What is agentic AI?

"Agentic AI" does not name a product or a specific model: it describes what a system does with a model. A chat that answers you is not agentic; a system that takes a request, decides the steps, goes into other sites or apps and hands back the finished job is. A store cares because some of the people who visit its product pages, compare prices and reach the cart are no longer people, but systems acting for them, with rules of their own. The part that does the work is the AI agent, and its application to shopping has a name of its own, agentic commerce. This entry answers the question that comes before both: what makes a system agentic.

How it works

Google Cloud defines it as a form of AI centered on deciding and acting with autonomy: where traditional AI responds to commands or analyzes data, agentic AI can set goals, plan and carry out tasks with minimal human intervention.⁠^{[1]Google Cloud (updated September 16, 2026). What is agentic AI? Definition and differentiators} It describes how it works as a five-step loop:⁠^{[1]Google Cloud (updated September 16, 2026). What is agentic AI? Definition and differentiators}

1. Perception. It gathers information from different sources: databases, interfaces, text or images.
2. Reasoning. A large language model (LLM) analyzes that information, understands the context and proposes solutions.
3. Planning. It sets goals, breaks them into smaller steps and decides how to meet them.
4. Action. It carries out the tasks, makes decisions or interacts with other systems.
5. Reflection. It evaluates whether what it did worked and uses the result to adjust its plans.

What sets it apart from generative AI

For Google Cloud, agentic AI is a subset of generative AI: the part that organizes and runs agents that use an LLM as a "brain" to act through tools. It goes beyond creating content and beyond tool calling (function calling) because it executes actions in the underlying systems to meet higher-level goals. Its example: generative AI writes the materials for a campaign; agentic AI publishes them, tracks their performance and adjusts the strategy based on the results.⁠^{[1]Google Cloud (updated September 16, 2026). What is agentic AI? Definition and differentiators}

More or less agentic

Anthropic uses "agentic systems" for all these variants and separates two kinds. Workflows orchestrate the model and its tools along paths set in advance in the code. Agents direct their own process and decide on the fly which tool to use and when.⁠^{[2]Anthropic, engineering blog (December 19, 2024). Building effective agents} That is why it helps to think in degrees: a system is more agentic the more it decides and the less is written in advance.

That autonomy has a price, and Anthropic says so plainly: agents fit open-ended problems where you cannot predict how many steps will be needed, but they cost more and errors compound from one step to the next. It recommends testing them thoroughly in sandboxed environments, with safeguards, stopping conditions (a maximum number of iterations, for example) and pauses to ask a person for input at checkpoints or when they get stuck.⁠^{[2]Anthropic, engineering blog (December 19, 2024). Building effective agents}

What it already does in a purchase

Commerce is one of the places where agentic AI has gone from demo to product. On November 13, 2025, Google began rolling out an agentic checkout in US Search: you say which product you want to track, with size, color and what you are willing to spend; it alerts you when the price drops within that budget and offers to have Google buy it for you on the store's site with Google Pay, always with your permission. It launched with Wayfair, Chewy, Quince and some Shopify stores. The same day it introduced another agentic feature that calls nearby stores to ask whether they have a product, at what price and with which promotions, in categories such as toys, health and beauty, or electronics, also in the United States.⁠^{[3]Google, The Keyword (November 13, 2025). Let AI do the hard parts of your holiday shopping}

Once the system pays, the question is who authorized it. The AP2 (Agent Payments Protocol) distinguishes two cases: the person approves the specific order, or signs a set of conditions beforehand and lets the agent buy on its own within them; if the order falls outside what was signed, the store can send it back to the person for approval.⁠^{[4]AP2, the protocol's official site. Flows}

In a store

A fictional example: a person asks an assistant with an agent mode for "waterproof trail running shoes, size 9, for under a hundred dollars, arriving before Friday." Here is what the system does in your store, step by step, and what it needs to find:

Each row is a point where the request can leave your store for another. Because errors compound,⁠^{[2]Anthropic, engineering blog (December 19, 2024). Building effective agents} one ambiguous piece of data in the reasoning step is enough for the system to drop your product page.

How to check

1. Make the request yourself. Ask an assistant with an agent mode to buy a specific product from your store, with a size and a budget, and stop it before payment. Note at which step it hesitates or stalls.
2. Read your product page without images or tabs. Price, stock, variants, shipping and delivery time have to be in the page's text.
3. Look in your logs for the visits a system makes on a person's behalf. Perplexity, for example, identifies its own as Perplexity-User, distinct from its crawler, and warns that, because a user requests them, they generally do not follow the rules in robots.txt.⁠^{[5]Perplexity. Perplexity Crawlers}
4. If you use an agent in your own store, for customer service or returns, give it stopping conditions and checkpoints where a person decides, as Anthropic recommends.⁠^{[2]Anthropic, engineering blog (December 19, 2024). Building effective agents}
5. If you sell in the United States on Shopify, ask your platform whether your products are included in Google's agentic checkout.⁠^{[3]Google, The Keyword (November 13, 2025). Let AI do the hard parts of your holiday shopping}

Related concept: AI agent

An AI agent is the unit: a system that receives a task and carries it out with its tools. Agentic AI is the property and the whole system: how much it decides on its own and how it coordinates one or several agents for a larger goal. Google Cloud compares it to a toolbox: the agents are the tools, and agentic AI is using them in a coordinated way to build the whole house.⁠^{[1]Google Cloud (updated September 16, 2026). What is agentic AI? Definition and differentiators} When the goal is buying, the term that covers it is agentic commerce.

To go further, the agentic commerce reading path lays out the agents, the protocols and the payments.