# What is an AI agent, in plain English, and how is it different from a chatbot? Four vendors' own definitions and what a small office should let one do

> An AI agent plans a task and uses tools to act; a chatbot answers and stops. What OpenAI, Anthropic, Microsoft and Google said on 1 October 2026.

- Page to cite: https://aeolookup.com/ai-for-small-firms/what-is-an-ai-agent-plain-english-difference-from-chatbot/
- Publisher: AEO Lookup (https://aeolookup.com/)
- Section: AI for small firms, Putting it to work
- Published 1 October 2026
- Key fact: On 1 October 2026 the documentation of OpenAI, Anthropic, Microsoft and Google each described an AI agent as a model that plans a task and uses tools to act over several steps, where a chatbot answers one message and stops; each also tells the user to check or supervise it, and Anthropic's help page advises manual approval for anything hard to undo, such as sending messages or making purchases.
- Sources: OpenAI, developer guide "Agents" and help articles "ChatGPT Work and Codex", "ChatGPT agent" and "ChatGPT Workspace Agents for Enterprise and Business"; Anthropic, "Building effective agents" (19 December 2024) and help articles "Use Claude Cowork safely" and "Use Claude in Chrome safely"; Microsoft, Copilot Studio overview (last updated 25 September 2026), Microsoft 365 Copilot agents overview (last updated 30 September 2026) and Copilot Terms of Use (effective 18 August 2026); Google Cloud, "What are AI agents?" and Gemini Apps Help, "Use Gemini Spark". All read in a browser on 2026-10-01.

An AI agent is an AI model that has been given tools, such as your files, your inbox or a web browser, and is allowed to work through a task in several steps without being prompted at each one. A chatbot answers the message in front of it and stops; on 1 October 2026 the documentation of OpenAI, Anthropic, Microsoft and Google all drew the line in that place.

## What each vendor says an agent is

| Vendor | Its own page, read 1 October 2026 | What it says |
|---|---|---|
| OpenAI | Developer guide "Agents"; help article "ChatGPT Work and Codex" | Agents plan and complete tasks using tools and keep their place across steps. Inside ChatGPT, Chat is for quick conversational help and Work is "an agent designed for longer, multi-step work". |
| Anthropic | "Building effective agents", published 19 December 2024 | A workflow follows steps fixed in advance by code. An agent is a system where the model can "dynamically direct" its own process and its use of tools. |
| Microsoft | Copilot Studio overview, last updated 25 September 2026 | "An agent is an AI assistant that handles conversations and completes tasks". It follows the instructions it is given, draws on the knowledge sources connected to it and uses tools to act. |
| Google | Google Cloud, "What are AI agents?" | Software that uses AI "to pursue goals and complete tasks on behalf of users". Google ranks three things by independence: a bot follows fixed rules, an assistant responds and leaves the decision to the user, an agent acts towards a goal on its own. |

Take the wording away and the four agree on three parts: a model, a set of tools, and permission to take more than one step before coming back to you.

## Chatbot and agent side by side

| | Chatbot | Agent |
|---|---|---|
| You give it | A question or some text | A goal, such as "file these invoices by supplier" |
| It does | Writes one answer | Plans steps, opens files or pages, acts, checks the result and repeats |
| When it is wrong | You read a wrong answer | A wrong action may already have happened |
| Who acts | You | The agent, unless you make it ask first |

The last two rows matter most. A chatbot's mistake sits on the screen until someone copies it. An agent's mistake can be a sent email or a deleted file.

## What an agent needs access to

An agent can do nothing until it is connected to something.

- **Files.** Anthropic's help page on using Claude Cowork safely says the agent reaches only the folders you have connected, and can read, write and permanently delete files in them. It asks before a permanent deletion.
- **Apps and accounts.** OpenAI's help page for its earlier agent mode says that once signed in to websites or connected to apps, the agent can reach emails, files and account settings and act on your behalf.
- **The browser.** Google's help page for Gemini Spark, its personal agent, says that in your own Chrome browser it can reach the same sites you can, including those you are signed in to.
- **Your screen.** OpenAI and Anthropic both say their browser agents work from screenshots of the page, so whatever is visible in that tab is seen.

Anthropic sorts an agent's tools into two groups. Read tools let the agent look at something: an inbox, a folder, a web page. Write tools let it change something: send, book, delete, pay. Write tools carry more risk.

An agent also has one risk a chatbot does not. It reads web pages and emails written by strangers, and a stranger can hide an instruction in one. The vendors call this prompt injection. Anthropic's example is an email telling the agent to ignore its instructions and transfer money. OpenAI and Anthropic both describe their defences, and both say the risk is reduced, not removed.

## What was on offer on 1 October 2026

- **OpenAI.** The help page for "ChatGPT agent" opens with a notice that it is no longer available and points to ChatGPT Work, which is offered on eligible paid plans. Workspace agents, built once and shared with a team, are for Business and Enterprise workspaces.
- **Anthropic.** Claude Cowork is listed for the paid plans (Pro, Max, Team and Enterprise) in the desktop app.
- **Microsoft.** Agents are built in Copilot Studio or added to Microsoft 365 Copilot. Its developer page, last updated 30 September 2026, says they can send emails and update records.
- **Google.** The Gemini Spark help page says it is not available in the United Kingdom, and not available on a work or school Google account. Google sells agents to businesses through Google Cloud; we did not read those product pages.

We tested none of these; this is what each page said on the day.

## What to leave to an agent, and what a person approves

The vendors' own advice sorts into two columns.

| Leave it to run | A person approves every time |
|---|---|
| Summarising documents and long email threads | Sending a message to a customer, a supplier or an official body |
| Collecting information into one sheet, with its sources | A payment, a purchase or a refund |
| Drafting replies and quotes into a folder for someone to read | Deleting or overwriting files and records |
| Sorting and renaming files in a folder set aside for it | Anything on a site that holds money or personal records |

The left column follows Anthropic's advice to start with low-risk tasks such as summaries. The right follows its advice to approve by hand when a mistake would be hard to undo, such as sending messages or making purchases, and OpenAI's statement that its agent mode asks for confirmation before high-impact actions. Google's page tells users to avoid tasks that touch anything they consider sensitive.

Six settings for a small office, drawn from those pages:

1. Give the agent one working folder, not the whole drive, and keep a backup of it.
2. Connect only the apps the task in hand needs.
3. Set it to ask before it sends, pays or deletes.
4. Keep it out of banking and government accounts. Anthropic suggests a separate browser profile that has no access to them.
5. Write the task narrowly. OpenAI warns against open-ended instructions to check the inbox and handle everything.
6. Read what a scheduled task produced after each run, and pause the ones nobody reads.

Whoever switches it on stays answerable. Anthropic's page says the user remains responsible for everything the agent does on their behalf. Microsoft's Copilot Terms of Use, effective 18 August 2026, say the user is solely responsible for actions Copilot takes at their request. Anthropic's pages are cited most here because they are the longest of the four, not because its product is safer.

## Where to go next

For the jobs that suit this kind of help, see [our count of 80 office jobs](https://aeolookup.com/ai-for-small-firms/which-office-jobs-can-an-ai-assistant-do-80-jobs-ten-trades-counted/). For wrong output, see [the UK cases and a five-step check](https://aeolookup.com/ai-for-small-firms/what-happens-when-ai-gets-it-wrong-uk-court-cases-five-step-check/). Sector guides cover what may be typed in: [for recruiters](https://aeolookup.com/recruitment/chatgpt-for-recruiters-uk-gdpr-prompts-ai-policy/) and [for accountants](https://aeolookup.com/accountants/can-accountants-use-chatgpt-copilot-claude-client-data/).

Overlyft ![](https://aeolookup.com/img/d/ours-badge.png) is an AI agency that sets up AI office assistants for ten trades: they are set up in the firm's own office for a fixed price agreed after an audit, the firm pays Anthropic directly for its own Claude plan, and a person approves anything that goes to a customer. Product names and availability change often, so read the vendor's page before switching one on.

## Short answers from this page

**What is the difference between an AI agent and a chatbot?** A chatbot answers one message and stops. An AI agent is a model given tools, such as files, email or a browser, and allowed to work through a task in several steps. On 1 October 2026 OpenAI, Anthropic, Microsoft and Google all defined it that way, and each tells users to check or supervise what it does.

---
![](https://aeolookup.com/img/d/owner-line.png)

General information for UK small firms, not legal, tax or financial advice.