Artificial intelligence (AI) agents are being built to do more than answer questions. They are expected to make phone calls, shop online, book appointments and complete other tasks without a person guiding every step. But when those tasks move into the real world, companies are still relying on humans to handle situations that AI systems cannot reliably complete on their own.
The issue has surfaced again. Meta is testing a "human concierge" for its personal AI agent Muse, with human contractors handling some phone calls that Muse is unable to complete on its own, according to internal company posts seen by Reuters. Muse's phone feature is designed to call US businesses for tasks such as booking appointments, checking product availability and getting contractor quotes.
The experiment highlights a problem that has appeared in other AI agent systems. Google's Duplex used human operators when its system could not complete certain phone tasks, while Perplexity has acknowledged human oversight for its shopping agent. The examples show the gap between an agent being able to perform a task in a controlled environment and being able to complete it when another person, a business or a website does not behave as expected.
When AI agents meet the real world
AI agents are designed to do more than generate text. They can browse websites, use applications, send messages, make bookings and, in some cases, complete purchases. The problem is that these tasks involve systems and people that the AI does not control.
A website can change its layout. A retailer can run out of stock. A customer-service employee can ask a question the agent was not prepared for. A business can refuse to deal with an automated caller.
Meta's Muse agent encountered such issues in testing. According to the report, one Meta employee told colleagues that an insurance company repeatedly hung up after recognising that the call was being made by AI. Meta's internal posts said the company enabled the human-concierge feature for half of its employees to address such problems. Under the test, Muse could hand a request to a trained human agent, who would place the call and complete it.
This is different from a chatbot failing to answer a question. An agent has been given a task, and the user expects the task to be completed. A failure therefore means the user may still have to intervene.
The test also raised privacy concerns among Meta employees. Internal posts cited by Reuters showed employees warning that sensitive information could be unintentionally shared with call-centre contractors. One employee wrote that the company was "one bug away from unnecessary information being leaked to human callers". Another employee reported that a contractor made a racist reference during a call to an internet and cable provider.
Meta's response to those concerns was more measured. Daniel Roberts, a Meta spokesperson, told Reuters that employee feedback on the test had been overwhelmingly positive. He said the purpose of the test was to gather feedback that would allow Meta to implement safety and privacy protections and improve the feature before releasing it publicly. Roberts also said Meta was working with merchants to improve the calling feature and would roll it out only when it was ready and with the proper disclosures.
While this was an internal test, a post cited by Reuters said a Meta vice-president reported that human callers had achieved a 95-98 per cent success rate in tests, compared with a lower success rate for AI-only calls.
Google had already encountered this problem
Google faced a similar problem when it began testing Duplex, an AI system designed to make phone calls for tasks such as restaurant reservations and appointments.
Google's research paper on Duplex said the system could complete the majority of its tasks autonomously. But it was also designed to recognise when it could not complete a task and signal a human operator. For new domains, experienced operators supervised the system in real time until it reached the required quality level.
The need for human intervention became visible when Duplex moved into real-world use. In 2019, Google told The New York Times that roughly a quarter of Duplex calls started with a live human voice. Of the calls that started with the AI, 15 per cent required human intervention, TechCrunch reported at the time.
Duplex was also deliberately limited to specific tasks. Google said its system worked within closed domains because it had to be trained deeply for those areas and could not handle general conversations.
Shopping creates another human checkpoint
Since Google's Duplex launch, AI agents have moved beyond making phone calls for narrow tasks. Newer agents can browse websites, interact with apps and complete transactions on behalf of users. That has also created more ways for an agent to get stuck when the information it sees does not match what is happening in the real world.
When Perplexity launched its shopping feature in the US in November 2024, it allowed users to search for products and place orders without visiting a retailer's website. The service also allowed users to store an address and credit card details for one-click checkout.
TechCrunch tested the system by asking it to purchase toothpaste. One attempt failed after three hours because the selected product was out of stock. A second purchase took eight hours to complete. The publication also reported that Perplexity's system involved human checkers.
Perplexity spokesperson Sara Platnick told TechCrunch there was "human oversight providing occasional support" to ensure transactions were completed on time and to avoid purchases of the wrong product. The company did not disclose how often humans intervened, how involved they were or whether they watched transactions in real time.
The problem in that case was not necessarily the AI's ability to identify toothpaste. The difficulty was completing the transaction when the information available to the agent did not match what was happening at the retailer.
TechCrunch found that Perplexity's system appeared to rely on information gathered from retailers' websites, which could create a gap between the product information shown in Perplexity and the retailer's actual inventory.
The human fallback also changes the privacy equation. An automated shopping agent may already have access to information such as a user's name, shipping address and payment details. If a person has to intervene, that information could potentially become visible to another human involved in completing the transaction.
Humans remain part of the system
Perplexity is not the only company to build human intervention into an AI agent. OpenAI's Operator, introduced last year, was designed to use a browser to perform tasks such as filling forms and ordering groceries. OpenAI said that when Operator encountered a challenge it could not resolve, it could hand control back to the user. Its safety framework also used human-in-the-loop safeguards for higher-risk actions.
These systems do not show the same type of human contractor intervention reported at Meta or Perplexity. But they show how companies are designing agents around human control, intervention and confirmation rather than assuming that AI can safely complete every task without supervision.
Why real-world tasks are harder for AI agents
The difficulty with AI agents is not only that they can make mistakes. They also have to operate in environments that are constantly changing and cannot be fully predicted in advance.
Demis Hassabis, co-founder of Google DeepMind, has pointed to this problem while discussing AI systems that interact with the physical world. Speaking on the Big Technology Podcast, Hassabis said models that try to understand physical environments are not perfectly accurate. Even if a model is accurate 99 per cent of the time, he said, small errors can compound when the system uses that model to plan many steps ahead.
That problem applies beyond robots. An AI agent making a phone call has to anticipate how another person will respond. A shopping agent has to work with websites, product listings, inventory and payment systems that can change while the task is underway. A travel agent may have to deal with availability, prices and booking rules that have changed since it started the task.
The agent therefore has to maintain an understanding of its environment while acting in it. If that understanding is wrong, the error can affect every subsequent step.
Research into physical AI points to the same problem. A 2026 survey of world models says systems operating in the real world have to make decisions with partial and noisy observations and deal with uncertainty about how an environment will change after an action.
The problem becomes harder as the number of possible environments grows. A robot operating in a controlled laboratory can encounter a limited set of objects and situations. An AI agent operating on the internet can encounter millions of websites, interfaces and human responses, many of which can change without warning.
The International AI Safety Report 2026 also notes that AI agents carry higher reliability risks because they can act autonomously and directly affect other systems or the physical world. It says current systems still struggle with complex tasks and that there is no combination of existing methods that guarantees the high reliability needed in critical applications.
 "Why human oversight remains crucial as AI agents become more autonomous")
Connect with us on WhatsApp