What is agentic AI?

You've probably seen the word agentic in an ad lately, usually next to a promise that an app will do the work for you. It spread across product launches faster than most people could learn what it means. Agentic artificial intelligence (AI) is the broad label for AI that plans and carries out work in several steps, with some independence. That's a useful idea. But the label alone doesn't tell you how much a product can do.¹

The word describes a way of working more than a kind of product. Generative AI, the kind that writes text or makes images, answers when you ask and then waits for you. An agentic system takes a goal and keeps going. It plans the steps, takes one, looks at the result, and decides what to do next. If the first try fails, it can try another way without you asking again.¹²

It helps to picture a dial.

• At one end, the tool suggests steps and you carry them out.
• In the middle, it carries out the steps, and stops for your approval before anything important.
• At the far end, it acts on its own and reports back when it's done.

Products that use the same word can sit anywhere on that dial. Many let you choose where.

Under the hood, most agentic products start with the same kind of language model that powers a chatbot. What changes is what's built around it: tools that let it reach outside itself, a running record of the task, and a loop that checks each result before the next step. Topic 12, "What is an AI agent?", walks through how an agent uses them.²

Because the word is popular, companies use it for very different things. One product called agentic can compare flights across several websites and fill in the booking form. Another sends a reminder on a schedule, which software has done for decades. Both may be useful. So when you see the word, look past it to what the product does.³

In 2025, Stanford researchers asked 1,500 workers which of their tasks they'd want an AI agent to take on. The ones they most wanted help with were repetitive and low in value, mostly to free up time for more important work. Think of filling in the same form again and again, or checking one long list against another. You still set the goal, and you still judge the result. That's "augment, not replace" in practice: the tool takes on the steps, and you keep the decisions that matter.³⁴

Three questions cut through most of the marketing.

• What can it touch? Your email, your files, and your payment details, or only a search box?
• What can it do without asking you first?
• How can you see what it did?

Clear answers are a good sign. Vague ones are a reason to start with less access, and add more once it has earned your trust.

The next time an app calls itself agentic, open its settings before you hand it a real task. Look for the one that controls when it stops to ask for your approval, and see what it's set to. Then decide whether that's where you want it.