What Is an AI Agent? (And What Most Vendors Actually Sell You)

Ask ten vendors what an AI agent is and you will get ten answers, most of them convenient rather than accurate. Here is a definition that does not come from anyone selling one.

The actual definition

The US National Institute of Standards and Technology describes agentic AI as systems that "function as autonomous agents capable of independently making decisions, learning from interactions, and adapting to changing environments," with "independent reasoning, goal-driven behavior, and the ability to interact dynamically with users, systems, and real-world scenarios." (NIST)

Academic surveys of the field converge on the same distinguishing features: an agent operates in a complex environment, takes instruction from multiple sources, responds to dynamic feedback, and pursues a goal autonomously across multiple steps using tools — as opposed to producing a single response to a single prompt. (arXiv survey, 2025)

In plain terms: a chatbot answers the question in front of it. An agent breaks a goal into steps, decides which tools it needs, acts, checks whether the action worked, and adjusts — without a human approving each step.

How this differs from what you already have

A chatbot is single-turn and conversational. You ask, it answers, the interaction ends. Even a chatbot with memory of prior turns in the same conversation is not planning multiple steps toward a goal — it is retrieving and responding.

RPA (robotic process automation) is deterministic. It follows a fixed script: click here, extract this field, paste it there. It is fast, cheap and reliable for stable, well-defined processes, and it breaks the moment the input format changes, because it was never built to adapt. (TechTarget)

An AI agent sits above both. It can use tools (including RPA scripts and chatbot interfaces as components), reason about which tool to use next, and adapt its plan when something doesn't go as expected.

The distinction matters commercially because these three things have wildly different price points, and a vendor selling you RPA at agent prices is the most common version of the problem below.

Agent washing

Gartner has named this practice directly: vendors rebranding existing chatbots, RPA tools and assistants as "agents" without the underlying autonomous capability. The firm predicts over 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls — a prediction that only makes sense if a large share of what is being sold as agentic today isn't. (Gartner, June 2025)

Gartner has been widely reported as estimating that only a small fraction of the thousands of vendors now claiming agentic capability have built something that actually qualifies — a figure repeated consistently enough across independent outlets that it is worth taking seriously, even though it does not appear verbatim in Gartner's public press release and should be attributed as a widely reported estimate rather than a direct quote.

Three questions that expose a rebrand

What happens when the system is uncertain? A real agent has a designed answer: escalate, ask, or stop. A rebranded chatbot or script usually doesn't, because uncertainty handling was never built.

Can it use more than one tool, and does it decide which one? If the "agent" only ever does one thing in one order, it is RPA with a new name.

What did it get wrong recently, and how did the system respond? A real deployment has an answer. A demo has never been tested against reality.

If a vendor cannot answer any of these cleanly, you are very likely being sold rebranded automation at agent prices — which is a fine thing to buy, at the right price, under the right name.

Kaizen Spark Tech designs and delivers software, AI, automation and digital infrastructure for businesses and institutions. Every statistic here is linked to its original published source.