What is agentic AI?

Ask whether a system is agentic and you have already framed it wrong. The word is not a badge a product wears or lacks. It names a quantity, how much of the deciding a system does for itself, and quantities come in amounts. The more useful question is how much agency a system exercises, and over which parts of a task.

Agent is the noun, agentic is the measure

An AI agent is a thing: a discrete system you can point at. Agentic is an adjective, and adjectives come in degrees. Warm describes a range of temperatures rather than a single one; agentic describes a range of autonomy the same way. A system can be barely agentic or strongly agentic or somewhere in the middle, and the same system can move along that range depending on how much of the work it decides without being told.

So the two words answer different questions. Agent asks what the entity is. Agentic asks how much agency that entity exercises.

Agentic AI describes a property, not a product category: the degree to which a system pursues a goal on its own, deciding its own next steps and acting on them with real consequence, rather than following a path fixed in advance. A system is more or less agentic along a continuum, and an AI agent is simply an entity that sits somewhere on it.

What makes one system more agentic than another

If agency is a matter of degree, the next question is what moves the dial. One line of research treats this head-on. In a peer-reviewed FAccT paper, Chan and colleagues decline to draw a binary boundary around the word and instead name four characteristics whose combination raises a system’s agency: how underspecified its instructions are, how directly it can affect the world, how goal-directed its behaviour is, and how far ahead it plans (Chan et al., 2023).

Read those together and agency stops being mysterious. A system told exactly what to do, acting only through a human, chasing no goal of its own, is barely agentic. A system handed a loose objective, able to act in the world directly, planning several moves ahead, is strongly so. Most real systems land between, high on one characteristic and low on another.

The idea of measuring autonomy predates the chatbots

Treating autonomy as a scale rather than a label is decades old, and it did not start with language models. In 1978, Sheridan and Verplank set out a ten-level scale of automation, running from a machine that offers no help at all to one that acts entirely on its own and tells the human nothing (Sheridan and Verplank, 1978). A ten-rung ladder makes the point on its own: autonomy was never on or off. It had rungs.

Parasuraman, Sheridan and Wickens later generalised the idea in a peer-reviewed IEEE paper. Rather than rate a whole system with one number, they split its work into four functions, acquiring information, analysing it, deciding on an action, and carrying that action out, and observed that each can be automated to a different degree, “from low to high, i.e., from fully manual to fully automatic” (Parasuraman et al., 2000).

That last move is the one worth keeping. A system need not be agentic at a single setting. It can decide for itself where a human once decided, while still leaving the final action for a person to approve.

Is it just a new word for automation?

Not quite, and the gap is the one Parasuraman’s functions expose. Ordinary automation runs a path someone fixed in advance; its autonomy sits low on the deciding function even when it acts fast and often. A system earns the word agentic as it takes over more of the deciding, choosing steps that were never laid out for it. Where that line falls is its own question, covered in AI agent vs automation. For the property itself, the test is where the choosing happens: in the code ahead of time, or in the system at runtime.

Can a chatbot be agentic?

Yes, a little. A plain chatbot that only answers what you ask sits near the bottom of the scale; it decides how to phrase a reply and little else. Give it the ability to look something up, call a tool, or take an action on your behalf and it has climbed the ladder, without turning into a different kind of thing. This is why the agent versus chatbot boundary is fuzzier than it first looks. The agency crept up a rung at a time.

The dial, not the switch

The practical question is never whether something qualifies as an agent. It is how much agency you have handed over, on which functions, with what check left in human hands. That turns a label into a setting you choose on purpose. A system that drafts an email for you to send has kept the final action in your hands; one that sends it itself has taken that rung too. Same underlying capability, different position on the dial. Knowing which position you are at, and on which parts of the work, tells you more than any yes-or-no verdict on whether the thing is “an agent.”

Sources

• Chan, A. et al. Harms from Increasingly Agentic Algorithmic Systems. FAccT ’23. arXiv:2302.10329, 20 Feb 2023.
• Parasuraman, R., Sheridan, T. B. and Wickens, C. D. A model for types and levels of human interaction with automation. IEEE Transactions on Systems, Man, and Cybernetics, Part A, 30(3):286-297, May 2000.
• Sheridan, T. B. and Verplank, W. L. Human and Computer Control of Undersea Teleoperators. MIT Man-Machine Systems Laboratory, 1978.