Agentic AI
Agentic AI describes the property of a system that plans, acts and evaluates across multiple steps; an AI agent is a system that has that property. The distinction is the one between "electric" and "electric motor", and in product literature the two words are used interchangeably. What carries practical weight is something else entirely: whether the system can only read, or can also act.
What is the difference between agentic AI and an AI agent?
Grammatical rather than technical. Agentic AI is the adjective, an agent is the thing. Vendors use both for the same products, which is why the question is asked so often and why the answer is unsatisfying.
The distinction that does matter is access. A system that reads your documents and proposes a draft is categorically different from one that triggers an order or amends a record — even when both are sold under the same word.
Is this a new capability or a new label?
Mostly a new label on a gradual capability improvement. Multi-step tool use existed before the term became common; what changed is that models follow multi-step instructions more reliably, which moved some workflows from demo to production.
The useful consequence of the hype is that tool-calling interfaces became standardised. The unhelpful consequence is that ordinary automation now gets sold as agentic, at agentic prices, for processes where a rule would be cheaper and auditable.
What should a buyer actually evaluate?
Three things, none of which appear in the word. What the system is permitted to change. How many unsupervised steps run before a human sees anything. And what the recovery path looks like when an intermediate step was wrong.
A supplier who answers those three precisely has built something. A supplier who answers with autonomy levels and capability tiers is describing a category, not a system.
What it is not to be confused with
Generative AI
Generative describes what the model produces — text, images, code. Agentic describes how a system organises work around it. A system can be both, either, or neither, and the words answer different questions.
Robotic process automation
RPA repeats a recorded sequence of interface actions. It breaks when the interface changes and never chooses a different path. Agentic approaches tolerate variation and are correspondingly harder to audit. Neither is superior; they fail differently.
Frequently asked
What does agentic AI mean?+
A system that takes a goal, decomposes it into steps itself, invokes tools to carry them out and assesses the results, instead of producing one output from one input. The word describes an architectural approach, not a capability threshold, and no standard defines where it begins.
Is agentic AI regulated differently?+
No. The EU AI Act classifies by use case and, for direct interaction with people, by transparency obligation. The architecture is not a classification criterion. An agentic system in a minimal-risk use case stays minimal risk; a single-step system screening applicants is high risk.
Should we wait for the technology to mature?+
For write access, in most cases yes. For read-only work — search, comparison, drafting — waiting costs more than it saves, and that work is where the returns currently are. The two decisions are separable and should be made separately.
How do we pilot this without committing?+
Pick one process, restrict it to reading, and require a human to approve the output. If that produces measurable time savings, the case for extension is made with evidence. If it does not, you learned that for the price of a small pilot rather than a platform.
Related terms
Before the vendor conversation
Write down what a system would be allowed to change in your environment. Most organisations discover the honest answer is "nothing yet" — which narrows the shortlist usefully and quickly.
Request a test