AI Agents Still Rely on Humans to Complete Complex Tasks

Artificial Intelligence agents are increasingly being marketed as digital assistants capable of handling everyday tasks, from searching the internet for products to arranging services and making appointments. However, despite advances in foundation models developed by companies such as OpenAI and Meta, some AI-powered services still rely on human workers to complete tasks that software cannot reliably handle.

Startups including Dazzle AI, Duckbill and Wajo AI are adopting a human-in-the-loop approach, combining artificial intelligence with human support to deliver services that fully automated systems may struggle to provide. The approach highlights a current limitation of AI agents: while they can gather information and suggest actions, executing complex tasks across websites, phone systems and real-world services remains challenging.

Why AI Agents Still Need Human Assistance

AI agents are designed to move beyond conventional chatbots by taking actions on behalf of users. They can browse websites, compare options and assist with tasks such as finding products or identifying service providers. However, problems often emerge when an assignment requires multiple steps, communication with another person or decisions based on changing real-world circumstances.

Tasks such as negotiating a cable bill over the phone, scheduling a doctor’s appointment or arranging a home cleaning service can involve unexpected questions and decisions. These situations require an ability to understand context, respond appropriately and adapt when the process does not follow a predictable path.

Marissa Mayer, former Yahoo chief executive and head of Dazzle AI, said current advanced AI models can identify suitable home renovation services but may struggle when asked to book and schedule a more complicated project, such as a rooftop upgrade.

Dazzle uses AI to generate recommendations and identify potential service providers, while human contractors help verify information and complete tasks when the software reaches its limits.

Dazzle AI Uses Human Contractors to Complete Complex Tasks

Mayer explained that Dazzle relies on advanced AI models to generate suggestions and candidate lists but uses human oversight to ensure that the service meets its standards. The company’s approach is based on recognising that a technically plausible recommendation does not always translate into a successfully completed task.

She described an incident involving a Dazzle user who asked the agent to coordinate several home renovation projects, including replacing a roof, repairing a broken window and purchasing a glass tabletop.

The AI system became confused between the glass required for the window and the material intended for the tabletop. According to Mayer, the agent began trying to determine whether the tabletop could fit into the window.

The problem was quickly resolved after Dazzle’s system flagged the task and transferred it to a human worker. Mayer said a person was able to understand the situation and find a solution within seconds, whereas the AI agent could have spent considerable computing resources without reaching a useful outcome.

The example illustrates why human intervention remains valuable when an AI system encounters ambiguity or needs practical judgment.

Duckbill Builds an API to Connect AI Agents With Human Workers

Duckbill, led by chief executive Meghan Joyce, is developing infrastructure that allows companies building personal AI agents to delegate difficult tasks to human contractors.

The startup recently introduced an application programming interface (API) designed to connect AI platforms with its network of human workers. Companies can use the system to transfer tasks that their agents are likely to struggle with, allowing people to complete specific parts of a larger assignment.

Joyce compared the operating model to Uber, where complex services are broken down into smaller tasks that individual workers can perform. Duckbill applies a similar approach by dividing assignments into manageable units that contractors can complete consistently.

The company also tracks the progress and outcomes of delegated tasks, allowing AI developers to incorporate human assistance into their services without handling every operational step themselves.

Joyce said that human judgment remains important for tasks involving real-world decisions and interactions. She believes consumer-facing AI assistants will need access to people if they are expected to complete assignments reliably.

Wajo AI Combines Artificial Intelligence With Human Support

Wajo AI is also incorporating human workers into its personal assistant services. Chief executive Shivani Poddar said people are used for tasks such as making phone calls and completing CAPTCHA tests, which websites use to distinguish human visitors from automated systems.

Poddar believes collaboration between AI systems and people is necessary to deliver practical results. If digital assistants repeatedly fail to complete the tasks users assign to them, consumers may lose confidence in the Technology.

The approach reflects a broader distinction between an AI agent’s ability to understand instructions and its ability to carry them out. A system may successfully interpret what a user wants but still require human assistance to navigate a complicated process.

Human Involvement Raises Privacy and Data Protection Questions

Using human contractors to complete personal tasks introduces another important consideration: user privacy. AI assistants can receive sensitive information about household purchases, healthcare appointments, financial arrangements and personal schedules. When a task is transferred to a human worker, that information may become accessible beyond the automated system.

Users may therefore want to know when a person is handling their requests, what information is shared and how that information is protected. Transparency is particularly important when companies describe their products as AI-powered personal assistants without clearly explaining the role of human contractors.

Meta employees previously raised privacy concerns about the company’s internal testing of human contractors making phone calls on behalf of its Muse AI agent. A Meta spokesperson said at the time that such testing was part of the product development process. Meta currently does not use humans to complete users’ Muse requests.

Duckbill said it conducts background checks on its 1099 contractors, including people seeking additional income through flexible work. The company also works with businesses that typically provide contact centre services.

Mayer said Dazzle screens its contractors and looks for qualities associated with personal assistants, including discretion, sound judgment and the ability to handle private information responsibly.

AI Companies Face Challenges With Real-World Tasks

The growing use of human support highlights the gap between AI demonstrations and everyday execution. An agent may be able to search for a product, compare prices or identify a suitable service provider, but completing the entire process can involve several unpredictable steps.

Phone calls are particularly difficult because conversations can change direction depending on the responses received. Scheduling appointments may also require dealing with availability changes, follow-up questions and service-specific conditions.

These challenges create operational costs for companies developing AI assistants. They must decide which tasks can be automated, when human intervention is necessary and how to transfer work without disrupting the user experience.

Human assistance can improve task completion, but it also means that some services marketed as automated are not entirely independent of human labour. The balance between Automation and manual support will remain an important consideration as these products develop.

Human-in-the-Loop Model Could Shape the Future of AI Agents

Mayer described the current stage of AI development as a transition period, during which foundation models are improving but cannot yet perform every complex web-based task or phone interaction as efficiently as people.

For startups such as Dazzle AI, Duckbill and Wajo AI, human involvement is being presented as a feature of the service rather than a temporary workaround. The model allows companies to use AI for information gathering and routine processing while relying on people for situations that require judgment, communication and adaptability.

As AI agents become more widely available, their success may depend not only on the capabilities of the underlying models but also on how effectively companies manage tasks that automation cannot complete. For now, the combination of AI systems and human workers is helping bridge the gap between digital assistance and real-world execution.