The promises of AI agents are everywhere. Before you get started, three concrete problems that the demonstrations do not reveal – and which can be very expensive.
The videos are multiplying. On LinkedIn, YouTube, in tech newsletters: AI agents do everything for you. Claude Cowork is the most visible example at the moment — and the promises are attractive.
Manage your emails. Sort your files. Generate a full report from a voice recording. Fill out a spreadsheet. Control your computer from your phone while you are in a meeting. Search for information on the web, synthesize it, format it. All this, autonomously, without human intervention.
The effect is real. Power too. And the reflex that follows is natural: identify your own repetitive task and try to automate it.
This is exactly where things get complicated.
Before the agent, understand what automation is
Deploying an AI agent without having ever built simple automation is like driving on the highway without having learned how to shift gears. It’s not a question of technical level — it’s a question of understanding the mechanism.
Automation, in its simplest form, is a series of conditional instructions: if this happens, then do that. A received email triggers an action. A file dropped into a folder generates a notification. A completed form feeds a spreadsheet. These basic mechanisms are accessible to any non-technical professional, with tools like Zapier or Make. And building them, even just once, profoundly changes the way we think about what an agent can — or cannot — do.
Because an AI agent is an automation that reasons. He makes decisions at each stage, chooses the path, adapts his actions to the context. This is what makes it powerful. It’s also what makes it unpredictable — and expensive — when you don’t understand what’s going on under the hood.
First problem: the spiraling cost
What the 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, parsing a page, writing a line, checking a result — consumes tokens, the basic unit that determines the cost of using the model. And because agents work step by step, stringing together dozens of micro-decisions to accomplish a seemingly simple task, costs add up quickly — often far beyond what was anticipated.
It’s a bit like a personal assistant that charges for every thought. The more complex tasks you give him, the higher the rating goes up. Claude Cowork users have seen their credits run out in a few hours, on workflows that they thought were innocuous — simply because they had not measured the real volume of processing involved.
Before deploying an agent, the right question is not “does it work?” » — it’s “how many actions will he carry out, and at what unit cost?” »
Second problem: invisible latency
The second problem is less visible, but just as concrete: the agents are slow.
Not because the model is limited — but because the very nature of sequential reasoning takes time. Each step awaits the previous one. Each decision is preceded by an analysis. Imagine a team of assistants passing notes to each other for every small decision: even if each note only takes a few seconds, the sequence can turn a two-minute task into a twenty-minute process.
The JDN documented this in its own tests: in the case of the balance sheet in Google Sheets, the agent spent fifteen minutes where a human would have been faster. This is not an anomaly — it is the normal behavior of an agent navigating through screen vision in an unconnected web application, capturing the screen at each step, inferring structure, acting cautiously.
Another telling example: asking an agent to book a plane ticket. The task seems simple. In practice, the agent must open a browser, load the site, analyze the page, fill in the fields one by one, check each result, manage pop-ups, navigate between payment steps. What you would do in five minutes might take twenty-five—with the risk of failure at every link.
Latency is not a bug. It is a structural characteristic that must be integrated before deciding what to entrust to an agent.
Third problem: the agent is not always the right tool
This is perhaps the most underestimated problem: Sometimes simple automation is enough — and deploying an agent is not only unnecessary, but counterproductive.
Let’s take a concrete case. You receive an Excel file from your sales team every week. You want data to be automatically integrated into your CRM. An AI agent can technically do this — reading the file, analyzing the columns, deciding how to map the data, making the necessary calls. But a classic automation, configured once in twenty minutes, does exactly the same thing — faster, without token costs, without latency, without risk of reasoning errors.
The rule is simple: when the task is repetitive, predictable and well-defined, traditional automation is almost always more effective. The agent brings value where there is variability, ambiguity, decisions to be made. Mistaking it for a universal tool is like using a jackhammer to drive a nail.
What changes when you understand before deploying
A manager who asks his team to deploy an AI agent without understanding how it interacts with existing systems is not leading a project. He signs a blank check — in time, in computing resources, in internal credibility when the project sticks.
It’s not a question of technical skill. It is a question of understanding the fundamental mechanisms: how an agent perceives its environment, through which interface it acts, what each action really costs it. Distinctions that cannot be learned by watching demonstrations — because demonstrations never show the conditions in which they were prepared.
Understanding the system means knowing that a desktop agent excels on local files but is not designed to handle 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 row into a spreadsheet without consuming a single token.
These distinctions exist. They are accessible. But they assume you’ve taken the time to understand what you’re using — before you start, not after three hours of wall-to-wall training.
The blank check has an alternative. Understand what you are signing before you sign it.
