Why are AI agents handing tasks to humans?
On 31 October 2026, CNBC reported that several personal AI assistants rely on people for work their software cannot finish. The companies named included Dazzle, founded by former Yahoo chief executive Marissa Mayer, along with Duckbill and Wajo AI. The report described assistants that can plan a task and then pass it to a person.
The tasks are ordinary: negotiating a cable bill, booking a dentist, waiting on hold, solving a CAPTCHA. They are also the tasks that AI agents are marketed as handling alone. The gap between the demo and the finished job is where the human enters.
The handoff is not always concealed. Some startups now sell human backup as a feature. Meta's internal test of the same approach showed why disclosure and privacy are the difficult parts. It also showed how quickly a product claim can become a trust problem.
For buyers, the appeal is completion. An agent that cannot finish a booking is worse than a form. For users, the risk is that a private request becomes a stranger's task. Both matter as assistants move from demos into daily use.
Which AI agents still use human workers?
Dazzle, which TechCrunch profiled on 29 September 2026, raised an $8 million seed round in December 2025. The assistant builds context from a user's camera roll rather than email or calendar history. It turns photos into calendar entries and local recommendations, and users reach it through an app or text message.
Mayer told CNBC that Dazzle sends some jobs to human contractors when the model's suggestions need checking. She described a user who asked for help with a roof replacement, a broken window and a glass tabletop. The assistant confused window glass with tabletop glass and tried to work out whether the tabletop would fit in the window.
Duckbill takes a different route. Chief executive Meghan Joyce, a former Uber executive, said the company breaks tasks into small pieces for human workers. Her startup recently released an API that lets other companies' agents hand work to Duckbill's contractors. The company calls the service a human API.
Duckbill's site describes more than 200 trained, vetted operators who make calls and chase refunds. Consumer plans run from $49 a month for 20 credits to $350 a month for 150 credits. The operators are 1099 contractors, Joyce told CNBC, and the company compares the model to Uber and DoorDash.
The service covers four broad areas: healthcare navigation, travel recovery, customer service disputes and physical errands. Examples include prior authorization calls, refund negotiations, subscription cancellations and package pickups. Duckbill also offers an MCP integration for agents built on Claude or ChatGPT.
Wajo AI, led by chief executive Shivani Poddar, sells an agent called Fo. Its website says Fo calls, emails and coordinates with people, and that it "works with trained executive assistants" when a task needs a human touch. CNBC reported that those people handle phone calls and even CAPTCHA tests.
Wajo's site says Fo can pay with single-use cards, send email from its own address and sign in to websites without seeing the user's password. Those features describe how the software limits data exposure. They do not remove the fact that a person may still join the call.
The common thread is not the model. It is the last mile: a phone tree, a hold queue, a person who can improvise when the script fails. The companies are selling access to that person as much as they are selling software.
What happened with Meta's Muse human callers?
Meta's Muse offered the clearest test. In September 2026, the company expanded a beta that let Muse call US businesses for reservations and appointments. 404 Media reported that the calls were sometimes placed by people in call centers.
The public signal came from Meta staff. Ryan Fox, a principal engineer on Muse, wrote on 16 September that the company had expanded the beta for outbound calls to US businesses. Chief AI officer Alexandr Wang also posted about the phone beta.
An internal post seen by 404 Media said Meta had "added a human agent layer for calls" and that "Muse is now able to hand requests to a trained human agent." The post said Muse could handle the conversation, complete the request and report back with a transcript and summary.
The same post said Muse "no longer works alone" and can pass a request to a trained agent who places the call and works it through. Employees were asked not to share the feature outside the company. The post described the arrangement as confidential and pre-launch.
Staff questioned both privacy and perception. "This has potential for so much negative PR," one employee wrote. Another called the plan a "bad bad idea." A tester was not told a human had made the call until afterward, according to 404 Media.
The internal criticism focused on consent. Users would share medical or financial details with what they believed was software. Handing those details to a contractor changes the privacy calculation, even if the contractor has received training.
Meta told 404 Media that internal testing is "core to the product development process" and said it would "only roll it out when it's ready and with the proper disclosures." The company also said it was working with merchants to improve the potential calling feature.
In its 31 October report, CNBC said Meta does not use humans to complete Muse requests. The report did not describe a human layer for Muse. What remains is the internal record of how the feature was built and how employees reacted to it.
The case matters because Meta operates at a scale the startups do not. A default-on human layer at Meta would reach millions of users before most of them read a policy page. One employee warned that launching it as a default would damage the goodwill from early adopters.
How does the human handoff work in practice?
The handoff can look like an API call. Duckbill's developer page shows an agent sending an intent such as rebooking a flight, with constraints attached. The service then routes the step to a person who waits on hold, negotiates and reports a result.
That design turns human labor into a component other agents can call. It also means the end user may never choose the handoff. The agent developer decides when to use it, and the user sees only the outcome. Duckbill's pitch to developers is that abandoned tasks cost them users they already paid to acquire.
The API model also changes who is responsible for a failed task. The agent vendor sets the rules, Duckbill supplies the worker, and the user sees a single answer. Contracts and quality scores are meant to settle what happens when the handoff goes wrong, and users rarely see those terms.
Vetting varies. Duckbill says its operators are background-checked and based in the United States. It also says operators are quality-scored on every task. Mayer said Dazzle screens contractors for discretion and judgment, comparing the search to hiring a personal assistant.
"We're not even trying to hide it, we think the human in the loop is a benefit," Mayer told CNBC. Joyce compared Duckbill's model to Uber, where tasks are broken into pieces that workers can accept. She said the company uses 1099 contractors and also draws on contact center businesses.
Duckbill publishes performance claims. It says its human-backed system completes 96% of multi-step tasks end to end, against less than 0.01% for an agent alone after 20 steps. Those figures are the company's own, not an independent audit, and the comparison assumes a long chain of steps.
The pricing hints at the economics. Human time is expensive, so the service bundles credits rather than promising unlimited work. A task that needs 40 minutes on hold has a real cost that an API call does not. That cost shapes which tasks get a person and which are left to software.
Why can't AI agents finish these tasks alone?
Phone calls remain a weak point. Automated menus, hold queues and verification questions demand patience and improvisation. CAPTCHAs are designed to block software outright, which is why Wajo's use of people to solve them is notable.
CNBC reported that frontier models from Meta and OpenAI still cannot reliably take actions across the web, especially calls. The gap is practical rather than philosophical: the agent has to finish the transaction, not just describe it.
Agents also hit login walls, payment forms and identity checks. Those steps often require a code sent to a phone, a card held by a person or a judgment about whether a charge is legitimate. Joyce said humans still trump AI when a task needs real-world judgment.
Mayer's glass example shows the failure mode. The agent confused window glass with tabletop glass and tried to work out whether the tabletop would fit in the window. A person resolved it in seconds, she said.
Duckbill argues that errors compound across long tasks, which is why it inserts people where agents stall. That argument favors human backup, but it also describes the limits of the current products. It is a sales pitch built on the failures of the software.
The gap may narrow. Agents are improving at browsing, forms and scheduling, and some tasks now run without a person. The honest position is that completion rates vary by task, and the public numbers come from vendors with an interest in the answer.
Some users may prefer a person for sensitive calls. The companies lean on that preference, presenting human help as a service upgrade rather than an admission. That framing works best when the handoff is visible and worst when it is discovered after the fact.
What are the privacy and disclosure risks?
The stakes are higher than they look. People ask assistants to dispute insurance claims, book medical appointments and negotiate bills. Those requests carry names, dates, account numbers and health details. A call that sounds routine can expose information that a user would not post publicly.
When a person takes over, that information moves to a contractor. Meta's internal reviewers worried about exactly this. One noted that the training contractors received was "not a security mechanism," according to 404 Media.
Disclosure is inconsistent. Wajo and Duckbill describe human involvement on their public sites. Dazzle's Mayer said the company does not hide it. The Meta test is the outlier in the reporting, with at least one tester learning afterward that a human had made the call.
Background checks address some risks but not all. A check verifies identity and history, not how someone handles a caller's medical details or whether they keep a transcript. The reporting does not establish retention rules for these handoffs.
The contractor pool is also fragmented. Duckbill draws on 1099 workers and contact center businesses, Joyce said. That mix can make oversight harder than a single call center with one set of rules.
The companies have designed some privacy controls. Wajo says it uses single-use payment cards, sends email from its own address and never stores website passwords. Dazzle says it discards personal information it flags as sensitive, according to TechCrunch. Those controls limit what the software holds, not what a human hears.
The pattern has a name. Mary Gray and Siddharth Suri called it ghost work: human labor that customers mistake for automation. Sarah Roberts documented a related problem in content moderation, where people absorb sensitive material that platforms keep out of view.
For users, the practical question is consent. Did they agree to a person hearing the request, and would they have phrased it differently if they knew? Those questions apply to every assistant that offers a human fallback.
What should users and businesses watch next?
Watch Meta's next move. The company has said it will roll the feature out only with proper disclosures. Whether the label appears on every call, or only in a policy document, will set a benchmark for rivals. A silent handoff would make the earlier internal warnings look prescient.
Watch Duckbill's partners. Its API is aimed at companies that build agents, not at consumers directly. Adoption there could spread the handoff without users ever seeing the brand. The same applies to Dazzle's developer offerings and any similar service that sells completion as a utility.
Watch the pricing. If human help is bundled into premium plans, the market may split into cheap agents that fail and expensive ones that finish. That would make completion a paid feature rather than a baseline. It would also put pressure on companies that promise autonomy at a low price.
For buyers, the questions are simple. Does the assistant say when a person steps in? Who sees the transcript? How long is it kept? What happens when a task involves health or money? A vendor that cannot answer those questions is asking for trust it has not earned.
For users, the safest assumption is that a human may be listening. The handoff can make an agent more reliable. It also turns a private request into someone else's work, and that changes the deal even when the result is good.