AI agents for business are the next step after chat assistants: AI that does things, not just writes them. This lesson explains what an agent is, how much freedom to give one, and where a person must stay in charge.
In this lesson
- The difference between an assistant that answers and an agent that acts.
- The parts of an agent: a goal, a model, tools, memory and a loop.
- Five levels of autonomy, from “answers questions” to “acts alone inside set limits”, with a small-business example of each.
- How agents connect to your software, including the Model Context Protocol, in one paragraph.
- What agents are good at today, where they stumble, and how to start safely.
What is an AI agent? Chat answers, agents act
Everything in Lessons 1 to 6 was mostly about AI that answers: you ask, it drafts, you decide what to do with the draft. An agent goes a step further. It is given a goal, access to tools, and permission to take several steps on its own to reach that goal, checking its progress as it goes.
Ask a chat assistant “find me a time next week for a call with Sam,” and it might suggest how to word an email. Ask an agent the same thing, with access to your calendar and email, and it can look at your diary, write to Sam, read the reply, and put the call in both calendars.
That is the whole difference, and it matters for one reason: chat mistakes stay on the screen until a person acts on them; agent mistakes have effects. A wrong draft is a nuisance. A wrong email sent, payment made or record deleted is a real problem. Everything in this lesson follows from that.
The anatomy of an agent
Every agent, however it is sold, has the same parts.
You give the agent a goal and instructions. The model works in a loop: plan the next step, act with a tool, check the result, and repeat. It ends either done, reporting back to you, or by asking a person when it is stuck or unsure. Its tools are email, calendar, web browser, spreadsheet and accounting software; reading and looking up go straight through, but an approval gate means a person approves any action that sends, pays or deletes. Memory sits underneath: notes it keeps between steps and files you let it read.
- A goal and instructions. What you want done, and the rules: what it may and may not do, when to stop, when to ask.
- A model. The same kind of model as in a chat assistant (Lesson 1), doing the thinking: working out the next step.
- Tools. The software it can use: email, calendar, a web browser, a spreadsheet, your accounting or booking system. Using them is called tool use. An agent can only do what its tools let it do.
- A loop. Plan a step, act with a tool, check the result, then plan the next step, until the goal is met or it gets stuck.
- Memory. Notes on what it has done so far in this task, and sometimes what it has learned from earlier tasks.
- A person. At the points that matter, it stops and asks. In the diagram, the amber gate before anything that sends, pays or deletes is where your approval goes.
Agents use more computing than chat, because each step is another round with the model. Anthropic, which makes Claude, reported that “agents typically use about 4× more tokens than chat interactions” in its data (Anthropic). That shows up in price, especially on usage-based plans (Lesson 11).
How agents connect to your software
To act, an agent needs a way into your systems. Some products build those connections themselves: an assistant inside your email or accounts can already see them. Others use a shared standard. The best known is the Model Context Protocol (MCP), which its own site describes as “an open-source standard for connecting AI applications to external systems,” adding: “Think of MCP like a USB-C port for AI applications” (modelcontextprotocol.io). Anthropic introduced it in November 2024 (Anthropic) and in December 2025 gave it to the Agentic AI Foundation, “a directed fund under the Linux Foundation, co-founded by Anthropic, Block and OpenAI” (Anthropic). For a business owner, the point is simple: if your software offers an MCP connector, many different AI tools can use it, and you decide which ones and with what permissions.
Five levels of autonomy
How much freedom to give an agent is your choice, and it should depend on the task, not on what the product can do. A useful way to think about it is a ladder.
- 0Answers questions"What is our refund policy?" answered from the staff handbook.Youask, then decide what to use.
- 1Drafts for you to sendWrites replies to online reviews.Youedit and press send.
- 2Acts after you approveBooks a supplier delivery once you click approve.Youapprove each action.
- 3Acts and tells youFiles receipts into the accounts and sends a daily summary.Youreview what it did and undo mistakes.
- 4Acts alone inside set limitsReorders stock under a set amount from approved suppliers.Youset the limits and check its record regularly.
- Level 0: answers questions. It can look things up in your information and answer, but takes no action. Example: a staff assistant that answers “what’s our refund policy?” from the handbook. The person does everything.
- Level 1: drafts for you to send. It prepares the email, quote or invoice; you review and send it yourself. Example: an agent that reads a customer enquiry and drafts a quote from your price list. The person checks and acts.
- Level 2: acts after you approve. It proposes an action and does it once you click yes. Example: a scheduling agent that proposes tomorrow’s rota and sends it to staff when the manager approves. The person approves each action.
- Level 3: acts and tells you. It acts on its own and reports what it did, so you can step in. Example: an agent that sends appointment reminders and tells you who has not confirmed. The person reviews afterwards and spot-checks.
- Level 4: acts alone inside set limits. It acts without reporting each step, within firm limits: a spending cap, a list of allowed actions, certain customers only. Example: an agent that reorders stock automatically below a set level, up to a set amount per week. The person sets the limits and audits regularly.
Most small businesses should start every agent at Level 1 or 2 and move up only when the logs show it gets the task right, reliably, over weeks. Anything that sends money, signs, deletes or speaks for the business publicly should stay at Level 2 unless there is a very good reason.
What agents are good at today, and where they stumble
Good fits are multi-step tasks that follow a pattern and happen in software:
- admin across several systems, such as copying details from emails into your customer records and booking system;
- research, such as gathering information on a list of suppliers or prospects and summarizing it;
- data entry and tidying, such as matching invoices to orders and flagging the ones that do not match;
- routine follow-ups, such as chasing missing documents or confirming appointments.
Where they stumble:
- Long tasks. The more steps, the more chances to go wrong, and an early mistake can carry through.
- Unclear instructions. People fill gaps with common sense; agents fill them with guesses.
- Unexpected screens. A changed website, a pop-up or a login can throw an agent that works in a browser.
- Being tricked by what they read. An email or web page can contain hidden instructions that the agent follows. This is called prompt injection, and Lesson 10 explains how to guard against it: limit what the agent can do, and approve actions.
- Knowing when to stop. Agents can keep going when a person would have stopped and asked.
AI agents for business: where you’ll meet them
You may never build an agent. More likely, you will meet them inside products you already use or are offered: assistants that can now take actions, and agent features in office, accounting, help desk and booking software. The main AI makers’ agent features, as each describes them on its own pages, with the date we read them:
| Maker | Assistant | Agent features, as the maker describes them | Date read |
|---|---|---|---|
| Alibaba (Qwen) | Qwen Chat | Not stated on the pages we read. | October 6, 2026 |
| Anthropic | Claude |
| October 6, 2026 |
| DeepSeek | DeepSeek (chat at chat.deepseek.com) | DeepSeek Harness: An agent harness for developers, in developer preview; the API docs also list guides for running DeepSeek models inside other makers' coding agents. | October 6, 2026 |
| Gemini (Gemini app; Gemini Enterprise app for business) |
| October 6, 2026 | |
| Meta | Meta AI |
| October 6, 2026 |
| Microsoft | Microsoft Copilot (formerly Microsoft 365 Copilot); Microsoft Copilot Chat (formerly Microsoft 365 Copilot Chat) |
| October 6, 2026 |
| Mistral AI | Mistral Vibe (formerly Le Chat) |
| October 6, 2026 |
| OpenAI | ChatGPT |
| October 6, 2026 |
| xAI | Grok |
| October 6, 2026 |
Features change quickly. Treat this as a snapshot and check the vendor’s own page.
How to start
- One agent, one task. Pick a task from your Lesson 4 list that is frequent, in software and easy to check.
- Write the goal and rules down: what it may do, what it must never do, when to stop and ask.
- Give it only the access it needs. A read-only connection where possible; no payment or delete permission unless the task requires it.
- Approve every action at first (Level 2), and keep the log.
- Loosen slowly. After a few weeks without mistakes, let it act and tell you (Level 3) for the lowest-risk part of the task. Keep the approval gate for anything that sends, pays or deletes.
- Name the person responsible, on your team chart from Lesson 5, and in your AI use policy from Lesson 10.
Key terms
- AgentAn AI given a goal, tools and permission to take several steps on its own, such as booking, filling in forms or updating records.Because agents act, their mistakes have effects. Start with a person approving every action.Lesson 7: From chatbot to agent · More in the glossary
- Tool useWhen an AI uses other software to do something, such as searching the web, reading a calendar or sending an email.More in the glossary
- Model Context Protocol (MCP)An open standard for connecting AI applications to other software and data, so one connector works with many AI tools.Anthropic introduced it in November 2024 and gave it to the Agentic AI Foundation, under the Linux Foundation, in December 2025.modelcontextprotocol.io · More in the glossary
- Prompt injectionTricking an AI with instructions hidden in what it reads, such as an email or web page, so it does something it should not.Limit what an agent can do, and have a person approve anything it sends, pays or deletes.OWASP: Prompt injection · More in the glossary
- Human in the loopA named person who checks or approves AI’s work before it takes effect, and answers for the result.Lesson 10: Risks and rules · More in the glossary
- TokenA small piece of text, roughly a short word or part of one, that a model reads and writes. Usage is often priced in tokens.More in the glossary
Every term is explained in the glossary.
Try this
Take one multi-step task from your list, such as booking a new customer’s first appointment. Write down every step a person takes, every piece of software they touch, and every point where they make a decision. Then mark which steps an agent could do, which need a person’s approval, and which should stay with a person. That is the design for your first agent, and you can hand it to any vendor or developer.
Sources
- Model Context Protocol, “Introduction” (2026-07-28 docs). Read 6 October 2026. https://modelcontextprotocol.io/docs/getting-started/intro
- Anthropic, “Introducing the Model Context Protocol”, 25 November 2024. Read 6 October 2026. https://www.anthropic.com/news/model-context-protocol
- Anthropic, “Donating the Model Context Protocol and establishing the Agentic AI Foundation”, 9 December 2025. Read 6 October 2026. https://www.anthropic.com/news/donating-the-model-context-protocol-and-establishing-of-the-agentic-ai-foundation
- Anthropic, “How we built our multi-agent research system”, 13 June 2025. Read 6 October 2026. https://www.anthropic.com/engineering/multi-agent-research-system
- Vendor agent features: each maker’s own pages, as listed with their dates in our vendor data.