An AI agent is software that can read a request, decide what to do about it, and use your tools to get it done. A chatbot talks. An agent also acts.
That one difference is why "AI agents" went from a research term to something every software company is selling in about two years. It's also why the term gets used for almost anything. This guide explains what an agent actually is, how it differs from tools you may already have, and where agents earn their keep in a small or medium business.
What is an AI agent, in plain English?
Think of a new hire on their first week, given one narrow job and access to a few systems. You tell them: "When someone asks about a viewing, reply, offer them a time from the diary, book it and log the lead."
An AI agent works the same way. It has three things:
- A goal: what it's there to do, written in plain language.
- Tools: the systems it's allowed to use, such as your calendar, CRM, email or accounts package.
- Limits: what it must never do, and when it must hand over to a person.
When a request comes in, the agent reads it, works out the steps, uses its tools to carry them out and checks the result. If something doesn't fit, such as a complaint, an unusual request or missing information, it stops and passes the job to a person with a summary.
AI agents vs chatbots vs automation
| Option | Can understand messy requests | Can take actions in your systems | Decides the steps itself |
|---|---|---|---|
| Chatbot | Yes | No | No |
| Automation (e.g. Zapier) | No | Yes | No, it follows fixed steps |
| AI agent | Yes | Yes | Yes, within limits you set |
In practice, the best setups mix all three. Fixed automation handles the predictable steps cheaply and reliably. The agent handles the parts that need reading and judgement. A person handles anything that matters.
Six AI agent examples for small businesses
1. The enquiry agent
Replies to every new lead within a minute, day or night, answers common questions from your own information, and books a call or appointment from your calendar. For most service businesses this is the first agent worth building, because slow replies lose work.
2. The AI receptionist
Handles the questions your front desk answers fifty times a day, such as opening hours, prices, parking and rescheduling, through web chat, email or WhatsApp. Clinics, salons and garages see the biggest difference.
3. The CRM agent
After a sales call, it writes up the notes, updates the deal, creates the follow-up task and drafts the email. Your CRM stays accurate without anyone typing into it.
4. The back-office agent
Reads incoming invoices, receipts and order forms, pulls out the details, enters them into your accounts or stock system and flags anything that doesn't match.
5. The internal knowledge agent
Staff ask questions in Teams or Slack ("what's our returns policy for damaged items?") and get the answer from your own documents, with a link to the source.
6. The reporting agent
Every Monday morning, a short summary of last week: leads, bookings, revenue, overdue jobs and anything that needs a decision.
What AI agents shouldn't do (yet)
Agents are good at the repetitive 80% of a job. They are not good at the 20% that depends on relationships, judgement or accountability. We don't build agents that:
- make decisions with legal, medical or financial consequences without a person approving them;
- handle complaints or upset customers on their own;
- send anything important to a customer before you've seen the agent work reliably for a few weeks.
The research backs this up. In a large study of customer support staff, an AI assistant raised the number of issues resolved per hour by 14% on average, and by 35% for the least experienced staff. The gains came from people working with the AI, not from removing them.
How to keep an AI agent safe
- Least access: each agent can only see and change what its job needs.
- Business accounts: use AI providers' business plans, which don't train on your data by default.
- Approval first: customer-facing messages wait for a person until you're confident.
- Logs: you can see everything the agent did, and why.
- An off switch: pause it any time; the work falls back to your team.
Under GDPR, an agent handling personal data is no different from any other system that does: you need a lawful basis, a processor agreement with the vendors involved, and to keep only what you need.
Where to start
Pick the job your team repeats most and likes least. For most businesses that's answering enquiries or booking appointments. Build one agent for that, keep a person approving its replies for a few weeks, then widen what it's allowed to do.
If you'd like help choosing, our AI agents service starts with exactly that conversation, and the AI capacity calculator gives a rough idea of the hours involved.