AI Agents: Alluring Promises, but Costly Traps to Avoid

The Alluring Promises of AI Agents

AI agents, such as Claude Cowork, are increasingly present in technological discussions, promising to automate various tasks like email management, file sorting, or report generation. These tools boast the ability to remotely control computers or synthesize information from the web, all without human intervention.

Demonstration videos are proliferating on LinkedIn, YouTube, and in tech newsletters, showcasing how these agents can autonomously handle complex tasks. The effect is impressive, and the temptation to delegate one's repetitive tasks to these agents is strong. However, behind these promises lie significant challenges that are not always evident during demonstrations.

Understanding Automation Before Diving In

Before rushing to use an AI agent, it is essential to understand the basics of automation. Simple automations rely on conditional instructions, such as triggering an action upon receiving an email or populating a spreadsheet from a filled-out form. Tools like Zapier or Make allow users to create these automations without advanced technical skills.

An AI agent goes further by making decisions at each step, which makes it powerful but also unpredictable and potentially costly. Deploying an AI agent without ever having built a simple automation is like driving on a highway without having learned how to shift gears. It’s not a matter of technical level but of understanding the mechanism.

The Hidden Cost of AI Agents

AI agent demonstrations do not highlight the cost of usage. Each action performed by an agent consumes tokens, and costs can quickly add up. Users of Claude Cowork have found that their credits deplete rapidly, often due to the complexity of the tasks assigned to the agent.

What demonstrations never show is the bill. An AI agent processes a significant amount of information to decide what to do at each step. Each action—reading a file, analyzing a page, writing a line, checking a result—consumes tokens, the basic unit that determines the cost of using the model. Because agents work step by step, chaining dozens of micro-decisions to accomplish a seemingly simple task, costs accumulate quickly—often far beyond what was anticipated.

Latency: An Underestimated Obstacle

Another major issue is latency. AI agents take time to execute tasks, as each step requires a decision based on prior analysis. For example, an agent may take fifteen minutes to complete an accounting summary in Google Sheets, whereas a human would be faster.

JDN documented this slowness in its own tests. It’s not an anomaly—it’s the normal behavior of an agent navigating through a web application by screen vision, capturing the screen at each step, inferring a structure, and acting cautiously.

A telling example is booking a plane ticket. The task seems simple, but in practice, the agent must open a browser, load the site, analyze the page, fill in fields one by one, check each result, manage pop-ups, and navigate through payment steps. What you could do in five minutes may take it twenty-five—risking failure at every link.

When the AI Agent Is Not the Best Solution

Sometimes it is wiser to use a simple automation rather than an AI agent. For example, to automatically integrate data from an Excel file into a CRM, a classic automation is often faster and less costly.

AI agents are better suited for tasks involving variability and complex decisions, but they should not be considered universal solutions. Sometimes, a simple automation suffices—and deploying an agent is not only unnecessary but counterproductive. Take a concrete case. You receive an Excel file from your sales team every week. You want the data to be automatically integrated into your CRM. An AI agent can technically do that—by reading the file, analyzing the columns, deciding how to map the data, and making the necessary calls. But a classic automation, set up once in twenty minutes, does exactly the same thing—faster, without token costs, without latency, and without the risk of reasoning errors.

The Importance of Understanding Before Deployment

Leaders must understand how an AI agent interacts with existing systems before adopting it. This avoids signing a "blank check" in terms of time, resources, and credibility.

Understanding the fundamental mechanisms of AI agents allows for optimizing their use and avoiding the costly pitfalls of misleading demonstrations. A leader who asks their team to deploy an AI agent without understanding how it interacts with existing systems is not managing a project. They are signing a blank check—in time, computational resources, and internal credibility when the project stalls. It’s not a matter of technical skill. It’s a matter of understanding the fundamental mechanisms: how an agent perceives its environment, through which interface it acts, and what each action truly costs. These distinctions are not learned by watching demonstrations—because demonstrations never show the conditions under which they were prepared.

Understanding the system means knowing that a desktop agent excels with local files but is not designed to manipulate a live web application. That Zapier or Make rely on stable connections where screen navigation fumbles pixel by pixel. That a simple script inserts a line into a spreadsheet without consuming a single token. These distinctions exist. They are accessible. But they require taking the time to understand what you are using—before starting, not after three hours of walls.