(Credit: Cole Kan/PCMag/Getty)
Love or hate them, AI chatbots are seriously impressive tools with which you can have lifelike conversations about any and every topic imaginable. AI agents sound even more futuristic: autonomous assistants that can carry out any task for you, from making reservations to ordering groceries. However, the unfortunate reality is that they just don't deliver on their core premise and come with hard-to-stomach privacy concerns. Here's what I found out based on my testing.
What Are AI Agents?
As mentioned, AI agents can independently do things for you. They can book flights, plan vacations, or otherwise perform most tasks you can achieve with a computer or mobile device. The idea is that they save you time and the trouble of doing things yourself.
The type of AI agents I’m referring to in this article, as well as the ones gaining popularity, take control of your apps, such as web browsers, or operate in virtualized instances to perform things on your behalf. For example, ChatGPT’s Agent exists within a dedicated virtual web browser window, while Perplexity’s Comet browser AI agent takes direct control of your browser. As I discuss later, both types often fail to fully carry out your instructions.
However, there isn’t a strict definition of an AI agent, so many companies brand different AI features as agents or agentic. One example is an AI customer service agent, which is basically a purpose-built AI chatbot that interacts with customers. Microsoft Copilot has AI agents, such as its Researcher agent, which essentially performs AI chatbot deep research. These aren't the kinds of AI agents I take issue with, and they more or less work as advertised.
How Do AI Agents Work, and Where Can You Find Them?
Large language models (LLMs) power AI agents just as they do AI chatbots and pretty much every mainstream AI product or service. Think of an LLM like a complex equation designed to take in prompts and spit out responses. With an AI chatbot, an LLM takes in a question and responds with an answer. But, with an AI agent, the LLM takes in an instruction and responds by following it. Keep in mind that AI agents are not conscious, nor do they possess intelligence.
(Credit: OpenAI/PCMag)AI agents, at the time of writing, primarily exist as AI chatbot features, such as ChatGPT’s Agent or Google’s Project Mariner, or as part of web browsers, such as ChatGPT’s Atlas, Microsoft’s Edge, Opera’s Neon, or Perplexity’s Comet. Many AI agents cost money, but some don't. For example, you need to sign up for ChatGPT’s $20-per-month Plus plan to use its AI agent, but you can use Microsoft’s Edge AI agent, called Actions, for free.
AI Agent vs. AI Chatbot: What's the Difference?
An AI chatbot can answer questions, generate images and videos, help you code, search the web, solve problems, and more. They are tools you can use to accomplish goals or conversation partners, not independent actors. As mentioned, AI agents go further by actually performing actions for you.
For example, you can ask ChatGPT for a recipe, and it will provide one. With ChatGPT Agent, you can ask it to not only find a recipe but also open your Instacart, add the ingredients to your cart, and order them.
The Big Problem With AI Agents: They Don't Do What They Say
The premise of AI agents sounds great in the same way that the idea of a robot that cleans your house does. The issue is that AI agents just don't deliver on their core promise. They struggle to do the virtual tasks they are designed to perform. Agents in web browsers, for example, often can't solve CAPTCHAs, fail to navigate sites, or routinely get stuck on certain steps, making them far more frustrating than convenient.
Even when they do work, they usually don't save you time, either. They almost always take longer than you would to do the same tasks. And they still require significant babysitting. You might have to click a pop-up ad that an agent can’t seem to close or input a password it needs to do something. If adding groceries to an online cart only takes you a few minutes, what good is an AI agent doing that in twice the time while you keep an eye on it?
AI agents are at least getting better. ChatGPT’s Agent, for example, is much faster and less error-prone today than Project Mariner was back when I first tested it less than a year ago. And I'm sure someone has found a specific use case for AI agents that works reliably, even if most people should still wait for broad improvements.
AI Agents Also Have Serious Privacy and Security Issues
Sharing your data with an AI agent is about as safe as sharing it with an AI chatbot, so there’s little difference in telling ChatGPT something as opposed to ChatGPT Agent. The bad news is that AI companies almost always collect significant amounts of data, often using it for model training. Always keep in mind that the information you share with an AI agent isn't at all private.
AI agents also bring about new kinds of security vulnerabilities, such as prompt injection attacks. Furthermore, some AI agent services, such as AI.com, put you on the hook for any actions an AI agent takes on your behalf without guaranteeing that said agent will always follow the law. Even if there's only a small chance you get in trouble for something an AI does on your behalf, is that really worth the risk?
AI Agents Might Be the Future, But the Future Isn't Now
Making the jump from a tool you can use to an end-to-end technology that actually performs tasks for you seems like an inevitable leap forward for the AI industry. And while the idea of AI agents is undeniably compelling, the technology just isn’t there yet, even if it is continuously improving. The future utility of AI agents (especially privacy-respecting ones) is worth looking forward to, but you should think carefully before sinking too much time or money into AI agents in the here and now.
Disclosure: Ziff Davis, PCMag's parent company, filed a lawsuit against OpenAI in April 2025, alleging it infringed Ziff Davis copyrights in training and operating its AI systems.