Quick answer
An AI agent is software that you give a goal; it decides the steps itself, uses tools such as your calendar, email, phone or spreadsheet to carry them out, and checks the result. There is no single official definition: Anthropic, Amazon Web Services and Google Cloud each word it differently.
Key takeaways
- An AI agent is given a goal, chooses its own steps, uses your tools and checks the result.
- There is no single official definition. Anthropic, Amazon Web Services and Google Cloud each word it differently, and sellers use the word loosely.
- Three tests (this site's rule of thumb, not a standard) tell a real agent from a renamed product: it chooses its own steps, it uses your tools, and it checks its work or hands over when unsure.
- If the steps are the same every time, you want an automation, not an agent.
- This page explains the word. For which agent jobs are ready to hand over, what a first one costs and a starter budget, use the live agents page.
What is an AI agent?
An AI agent is software that you give a goal; it decides the steps itself, uses tools such as your calendar, email, phone or spreadsheet to carry them out, and checks the result. Unlike a chat assistant, which only replies when you ask, an agent is closer to handing a colleague a job than asking a colleague a question.
An example makes it concrete. "Draft a reply to this customer" is a task. "Get this customer a time on Thursday and confirm it" is a goal. To meet it, software has to read the calendar, pick a slot, send the message and notice whether the customer answered. If it does those steps without you scripting each one, you are looking at an agent.
Who defines "AI agent", and do they agree?
There is no single official definition, so treat every one as one company's wording. The word is also used loosely in marketing. 3 vendors that build the underlying technology put it this way:
| Source | How it words the idea | What an owner should hear |
|---|---|---|
| Anthropic, "Building effective agents" | Agents are "systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks." | The software chooses the path. |
| Amazon Web Services, "What are AI agents?" | "Humans set goals, but an AI agent independently chooses the best actions it needs to perform to achieve those goals." | You set the goal; the software chooses the actions. |
| Google Cloud, "What are AI agents?" | "AI agents are software systems that use AI to pursue goals and complete tasks on behalf of users." | You give the goal; the software acts for you. |
The wording differs. Amazon Web Services and Google Cloud both start from a goal; Anthropic describes the software directing its own process and tool use.
How can you tell a real AI agent from a renamed product?
Run three tests; a product that fails the first one is an automation or a plain text tool, whatever its label says. The tests are this site's rule of thumb, not an industry standard. Ask the seller, or read the setup screen:
| Test | What a pass looks like | What a fail means |
|---|---|---|
| 1. Does it choose its own steps? | You describe the goal and it works out the route. | You drew the flowchart and it only follows it: an automation. |
| 2. Does it use your tools? | It acts where your work lives: calendar, inbox, phone line, spreadsheet, booking or accounting software. | It only produces text for you to paste somewhere: an assistant. |
| 3. Does it check its work or hand over when unsure? | It reports back, or stops and asks you at a point you can set. | It never stops and never reports: a risk, not a feature. |
Each vendor above says something on these points. On choosing steps and tools, Amazon Web Services writes that "the agent identifies when a task requires a tool and then delegates the operation accordingly," and Google Cloud writes that "tools are functions or external resources that an agent can utilize to interact with its environment." On human checks, Anthropic says agents "can then pause for human feedback at checkpoints," and Google Cloud says its AI assistants, which it defines as agents built to work directly with users, "reason and take action on the users' behalf with their supervision." Amazon Web Services lists "safeguards, such as human reviews" against biased or inaccurate output, which is a check on results, not an approval step before an action.
Passing all three tests does not mean the agent is good, only that the label fits. How well it does a given job is a separate question.
What should you ask a seller who says "agent"?
Ask three questions, one per test. Use them to cut through the label:
- "When I give it a goal, who decides the steps, you or the software?"
- "Which of my tools does it work in, and what exactly is it allowed to do there?"
- "Where does it stop and ask me, and can I change that?"
A clear answer to all three means the label probably fits. A vague answer to the first means you are likely buying an automation; to the third, plan to check every result yourself.
When do you want an agent, and when an automation?
If the steps never change, you want an automation, not an agent. When the path differs from case to case, read the AI agents for small businesses page, which covers when an agent is the wrong tool.
Where does this guide go next?
This is the hub of our "AI explained" series; each block below groups the questions an owner usually asks next. The table lists the questions only; it does not link to guides that are not published yet, and it sets no dates.
Start here: most readers begin with three questions: what agentic AI is, how Claude compares with ChatGPT, and what AI is.
| Block | The question it answers |
|---|---|
| AI basics | What is AI, and what is a language model? |
| Agents and the plumbing behind them | What is agentic AI, and how does an agent connect to your tools? |
| The assistants | What are ChatGPT, Claude and Gemini, and how do you choose between them? |
| Picking tools and what they cost | What does an agent cost, and can you build one without code? |
| Risks and the bigger picture | Are agents safe for a small business, and what should you not hand over? |
This page tells you what the word means; the AI agents for small businesses page tells you which agents are ready to use. We did not find a public dataset that counts how many small businesses use AI agents, so be wary of any adoption figure quoted for them. Our Census data on AI use by business size and industry covers AI use in any business function, and the search interest by city study measures how often people search Google for agent-related phrases. Neither counts agents in use.
How was this page made?
The definitions above are quoted from the Anthropic, Amazon Web Services and Google Cloud pages listed in the sources, read on October 7, 2026, and nothing here recommends a product.
We also looked for a definition of "agent" on OpenAI's agents guide (developers.openai.com) and on Microsoft Learn's Foundry Agent Service overview and agent orchestration pages; none of them states one, so those vendors are not in the table, which is not a judgment on them. This page was drafted with Claude, an AI model made by Anthropic, and checked against each vendor's published material on October 7, 2026. Anthropic is one of the three vendors quoted, and it has the same space and the same treatment as the other two.
Your next step: use this page as a filter. Pick the block above that matches your question, and the next time a seller says "agent", ask which of the three tests their product passes.
FAQ
Is an AI agent the same as an AI agency?
No. An AI agent is software. An AI agency is a company that sells AI services, such as building automations or setting up tools for clients. An agency may build agents for you, but the two words name different things.
Do I need an AI agent to use AI in my business?
No. Many useful jobs, such as drafting an email or summarizing a call, need only a chat assistant or a fixed automation. An agent earns its place when the steps differ from case to case and you are willing to check its work.
Does an AI agent act without asking me?
It can if you give it the access and the permission, which is the reason to decide in advance what it may do alone. Anthropic describes agents pausing for human feedback at checkpoints. Google Cloud describes its AI assistants, which it defines as agents built to work directly with users, as acting with the user's supervision. Amazon Web Services lists human reviews as a safeguard against biased or inaccurate output, which is a different kind of check. Ask any product where that pause is and whether you can set it.
What does an AI agent need from me to get started?
A goal in plain words and access to the tools it is meant to use. Google Cloud describes tools as the functions or external resources an agent uses to interact with its environment, so in practice that means permission to your calendar, inbox or spreadsheet. Give the least access the job needs.
Is “AI agent” a regulated or official term?
This site found no single official definition. Each vendor words it differently, as the table above shows, and sellers apply the label to very different products. Judge a product by what it does, using the three tests.