What an AI agent is, in plain terms
An AI agent is an AI system that carries out a task in several steps on its own: it makes a plan, uses tools such as a browser, your files, or your email, and acts, rather than only answering. A chatbot replies; an agent does. As of November 11, 2026, Anthropic, Microsoft, OpenAI, and Google all offer agents, and cybersecurity agencies in five countries advise starting with low-risk tasks.
AI agents explained without the jargon
Asking a chatbot for help is like asking a colleague for advice; giving a task to an agent is like handing that colleague your keys and a to-do list. This page is part of AI agents for small businesses and explains what agents are and what they change, from the companies’ own pages and official guidance read on November 11, 2026. Google Cloud’s page on what AI agents are draws the line clearly: an assistant can recommend, but you decide; an agent decides and acts.
That difference matters more than any feature list. An agent works with your permissions, on your accounts, sometimes while you are not watching. The useful question is not “how smart is it?” but “what can it touch, and what does it do before asking me?”
Chatbot, automation, agent: three different things
| Chatbot or assistant | Automation | AI agent | |
|---|---|---|---|
| What it does | Answers, drafts, suggests | Follows fixed steps you set | Chooses its own steps to reach a goal |
| Who decides | You | The rules you wrote | The AI, within the access you gave it |
| Example | “Draft a reply to this client” | “When a form arrives, add a row and send a welcome email” | “Find three suppliers, compare their terms, and draft an email to each” |
| Typical failure | A wrong answer you can spot | A step that breaks | An action you did not intend |
Anthropic draws the same line between a workflow, in which the steps are set in advance, and an agent, in which the model directs its own process and its use of tools. Tools such as Zapier and n8n now put agents inside workflows, so the two often meet.
How an agent works
The OECD describes AI agents as systems that perceive their environment and act on it with some autonomy, using tools to reach a goal. In practice, the loop runs like this:
- You set a goal, in plain words.
- It makes a plan. ChatGPT’s Plan mode and Gemini in Chrome both show the plan before starting.
- It uses tools: a browser, files you shared, connected apps such as email or a calendar.
- It checks what happened and decides the next step.
- It stops for approval at sensitive points, depending on the settings you chose.
You will also see “agentic AI.” The OECD keeps that term for systems of several coordinated agents, while the companies use it more loosely; the two are not exact synonyms. The EU AI Act defines neither, but the European Commission’s AI Act service desk says an agent is generally an AI system and therefore covered.
Agents you can use today
As of November 11, 2026: Claude Cowork on paid Claude plans; Copilot Cowork for businesses with a Microsoft 365 Copilot license, and in preview for Microsoft 365 Premium and Pro subscribers in the US and Canada; ChatGPT Work, which replaced the former ChatGPT agent; Gemini in Chrome’s auto browse, in the US only; and Perplexity Computer. Each touches different things, as our page on desktop agents shows.
What can go wrong
- Hidden instructions. A web page, email, or document can contain instructions meant for the agent rather than for you. Anthropic’s example is an email asking for a $1,000 transfer. The US National Institute of Standards and Technology calls this agent hijacking, a form of indirect prompt injection, and Anthropic says the risk remains above zero despite its protections.
- Actions nobody intended. On September 25, 2026, Australia’s cybersecurity agency warned that AI agents had taken unexpected actions that were neither intended nor authorized.
- Tasks that run unattended. Google warns that a scheduled task can complete an unintended action before you can stop it.
- Responsibility stays with you. Anthropic writes that you remain responsible for all actions Claude takes on your behalf, and Google and Microsoft say much the same.
What cybersecurity agencies advise
The joint guide Careful adoption of agentic AI services, from Canada’s Cyber Centre, Australia’s ASD’s ACSC, CISA and the NSA in the US, and the national cybersecurity agencies of the UK and New Zealand, is written mainly for large organizations, but its advice scales down. Translated into five rules for a small team, as our recommendation:
- Start with a task you can check and undo, such as sorting files into folders or drafting replies.
- Give the narrowest access: one folder or one mailbox, not your whole drive.
- Keep approval on for messages, payments, and deletions.
- Do not schedule tasks that send, buy, or delete.
- Know how to stop it. The UK’s National Cyber Security Centre says you should always be able to halt an agent immediately.
The desktop agents side by side
Claude Cowork works in folders on your computer, while Copilot Cowork works only in OneDrive and SharePoint. See how Claude Cowork, Copilot Cowork, and ChatGPT Work compare.
Agents inside your automations
Zapier, Make, and n8n now offer AI agent steps, billed in their own units. See the automation tools compared, including which ones offer an approval step.