands to pursue their own aspirations."

For Muse to work, though, it needs access to users' data-their emails, calendars, e-commerce sites and bank accounts. And that is where Meta's strategy seems destined to face hurdles.

People don't trust big tech companies much, but they are especially wary of Meta. An Oppenheimer & Co. survey of U.S. consumers found just 8% would trust Meta with their passwords, less than a third of the number who would trust Google.

"We see any questions or hesitancy from consumers around data privacy and security as likely the biggest potential headwind for broad-based adoption of consumer AI agents," Goldman Sachs analyst Eric Sheridan said in a note after Muse's launch.

Given that backdrop-and a recent groundswell of concern about AI agents going rogue-Meta can ill afford any user-privacy hiccups in Muse's rollout.

Meta's announcement of the tool took pains to underline how rock-solid it was. The company described it as "safe, secure, private," adding that Muse would run on a digitally walled-off virtual machine that no one else's agents can access.

Zuckerberg said in a post on X Tuesday that the company delayed Muse's release for several months "to focus on safety and security." He dismissed the idea of slowing the pace of frontier AI development in concert with Anthropic, OpenAI and SpaceX, suggesting there were natural incentives for it to keep AI agents in line.

It isn't yet clear whether users are buying this. But getting a large segment of Meta's 3.6 billion users to sign up seems daunting.

Only six years ago, after all, Meta was paying a record $5 billion fine to the Federal Trade Commission for user-privacy breaches in the Cambridge Analytica scandal. The following year, 533 million users' information was leaked. Transfers of European users' data to U.S. servers led to a record $1.3 billion fine under the Continent's stringent data-privacy regulations in 2023. User-safety litigation, including a recent $18 billion settlement of a case about harm to teens in the U.S., hasn't helped its reputation.

Meta does have a few advantages. It has unparalleled social-media data and can integrate agents with its Marketplace e-commerce platform. It is also the first mover in the personal AI assistant game: No other big AI player yet offers anything similar.

Yet being fast out of the gate could come back to bite Meta, too.

If agents make mistakes as Meta works out the kinks, the risk is high that users will lose trust and flee. The stakes with a personal agent are elevated because agents are empowered to buy products, set appointments and book travel by themselves-all of which could cost users money if the agent errs.

And Meta acknowledges that there will be glitches: When users register for Muse, a warning pops up saying the agent "may make mistakes or take unexpected actions, so review its work." Even if a user gives the agent unclear instructions, having it buy a $1,000 item when the user wanted a $10 item might not sit well.

Competitors could also jump in if Meta's Muse idea takes off. Google is in an especially good position to capitalize on agents, given that it offers email, calendar and cloud-storage apps that could feed them, paired with a cutting-edge AI model.

And even if Meta can overcome its privacy record and Muse becomes a success, its financial impact isn't likely to be huge.

Light usage of Muse is free, but the company charges $20 or $100 a month for heavier users. To increase Meta's earnings per share by roughly 20%, Oppenheimer estimated that the company would have to gather 115 million Muse subscribers paying $20 a month-about double the number of paying subscribers of OpenAI's ChatGPT.

That would be an ambitious goal even without Meta's trust issues, given that many of the people willing to pay for AI are already customers of OpenAI, Google and others.

But Meta needs it to work. The company is spending big on AI, including an expected $137 billion of capital expenditures this year, fueled by a lucrative and fast-growing ad business.

Beyond its successful use of AI in ads, the company is trying to generate returns on that investment with subscriptions for extra features on Facebook, Instagram and WhatsApp. It is selling outsiders access to its most advanced AI models, and is making a potential cloud-computing push. Muse is another major cog in the return-on-investment engine.

Investors have welcomed the Muse rollout: Meta's stock is up about 10% since its launch.

But Zuckerberg's move-fast-and-break-things mantra has broken a lot of things over the years, including his users' trust. That might cost Meta dearly in the AI era.