Agentic AI

Open full sized image

When artificial intelligence goes from answering to acting.

Encyclopædia Britannica, Inc.

Many people use artificial intelligence to answer questions, generate text or images, and get recommendations. You can ask AI to write an email, summarize a document, or suggest which laptop fits your budget. Agentic AI takes the next step. Instead of telling you what to do, it can take actions on your behalf.

Given a goal, an agentic AI system can determine the steps needed to accomplish it, use web browsers, apps, and other tools to carry out those steps, and adjust its approach as it works. For example, an AI agent can do more than suggest several hotels for an upcoming trip. It can search available rooms, compare prices based on your budget and preferences, and potentially make the reservation.

Agentic AI refers to artificial intelligence systems designed to take action with some degree of autonomy. Instead of requiring a human prompt for every step, an agentic system can determine what needs to happen next and continue working toward the objective it’s been assigned.

An AI agent is the software that carries out these tasks. It can receive information, make decisions, and take actions toward a goal.

The amount of autonomy can vary depending on the agent and the limits placed on it. One agent may perform several steps independently but require your approval before sending an email or spending money. Another could operate largely on its own within predetermined rules.

Although systems vary, an AI agent uses an AI model to make decisions and relies on software tools to gather information and carry out tasks.

An AI agent can repeat these steps until it accomplishes the goal or determines it needs further instructions.

The distinction between generative AI and agentic AI mostly comes down to what happens after the AI produces an answer.

Generative AI can create text, images, audio, video, and computer code. Ask a chatbot to compare three laptops, and it can generate a table explaining their differences.

Agentic AI can go a step further and act on that information. Give an agent a $1,000 laptop budget and permission to make a purchase, and it could research models, find one that meets your specifications, compare prices, and place the order.

These technologies aren’t mutually exclusive. Agentic systems often use generative AI models as part of their decision-making process.

The main appeal of agentic AI is delegation. Instead of using AI to determine your next step and then completing it yourself, you can assign an entire task to an agent.

AI agents could save you time by handling repetitive administrative work, such as organizing a calendar, responding to routine emails, or preparing expense reports. They could also help manage more complicated tasks that involve gathering information from multiple sources, making decisions, and adjusting plans when circumstances change.

Agentic commerce extends these capabilities to shopping and payments, allowing AI agents to find products and make purchases within limits you set.

Giving an AI agent the ability to act creates risks that don’t typically exist when a chatbot generates an answer. If a chatbot recommends a flight, you can check whether it meets your needs before booking it. But if an AI agent books the wrong flight on your behalf, correcting the mistake could cost you money. A mistake made early in the process can lead to more mistakes as the agent works through a task.

Other potential risks include:

One way to reduce these risks is to follow the principle of least privilege, a security practice that means giving a system only the access and permissions it needs to perform a task. You might allow an AI agent to research purchases and add items to a cart, for example, but require your approval before checkout.

You can also set spending limits, restrict account access, review activity logs, and require confirmation before an agent takes actions that involve money or sensitive information. The more consequential the action, the more important human oversight becomes.

Agentic AI promises to let software carry out increasingly complex tasks on your behalf. But the more independence you give an AI agent, the greater the importance of knowing what it can do—and when it needs your approval.