"AI agent": decoding a catch-all concept that covers five very different realities

Chatbots, copilots, automation, autonomous agents: for the past two years, almost everything has been sold under the "AI agent" label. The word has become a catch-all, used for anything from a simple chat tool to a system that can act on its own inside a company's processes. For a business leader, this confusion is not just a matter of vocabulary: it distorts comparisons, budgets and decisions.

A quick comparison of studies shows how blurry the picture is. In April 2025, 79% of US executives surveyed by PwC said AI agents were already being adopted in their companies. In July of the same year, the Capgemini Research Institute found that only 14% of organizations had begun implementing agents, and just 2% had scaled them fully. McKinsey, for its part, counted 62% of organizations experimenting with agents in 2025, but only 23% scaling them across the enterprise.

These gaps are not necessarily errors. They show that people are not talking about the same thing. When a leader says "we use agents", they may mean a writing assistant open to all employees or a system that handles customer cases end to end on its own. Before asking how many agents to deploy, you need to know what you are talking about.

A word turned into a sales pitch

The phenomenon now has a name: "agent washing". Gartner describes it as rebranding existing products, such as AI assistants, robotic process automation and chatbots, without adding substantial agentic capabilities. According to the firm, of the thousands of vendors claiming to offer agentic AI, only about 130 actually do.

The consequences are tangible. Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value or inadequate risk controls. Some of these failures stem from an initial misunderstanding: the company bought an "agent" expecting autonomy and received a tool that requires as much human intervention as before.

Why the term has become so blurry

The first reason is commercial. "Agent" has become the word that sells, as "cloud" and "big data" did in their day. For a software vendor, the priority is not to be seen lagging behind, even if that means stretching the definition.

The second reason is technical: autonomy is not binary; it is a spectrum. Between a tool that suggests an answer and a system that completes an entire task without sign-off, there are many intermediate stages. Anthropic, one of the leading AI model developers, draws a useful distinction: workflows, in which AI follows a path predefined by humans, and agents, in which AI decides for itself which steps to take and which tools to use to reach a goal. Many products sold as "agents" actually belong to the first category.

The third reason lies in how things are measured. Surveys rarely ask the same question: some count experiments, others production deployments, others mere intentions. The same leader can therefore, in good faith, report using agents in one survey and not in another.

Five realities behind a single word

To make sense of it, it helps to distinguish five levels, from the simplest to the most autonomous.

The first is the conversational assistant. It answers questions, drafts and summarizes, but it does not act: every answer is read, then used or not, by a human. It is the most widespread use, and the one most often relabeled as an "agent" for no good reason.

The second is the embedded copilot. Built into existing software such as email, spreadsheets or a CRM, it suggests actions in a specific context: a draft reply, a formula, a follow-up. A human approves each suggestion. The productivity gain is real; the autonomy remains limited.

The third is the AI-enhanced automated workflow. A predefined process, such as invoice processing, request triage or report generation, uses AI at certain steps. The system can run without intervention, but it follows a path designed by humans. It is reliable and predictable precisely because it does not improvise.

The fourth is the truly autonomous agent. It receives a goal rather than a procedure, chooses its own steps, consults sources, uses tools and adjusts its approach based on the results. This is the level that justifies the word "agent", and it is also the one that requires the most oversight.

The fifth is the multi-agent system. Several specialized agents, each dedicated to one task, are coordinated to produce an overall result: one gathers, another analyzes, a third checks. This model can tackle complex problems, provided that coordination and control are designed in from the start.

Questions to ask before buying or deploying an "agent"

When faced with a solution or a project billed as agentic, three questions are often enough to clarify what is really on the table.

The first is about decisions. Who decides on the steps: the system, or a human who programmed them in advance? If the path is entirely predefined, it is an automated workflow. That is not a flaw, but it justifies neither the same price nor the same promises.

The second is about action. Does the system merely suggest, or does it actually act: send a message, change data, trigger a transaction? The more it acts, the more central the questions of access rights, traceability and accountability become.

The third is about control. Who checks what the system produces, at what point and against what criteria? An agent whose results nobody can explain or verify does not give the company autonomy. It exposes it to risk.

Naming accurately to decide accurately

Untangling the vocabulary is not an academic exercise. It is what makes it possible to compare offers, estimate return on investment and set the right level of oversight. A conversational assistant and a multi-agent system call for different budgets, different governance and different expectations.

Truly autonomous agents are nonetheless set to play a growing role. Gartner estimates that by 2028, 15% of day-to-day work decisions will be made autonomously through agentic AI, and that 33% of enterprise software applications will include agentic capabilities, up from less than 1% in 2024. All the more reason to learn now to tell a real agent from a mere label.

For a business leader, then, the right question is not "do we have agents?" but "what exactly do they do, who decides their actions, and who checks their results?" As long as those three answers remain vague, the word "agent" serves the narrative more than the strategy.

Hymeria's position

At Hymeria Consulting, AI agents are at the heart of our approach, which is precisely why we are careful to name them accurately. Our analyses draw on more than 40 specialized agents, each dedicated to a specific analytical task and coordinated with one another. But none of them decides alone what reaches the client: every deliverable is reviewed, challenged and validated by a senior expert. Studies, analyses, business plans: leaders receive a sourced, actionable deliverable in 5 to 10 days, within a fixed budget agreed before the project starts.

Related service: Innovation & Technology, technology landscape, make-or-buy analysis and product roadmap in 5 to 10 days.