What an AI agent is, explained for small business owners

"Agent" has become a marketing word. Software vendors use it for everything from a website chatbot to a system that supposedly runs your whole back office. That makes it hard for an owner to judge what is on offer.

This article explains what an agent is in plain terms, how it differs from a chatbot and from ordinary automation, where it helps a small business, and where a simpler tool would do the job better.

What is an AI agent, in plain language?

An AI agent is an assistant that can use your tools to complete steps of a task, not just talk about it. You give it a goal, such as "prepare a quote for this enquiry", and it decides which steps to take: look up the customer in the CRM, check the price list, draft the quote, and put it in a folder for review.

The difference from a chat assistant is access and action. ChatGPT in a browser tab can only work with what you paste into it. An agent is connected to your systems and can read from them, and sometimes write to them. That is what makes agents useful, and it is also why they need limits.

Chatbot vs automation vs agent

These three are often confused. The simplest way to tell them apart is to ask who decides the steps.

Type	Who decides the steps	Good for	Example
Chatbot or chat assistant	No steps. It answers questions in a conversation.	Drafting, answering questions, summarizing what you give it	A website chat that answers questions about opening hours and services
Automation	You do, in advance. It follows fixed rules every time.	Predictable, repeated tasks with clear inputs	When a form is submitted, add the contact to the CRM and send a confirmation email
AI agent	The AI chooses steps within limits you set.	Tasks where the inputs vary and some judgment is needed	Read an unstructured enquiry, find the right products and draft a quote for approval

When a simple automation is better than an agent

If you can write the steps down as fixed rules, use an automation, not an agent. Automations are cheaper to run, behave the same way every time, and are easier to test and fix. "When an invoice is 30 days overdue, send reminder template B" does not need AI at all.

Agents earn their place when the input is messy and the right next step depends on reading it: an enquiry written in free text, a supplier email with a changed delivery date buried in the third paragraph, a request that might be a quote, a complaint or both. Many good setups use both. An automation moves the data reliably and calls an agent only for the step that needs judgment.

A vendor who proposes an agent for a task that a rule could handle is adding cost and unpredictability for no gain.

What can AI agents do for a small business?

The most useful agents in small businesses do one narrow job well. Five realistic examples:

• Quote preparation: reads an enquiry, matches it to products or services and past quotes, and drafts a quote for staff to check.
• Invoice follow-up drafts: checks which invoices are overdue, reads any recent correspondence with that client, and drafts a suitable reminder. A person sends it.
• Enquiry triage: sorts incoming messages by type and urgency, fills CRM fields and routes each to the right person. Our article on responding to enquiries faster covers this in detail.
• Meeting preparation: pulls recent emails, open orders and notes about a client into a one-page brief before a call.
• Report compilation: gathers figures from several systems into a weekly or monthly summary, flagging anything unusual.

Example: a marine parts distributor with 15 staff

This is an illustrative example, not a client case. A South Shore distributor of marine engine parts gets quote requests by email from boatyards and fishing fleet operators. Requests are written loosely: "need impellers and a raw water pump for a 2015 Volvo D4, two boats". Staff spend a long time matching these to part numbers.

The agent reads each request, looks up likely part numbers in the product catalogue, checks stock levels in the inventory system, and drafts a quote in the accounting software as a draft, not a sent document. It flags anything it is unsure about. A staff member checks the parts, adjusts and sends. The agent can read the catalogue and inventory, and can create drafts. It cannot send emails, issue invoices or change prices.

Are AI agents safe? Why approvals and access limits matter

Agents are as safe as the limits you put on them. Two limits do most of the work.

Approvals. The agent proposes and a person approves anything that spends money, sends something to a client, or changes or deletes records. Drafts are cheap to throw away. A wrong email to a client, or a changed price in your accounting system, is not.

Access limits. Give each agent access only to the systems and actions its job requires. A meeting-prep agent needs to read email and the CRM. It does not need access to payroll or the ability to send anything. The OWASP Top 10 for LLM applications, a widely used security reference, lists "excessive agency", meaning an AI given more permissions or autonomy than it needs, as one of the main risks.

Other things that go wrong, and the guardrail for each:

• Wrong or outdated data: the agent is only as accurate as your price list and records. Keep the sources it reads current, and test it against real past cases.
• Loops and repeated actions: an agent can retry a failing step many times. Set limits on how many actions it can take per task, and alert a person when it hits them.
• Prompt injection: an email or web page can contain hidden text that tries to give the agent instructions. OWASP ranks prompt injection as the top risk for AI applications. Approvals and narrow access are the main defence: an agent that cannot send email cannot be tricked into sending it.

Keep a log of what each agent did, so you can check its work and trace any mistake.

What do AI agents cost?

Costs usually have two parts: licences for the AI tools, and usage charges that rise with how much the agent does. For example, Microsoft says that while Copilot Chat is available at no additional cost with eligible Microsoft 365 subscriptions, agents that work with your organization's data are billed by usage through an Azure pay-as-you-go subscription. Add setup time and someone to look after the agent as your tools and data change. Our article on what AI automation costs in Canada covers the full picture.

One do-everything bot or a team of specialists?

Several narrow agents, each with one job and its own limited access, are easier to test and safer than one agent that does everything. If the quote agent misbehaves, you switch it off without losing enquiry triage. Each one can also be given exactly the access its job needs. This is the approach we take in our AI workspace and agent team service.

How to start with one agent

1. Pick one task that takes real staff time, has messy inputs, and has an output a person can check quickly.
2. Check whether a fixed-rule automation could do it. If yes, use that instead.
3. List exactly what the agent needs to read and what it may create. Everything else stays off.
4. Decide which actions need approval. Start with all outgoing messages and all record changes.
5. Collect 20 to 30 real past examples and test the agent against them before it goes live.
6. Measure time spent on the task before and after. See how to tell whether an AI project worked.

If the job you have in mind is the owner's own inbox and schedule, our owner assistant is built around that.