An artificial intelligence agent is not a digital employee: it is probabilistic software with access to tools and permissions. The difference is not semantic. An employee has judgment, accumulated context, legal responsibility, and the ability to say, "This doesn't add up, I'll ask." An agent has a goal, a set of available actions, and a probability of being right.
The employee metaphor has become prevalent because it sells. And it sells because it allows you to compare the price of a license with the cost of a payroll, which is the most favorable comparison you can make. The problem arises when this metaphor moves from the sales pitch to the system design, because then you design as if there were judgment where there is only probability.
A chatbot responds. An agent acts. That's the whole difference, and it's huge.
When a model generates text, the worst possible outcome is an incorrect response that someone reads. When a model can perform actions—create an order, issue a refund, modify a record, send an email to a customer—the worst possible outcome is an incorrect action that has already occurred and needs to be undone.
An agent consists of four elements, and it is advisable to be able to name all four of them before approving any of them:
The first three are usually defined. The fourth is the one that is missing in most of the projects we audit, and it is the one that determines whether the agent is an asset or a liability.
McKinsey's 2025 survey found that 62.1% of organizations were experimenting with agents, but only 23.1% had scaled them in at least one function. The gap between these two figures is the subject of this article.
Gartner was more direct in June 2025: it predicts that More than 40% of agentic AI projects will be canceled before the end of 2027, This is due to rising costs, unclear business value, or inadequate risk controls. In the same analysis, Anushree Verma, senior director of analysis at the firm, estimated that of the thousands of suppliers presenting themselves as agent specialists, only around 130 truly are. The rest were dubbed with a term that should be added to the vocabulary of any purchasing committee: *agent washing*.
This is not a statistic against the agents. It is a statistic about what separates the projects that survive from those that don't: projected costs, defined value, and designed risk controls.
Attribute | Employee | AI Agent |
Legal responsibility | Own, within the labor framework | From the company that deploys it, always |
Judgment in the face of the unexpected | He recognizes that something doesn't fit and asks | Complete the task with the information you have. |
Accumulated context | Years of tacit knowledge of the company | Only what is provided to you in each execution |
Traceability | Reconstructible through conversation | It only exists if the record was designed. |
Cost of a mistake | Limited and detectable | Scalable: repeats the same mistake thousands of times |
The line between legal responsibility and liability is no longer theoretical. In the case of *Moffatt v. Air Canada*, decided in 2024 by the British Columbia civil court, the airline argued that its chatbot was a separate entity responsible for its own statements. The court called this argument remarkable—in the worst sense—and ordered the company to compensate the customer for the erroneous information provided by the system.
The operational conclusion is simple: the company is responsible for what its software says and does. Always. There is no delegation architecture that transfers that responsibility to the model provider.
The cost of error is the row that committees most underestimate. An employee who misinterprets a policy applies that error to every case they handle in a day. An agent applies it to every case until someone catches it. Scale, which is the selling point, is also the risk multiplier.
It would be absurd to write this article from a defensive position. Agents solve problems that previously had no good solution, and these are the patterns where they consistently perform well:
Tasks with ambiguous language and structure. Reading fifty emails from suppliers, extracting the order number and the promised date, and entering it into a system. Previously, this required rigid rules or manual work.
Orchestration between systems that do not communicate. Consult three sources, cross-reference them, and propose a course of action. The agent does not replace integration, but it reduces friction when full integration is not cost-effective.
First level of classification and routing. Decide which queue an incident goes to, what priority it has, and what information is missing to resolve it.
Preparing decisions, not executing decisions. The most profitable model we've seen: the agent gathers, compares, and proposes; the person approves. The savings are in the preparation, which is where the 80% of time was.
Note that in all four cases the agent operates within a defined process. None of them operates on a process that no one has mapped, for the same reason we explained when discussing Why AI can't fix a poorly designed company
We propose replacing "digital employee" with supervised automated capacity. It's uglier and less marketable, and it describes exactly what you're buying.
Changing the vocabulary changes the questions asked in the meeting. No one asks a "digital employee" what their maximum scope of action is, because employees don't have a maximum scope of action defined in a document. An automated capability, however, is asked:
A project that can answer all six questions is production-ready. One that cannot answer three or more is not ready, regardless of how well the demo performs.
The issue of identity and permissions deserves its own treatment, because that is where the real risk lies: we have developed it in non-human identities, AI agents, and cybersecurity
When someone proposes deploying an agent, the useful question isn't what it can do. It's What happens the day he makes a mistake?, Because it will make mistakes: it is probabilistic software and that is in its nature, not its quality.
If the response includes a complete record of what happened, a reversal mechanism, an identified responsible party, and clear limits on action, the project is well-designed. If the response is that it won't go wrong because the model is so good, what's being purchased isn't a system: it's an expectation.
A chatbot responds with text; an agent executes actions on real systems. An agent consists of a goal, a set of tools, a set of permissions, and a control mechanism. The difference is critical: a chatbot's worst mistake is an incorrect response, while an agent's worst mistake is an incorrect action that has already been performed.
Because it attributes judgment, accumulated context, and personal responsibility to a system that lacks them. An agent completes the task with the available information even if context is missing, fails to recognize when something doesn't fit, and repeats the same mistake on a large scale until someone detects it.
The company that deploys it. In the case of Moffatt v. Air Canada (2024), the court rejected the argument that the chatbot was an entity responsible for its own statements and ordered the airline to compensate the customer for the system's erroneous information.
According to McKinsey (2025), 62.1% of organizations are experimenting with agents, but only 23.1% have scaled them in any function. Gartner predicts that more than 40.1% of agentic AI projects will be canceled before the end of 2027 due to costs, unclear value, or insufficient risk controls.
In tasks with ambiguous language and structure (extracting information from emails or documents), in orchestration between systems that do not communicate, in classification and routing of incidents, and above all in preparing decisions that are later validated by a person.
Being able to answer six questions: what can and cannot be done, with what identity accesses the systems, what is recorded of each execution, what decisions require human validation, how an incorrect action is reversed, and who is responsible for the result.
Are you evaluating an agent project? We analyze the use case, permissions model, and risk controls before it becomes a deployment that's difficult to reverse. Two hours of consulting, no obligation. Let's talk → |
Español Español 
English
Français
العربية
Português
Deutsch
Italiano
