What is an AI agent, and what can it do for a small business?

"AI agent" is used today for very different things, from a simple chatbot to software that works on its own for hours. Below is the definition used by the people who build agents, how an agent differs from a chatbot and an automation, examples for a small business and the rules that keep it under control.

What an AI agent is

An AI agent is a language model, like the ones behind Claude or ChatGPT, that gets a goal and access to tools: email, a calendar, a spreadsheet, your invoicing software, web search. The model decides which tool to use, looks at the result and picks the next step, until the task is done or it needs a person.

Anthropic, the company behind Claude, draws a useful distinction:

• Workflows are systems where the AI model and the tools are connected through steps written in advance in code.
• Agents are systems where the model directs its own steps and its use of tools.

An example: "Find three free slots next week for a meeting with client X and propose them." The agent checks the calendar, reads recent emails with the client to see what they prefer, drafts the proposal and leaves it for approval. Nobody wrote down in advance in what order to do that.

Chatbot, automation or agent

Anthropic recommends finding the simplest solution that works and adding complexity only when it's needed. Agents suit open-ended problems where you can't predict how many steps it will take. For everything else, an automation is cheaper, faster and easier to check. Examples of automations are in the guide to AI automation for small businesses.

How widely businesses use AI

The US Census Bureau found that between mid-December 2025 and early May 2026, between 17% and 20% of US businesses used AI in some business function, and 20% to 23% expected to within six months. Use rises with size: 37% of firms with 250 or more employees used AI, against under 20% of firms with four or fewer. In the EU, Eurostat reports that 20% of businesses with 10 or more employees used AI technologies in 2025, up from 13.5% a year earlier.

What an agent can do in a small business

Quote requests. The agent reads the request from email or a form, looks up prices and stock, checks the client's previous orders and writes the quote. A person reviews and sends it.

Email. The agent sorts incoming messages, drafts replies to routine ones, puts requested meetings on the calendar and flags what's urgent. A person sends.

Rescheduling with exceptions. A client wants to move an appointment, but the requested day is full. The agent looks for alternatives, takes the client's preferences into account and proposes them.

Research and comparisons. "Find suppliers of compostable packaging in our state and compare prices." The agent searches, reads the pages and writes a summary with sources.

Questions about your data. "How much did we sell in September compared with August, by product?" The agent queries your sales system and answers with a table.

Answers from your documents. A new hire asks how returns are handled; the agent searches your procedures and answers with a reference to the document it read.

In every example, the agent prepares and a person decides. As the agent proves reliable at a task, it can be given more freedom.

How an agent is built

An agent has four parts:

1. The AI model, such as Claude or GPT, used through an API.
2. The instructions: what it does, what it doesn't, when it stops and asks for help.
3. The tools: connections to your business software, each with limited permissions. The Model Context Protocol (MCP), the open standard Anthropic launched in 2024, has become the usual way to connect them; since December 2025 it has been run by the Agentic AI Foundation, under the Linux Foundation, and it's used by ChatGPT, Gemini and Microsoft Copilot as well.
4. The log: everything the agent did, step by step, so it can be checked.

For simple tasks, platforms such as n8n and Make already have building blocks for AI agents. For tasks tied closely to your business software, the agent is written in code.

The risks, and how to keep them in check

OWASP, the organisation that publishes the reference lists for application security, ranks prompt injection first among the risks of applications built on language models: text the agent reads, in an email, a document or a web page, that gives it new instructions. Sixth is excessive agency: an agent with more permissions than it needs. Anthropic adds that an agent's autonomy means higher costs and errors that can compound from one step to the next.

The rules that keep the risks small:

• A person approves anything that can't be undone: sending an email, a payment, deleting data, an order to a supplier.
• Minimal access. The agent that writes quotes reads the price list; it has no access to bank accounts.
• A separate account for the agent, not a person's account, so it's clear what it did.
• Spending limits on model usage, so an agent stuck in a loop doesn't run up a large bill.
• Testing on real cases before use, and regular reviews of the log.

Your data and the law

• Privacy. If the agent handles personal data (customers, employees), check your AI provider's terms: whether data sent through the API is used for training, where it's stored and for how long. In the EU, GDPR requires a data processing agreement with the provider.
• Saying it's AI. If the agent talks to customers directly, by chat or email, California law makes it unlawful to use a bot to mislead people about its artificial identity in order to sell, and in the EU, Article 50 of the AI Act has required since 2 August 2026 that people be told they're interacting with an AI system.
• Responsibility. What the agent sends on your business's behalf is sent by your business. That's why human approval matters.

How to start

1. Pick one task that repeats, has volume and produces a result that's easy to check.
2. Try an automation first. If the steps can be written down in advance, you don't need an agent.
3. Let the agent only propose for the first few weeks; a person approves everything.
4. Give it minimal access and a separate account.
5. Measure the time saved and the mistakes caught in review.
6. Expand only what works.

We build AI agents and automations connected to the software your business already uses, as part of our AI automation work, with human approval wherever an action can't be undone.