Behind Some AI Agents Is Just a Human: Dazzle, Duckbill and Wajo Patch People In Where Models Still Fumble

The pitch for personal AI agents — Meta's Muse, OpenAI's Dots and a growing list of rivals — is software that does things for you: scout the web, book the appointment, make the call. But a quieter truth is emerging underneath the demos, CNBC reported on October 31: for many consumer agents, the bridge between what the model promises and what actually gets done is a human contractor working in the background.

Marissa Mayer, the former Yahoo chief who now runs agent startup Dazzle AI, says current top-tier models are excellent at narrowing the world down. "We're relying on the frontier models to give us good suggestions and a good candidate list, but we're checking that over to make sure that we're really delivering the type of service that we can stand behind," Mayer said, describing how Dazzle funnels certain tasks — especially anything involving a phone call or a real-world booking — to human contractors.

Her favorite failure case is a super user who asked Dazzle to handle three home renovation projects at once: a roof replacement, a broken window and the purchase of a glass tabletop. The agent confused window glass with tabletop glass and, in Mayer's telling, spent its effort "trying to figure out if he could make the tabletop fit in the window." Dazzle's systems flagged the task and handed it to a person. "Within a few seconds, the human had things sorted out. Whereas an agent would have burned a ridiculous number of tokens," she said. "Or they may never have gotten to a good solution."

Duckbill is building a business on exactly that gap. CEO Meghan Joyce, a former Uber executive, says the roadblocks that trip up agents typically involve phone calls, scheduling a doctor's appointment or reserving a cleaning service — tasks that need real-world judgment. Duckbill recently debuted an API that lets companies building personal agents hand fumbled tasks to human contract workers, using a model "not dissimilar from Uber, where we boil tasks down into atomic bite-sized pieces." She says the company runs full background checks on its US-based 1099 contractors and also draws on contact-center labor businesses.

Wajo AI goes further down the same road. CEO Shivani Poddar says her startup routes phone calls — and even CAPTCHA tests designed to keep bots off websites — to human workers when the software cannot pass. "I think AI plus humans, collaborative AI, is what you need to actually get the job done," Poddar said, with a blunt logic: users will simply stop using agents that cannot finish tasks.

The pattern has already caused friction at Meta. Employees previously raised privacy concerns about an internal test that used human contractors to handle phone calls on behalf of the Muse agent, according to 404 Media; a Meta spokesperson said at the time that this kind of testing is "core to the product development process." Meta currently does not use humans to complete users' Muse requests — which makes the startup approach a live question about where the industry's borders actually sit.

Mayer is unbothered by the optics. "We're not even trying to hide it, we think the human in the loop is a benefit," she said, calling the market a "transitionary period" until foundation models can handle involved web tasks and phone calls as reliably as people. The precedent is familiar: ride-hailing and robotaxi services leaned on remote human operators long before autonomy was real. Agent startups may be entering the same phase, where the demo is AI but the last mile is still a person with a headset.