Agentic AI is artificial intelligence that pursues a goal by taking actions: it plans steps, calls tools and systems, reads the results and decides what to do next, with limited human input between steps. Generative AI answers. Agentic AI does. That shift from producing text to changing records is why it needs a different kind of control.
What makes AI "agentic"?
Three things together. A goal rather than a single question, such as "resolve this billing dispute". Access to tools that change the world outside the conversation, such as updating a record, sending a message or initiating a refund. And a loop that lets the system decide its next step from what happened in the last one.
Remove any of them and you are back to a chatbot or a script. A model that drafts an email is generative. A model that drafts it, sends it, reads the reply and updates the case is agentic.
Why does agentic AI change the risk?
Because mistakes stop being words on a screen. A generative model that gets an answer wrong produces a wrong answer, which a person can ignore. An agent that gets a step wrong may have sent the message, posted the entry or moved the money before anyone looks. In a regulated business, the question shifts from "is the output good?" to "was this action allowed, by whom, on what evidence, and can it be undone?"
That is a question about authority, not intelligence. A better model does not answer it.
How is agentic AI used in practice?
Usually as agents built for specific jobs, often combined in an agentic workflow where some steps are fixed and others are left to the agent. In regulated firms, most start at read and draft: the agent researches, prepares and proposes, and a person approves. Routine, reversible steps move to the agent once it has proven it handles them well. Steps that move money or are hard to undo come last, if at all.
For example, a broker's operations team deploys an agent to chase missing documents on new policies. It reads the case, identifies what is missing, drafts and sends standard requests, and updates the checklist when documents arrive. It does not bind cover. That stays with an underwriter.
How is agentic AI different from generative AI?
Generative AI produces content: text, images, code, summaries. Agentic AI uses generative models as one component, but adds tools, a goal and a loop so it can act. Every agentic system is built on generative models; most generative AI uses are not agentic. The difference between an agent and the workflow it runs inside is covered in how an agentic workflow differs from an AI agent.
Which terms sit near agentic AI?
- Agentic payments: agentic AI applied to initiating or preparing payments.
- Can an agentic operating system work with any model?: why governance should sit outside the model.
- The seven jobs of an agentic OS: what the control layer has to do.
- The governed agentic operating system: the layer that decides what agents may do.
Questions people ask
Is agentic AI the same as autonomous AI?
Not quite. Agentic describes the ability to act; autonomy describes how much it may do without a person. An agentic system can run with almost no autonomy, drafting everything for approval, or with a great deal.
Do you need a new model for agentic AI?
No. Agentic systems use the same kinds of models as generative ones. What changes is the surrounding system: tools, memory, bounds on what the agent may do, and a record of every action it takes.