AI Agents Still Need Humans Behind the Scenes to Make Phone Calls

A new report on consumer AI agents from Meta, OpenAI and startups like Dazzle and Duckbill shows real people are quietly handling the tasks the software still can't.

AI agents still need humans behind the scenes to make the phone calls the software can't handle. That's the real story in a new CNBC report out today: the consumer AI agents from Meta, OpenAI and a wave of startups can browse the web and surface good suggestions, but when a task needs a phone call, a scheduled appointment, or a CAPTCHA solved, a real person is quietly doing that part.

The story names specific players. Marissa Mayer, the former Yahoo CEO who now runs agent startup Dazzle AI, told CNBC that today's top AI models are good at finding options online, like home renovation services, but can fumble the follow-through, like actually booking the job. Meghan Joyce, CEO of agent startup Duckbill, said the gap shows up most on tasks that require a phone call, a scheduled doctor's appointment, or booking something like a house cleaning. Shivani Poddar, CEO of personal agent startup Wajo AI, said her company also routes phone calls and CAPTCHA tests, the tests sites use to block bots, to humans.

When the AI Can't Tell a Window from a Tabletop

Mayer described a Dazzle user who had the AI handling three home projects at once: a roof replacement, a broken window, and the purchase of a glass tabletop. The AI agent mixed up the two kinds of glass and started trying to figure out if the tabletop could fit in the window frame. Dazzle's system flagged the mix-up and handed it to a person, who sorted it out within a few seconds. Mayer's point: an AI agent stuck in that kind of loop just keeps burning computing power without getting closer to an answer.

Duckbill built an API so other companies' AI agents can hand off the tasks most likely to go wrong. Joyce compared the model to Uber, where she worked before: break a task into small enough pieces that an average adult can do it reliably every time. Poddar put it plainly: AI plus humans, working together, is what actually gets a task finished. If agents can't close the loop, she said, people stop using them.

A New Kind of Gig Work Is Forming Behind the Chatbot

Here's what this means for jobs: a new layer of gig work is forming underneath the AI agent boom, and it looks a lot like the last one. Duckbill runs full background checks on the 1099 contractors who take these jobs and compares them to Uber and DoorDash drivers, people in the United States looking for additional income. It also draws on contract workers from businesses that already supply contact center labor. Dazzle screens for the same traits you'd want in a personal assistant: discreet, good judgment, and able to forget what they heard once the call ends.

This is a real job category, even if nobody is calling it one yet. It's not building the AI. It's being the fallback when the AI can't finish what it started, on someone else's schedule, paid by the task.

What This Looks Like From the Insurance Side

I run an insurance agency, so here's the part of this story most coverage will skip. When you take 1099 work like this, you are not an employee, and that changes what covers you if something goes wrong. A platform's background check tells a customer you're trustworthy. It does not mean you're carrying workers' compensation, because 1099 contractors generally aren't eligible for it the way employees are. General liability coverage isn't automatic just because you're doing gig work either; it depends on what you actually agreed to and what, if anything, you're carrying on your own.

In insurance, gaps like this almost always surface at the worst possible time, after something has already gone wrong, not before. Anyone picking up this kind of work should know, going in, what the platform covers and what it doesn't, the same way you'd ask before driving for a rideshare company.

What a Transitionary Period Means for Your Business and Your Paycheck

Mayer called this a transitionary period, her word for the stretch between agents that mostly work and agents that fully work on their own. That has two sides. If you run a business and you're buying or building AI agent tools, budget for the humans behind the curtain, because right now they're part of the real cost, not a rounding error. Pure software margins on an AI does it all pitch aren't accurate yet for anything that needs a phone call or real-world judgment.

If you're one of the people doing that human work, the income is real but it's tied to how bad the AI still is. Meta's own employees raised privacy concerns after the company tested using human contractors to handle phone calls for its Muse agent; a Meta spokesperson called that kind of testing core to the product development process, and Meta says it currently doesn't use humans to complete Muse requests. That tells you where this is headed: as the models get better, some of this work gets automated away, the same pattern that's hit other gig categories.

What to Watch and What to Do

A few practical things to take from this.

• Looking for extra income: size up an AI-agent human-in-the-loop job the same way you'd size up any gig platform, know what's background-checked, what's recorded, and what you're actually liable for.
• Running a business: if you're buying an AI agent product, ask directly whether a human ever steps in on your customers' behalf, and whether that's disclosed to them.
• Doing this work as a 1099 contractor: don't assume the platform's coverage protects you the way an employer's would, ask what happens if a call or task goes wrong.
• Watching the space: the companies calling this a transitionary period are telling you themselves that today's setup isn't permanent, don't build a business model around today's gap.