Are we building AI agents or just rebranding chatbots?

Cesar Schmitzhaus, Managing Director at Teltec Data, explains why AI agents must execute tasks, integrate systems, and deliver business value.

Every two or three meetings, someone introduces me to an AI agent. When I ask what it does, the description I often receive sounds more like a chatbot. I understand where the confusion comes from. I've seen this movie before, just under a different name. A few years ago, the term “digital transformation” became an umbrella for very different initiatives, ranging from replacing a server to digitizing a paper form. Over time, the concept lost precision. Today, I see “AI agent” beginning to follow a similar path.

The distinction may seem technical, but it has a direct impact on business expectations. When a solution consults a document repository and returns an answer in natural language, we have a chatbot with access to knowledge. Even when Retrieval-Augmented Generation (RAG) is involved, a technology that enables the system to retrieve relevant information from its own data sources before generating a response, its primary function is still to answer questions. That has value. It can reduce the time spent searching for a contract clause, an internal policy, or a procedure. Responding, however, is a smaller category of work than executing.

An agent, as I understand the concept, must be capable of performing tasks. It interacts with systems, queries APIs, triggers processes, updates records, and makes decisions within predefined parameters. The difference becomes evident after the response is generated. If nothing happens within the company's systems, we may still be looking at a conversational interface rather than agentic automation.

Have a simple test for evaluating these projects. Turn off the chat interface and observe what happens. Does the solution continue performing activities, interacting with systems, or moving a process forward? If the answer is no, it is worth asking whether we are truly dealing with an agent or simply a chatbot with an additional layer of intelligence.

This discussion is already emerging in the market. Gartner, for example, uses the term “agent washing” to describe the practice of presenting solutions as agentic AI when, in reality, they offer capabilities already familiar in AI assistants, robotic process automation, and chatbots. In an analysis published in 2025, the firm estimated that more than 40% of agentic AI projects could be canceled by the end of 2027, primarily due to rising costs, unclear business value, or insufficient risk controls. More important than the percentage itself is the reason behind it. An AI project can be technically functional and still fail to deliver business value. This happens when organizations start with the technology and only later try to determine which business problem it should solve.

This is where the conversation about agents must move beyond the technology department and reach the business areas that will actually be impacted by automation. Before asking which model will be used, which architecture will be adopted, or which vendor is behind the solution, organizations need to understand what work will be executed, which systems must be accessed, which decisions can be automated, and what outcomes are expected.

Building an agent that truly performs tasks also requires a different infrastructure than a chatbot. The solution must connect to existing systems, work with multiple data sources, operate within security and governance frameworks, and maintain an appropriate level of reliability. In many cases, this means integrating ERP, CRM, financial systems, customer service platforms, productivity tools, and internally developed applications.

This complexity helps explain why so many initiatives remain limited to conversational interfaces. It is relatively easy to place an AI layer on top of a document repository and allow employees to ask questions. It is something entirely different to enable the same technology to consult a system, make a decision within defined rules, execute an action, record the result, and escalate an exception to a human when necessary.

The challenge, therefore, is not simply making AI converse. It is making AI participate in real business processes.

This also changes how return on investment should be measured. The number of questions answered may be an interesting metric for a chatbot, but it does not necessarily demonstrate the value of an agent. For an agentic application, it makes more sense to track metrics such as process execution time, reduction of manual tasks, volume of automated transactions, number of required human interventions, and impact on costs or revenue.

Consider, for example, an agent used in a procurement process. A chatbot can explain the company’s purchasing policy and answer questions about approval limits. An agent could consult the policy, verify a request, gather information from the ERP, identify whether approval is required, route the request to the appropriate stakeholder, and record the outcome in the system. Both are AI applications, even if they share a conversational interface.

This distinction matters because it prevents organizations from confusing a technology demonstration with process transformation. A conversational interface that answers questions effectively can impress in a presentation. A system that eliminates manual steps in a critical business process must prove its value in day-to-day operations.

For that reason, before calling an initiative an AI agent, I would ask a few basic questions. What task does it perform? Which systems does it interact with? What happens after it generates a response? What decisions can it make independently? In which situations does it require human intervention? How does the organization measure the results of that automation?

These are simple questions, but they help distinguish different types of projects that are increasingly receiving the same label.

AI has the potential to go far beyond conversation. The next step for organizations should not be simply providing more AI interfaces to employees, but identifying where this technology can take ownership of concrete parts of the work, with security, governance, and measurable returns.

In the end, perhaps the most important question is not whether a solution is truly an agent. It is what that solution can do when nobody is typing a question.

*By Cesar Schmitzhaus, Managing Director at Teltec Data.