AI agents for small business: what they are and how they work, with 5 examples

AI agents for small business sound like something only big companies can afford. We run one at Adaxit, though, and it’s nothing grand. Every weekday morning an agent on Claude reads the inbox, sorts it, drafts replies and leaves me a summary that fits on one screen. It never sends anything on its own: I read, fix what needs fixing and decide. Still, it gives a fair idea of what an agent can already do in a company of one or two people, and where it should stop.

So what is an AI agent, and how does it work?

Don’t go looking for one neat definition; even the European Commission’s AI Act Service Desk says the term gets used loosely. Most agree on one thing, though. An agent takes in input from its environment and can act on it, say by calling a function in another program. The UK’s National Cyber Security Centre (NCSC) puts it this way: agentic systems “can access data sources, remember context, make decisions, use tools, and take actions in pursuit of a goal”.

1. A goal and some rules. “Sort the inbox into four groups and draft replies the way we write.”
2. Tools. Email, spreadsheets, documents, web pages. Only the ones you open up to it.
3. A loop. It works in small steps, checking each result before the next move.
4. A stopping point. You set it. A draft waiting for approval, say, or a summary to read.

Under the hood there’s usually a general-purpose AI model, the same kind you chat with. And the AI Act? It treats agents like any other AI system.

Is an AI agent really different from a chatbot or classic automation?

Talk of AI agents for businesses lumps three things together. A chatbot answers you inside a conversation and stops there. Classic automation is the “when an invoice comes in, save it to this folder” sort of thing: fixed rules, dependable right up until something unexpected turns up and it stalls. An agent gets a goal and picks its own steps and tools. It can deal with the unexpected, which is exactly why it needs closer watching.

	Chatbot	Classic automation	AI agent
How it works	Replies in the chat	Runs fixed rules	Chooses steps and uses tools
Unexpected cases	Answers, but does not act	Stops or fails	Tries to handle them, can get it wrong
The person’s role	Reads the answer	Sets the rules once	Approves what the agent produces
Example	“Summarise this text”	“Save attachments to a folder”	“Sort the inbox and draft replies”

AI automation won’t send the old kind into retirement, mind you. Use fixed rules where the process never changes and an agent where something has to be read and understood. Want to build a tool of your own without developers? Here’s how to build an MVP with AI and no code.

How many small businesses are already using AI?

Fewer than the noise suggests, but climbing fast. By Eurostat’s count, 20% of EU enterprises with 10 or more employees used AI in 2025; a year earlier it was 13.5%. Italy is a bit behind. Istat puts it at 16.4% of firms with at least 10 employees, with a wide gap by size: 15.7% for small and medium firms, 53.1% for large ones.

The top use across the EU? Analysing written language, at 11.8% of enterprises. That’s agent territory.

What can AI agents for small business do right now?

Why these five? They come round every week, they’re mostly text, and you can check them in a minute.

1. Sorting the inbox and drafting replies

Each morning the agent reads what came in, splits it into clients, suppliers, admin and noise, and drafts a reply wherever one is needed. It’s the one we run at Adaxit. You approve: every reply, before it goes out.

2. Finding prospects with a public need

Plenty of organisations say publicly what they need: a tender, a job ad, a post looking for a supplier. On TED, the EU’s public procurement portal, more than 3,000 notices go up every weekday, and that’s just one source. An agent can read the lot for you, flagging whatever matches what you sell, with a line on why. Writing to them is where cold-email rules kick in, wherever you sell. In Italy promotional messages need prior consent, even if someone’s details are public online. In May 2026 the data protection authority fined a company for marketing calls and emails sent without consent to contacts pulled from a data platform. The bigger picture is in our go-to-market strategy guide. You approve: who gets contacted, and how.

3. A weekly numbers report

Every Monday the agent goes through invoicing, the CRM or your sales sheet and writes you one page. What moved? What doesn’t add up? What do you need to decide? Choose the numbers worth tracking first, CAC and LTV for instance. The same page can feed your investor update. You approve: the decisions, once you’ve checked the key figures yourself.

4. First drafts of quotes and proposals

After a call, drop your notes in. The agent writes a first draft using your templates and price list. If you’re a professional in Italy, there’s one more rule: Law 132/2025, in force since 10 October 2025, limits your use of AI to a supporting role and requires you to tell clients. You approve: prices, terms and anything you’d be committing to. Always.

5. Answering customers from your own FAQ

Here the agent answers common questions from your own material alone (the FAQ, terms of sale, product sheets). If the answer isn’t in there, the case goes to a person. If you ever let it answer alone, customers have to know they’re talking to an AI. You approve: each reply before the customer sees it.

Our principle: a person signs off on everything that goes out. The agent reads, sorts and drafts. A person decides what leaves the building, whether it’s an email, an offer, a number or a reply to a customer. You get the agent’s work without handing it the consequences. The NCSC says much the same: keep real human oversight and give the agent only the access it needs.

What will an AI agent set you back?

You can start with one subscription to a tool that has connectors and scheduled tasks. Here’s where US list prices stood when we checked, on 30 October 2026:

• Claude Pro (Anthropic) is 20 dollars a month before tax, or 17 if you pay yearly. Connectors, skills and scheduled tasks are included. Team is 25 dollars per user per month (20 yearly).
• ChatGPT Business (OpenAI) wants at least two users, at 25 dollars each per month, or 20 if you’re billed yearly.
• Microsoft 365 Copilot Business bolts onto Microsoft 365 and starts at 21 dollars per user per month, provided you sign up for a year. Already a customer on an eligible plan? Until 31 December 2026 you can get it for 18 dollars in the first year.

Through an API you pay for what you use instead, based on the tokens processed.

What gets left out of the sums is your own time. Someone has to write clear instructions and hook up the tools, then review drafts for the first few weeks; budget for it. I’d be wary of anyone who promises you guaranteed savings.

What could go wrong, and which EU rules apply?

There are two risks I’d keep an eye on.

• Mistakes. It might misread a request or get a figure wrong, which is why a human looks at everything.
• Hidden instructions. An email or a web page can hide text written to steer the agent somewhere it shouldn’t go. It goes by the name prompt injection, and the NCSC warns it may never be fully fixed. So you limit the damage instead: fewer permissions, less harm. And no open access to sensitive data or critical systems.

The AI Act. It entered into force on 1 August 2024 and applies in stages, with a risk-based approach. The European Commission says most AI systems used in the EU, spam filters for instance, are minimal risk and fall outside its rules. Three points are worth knowing in a small company:

• Article 50 has applied since 2 August 2026. If people interact with an AI system, they have to be told at the start, unless it’s obvious.
• High-risk obligations, which cover things like sorting CVs for hiring and credit scoring, start on 2 December 2027. That’s later than first planned, after the Digital Omnibus on AI (Regulation (EU) 2026/1744, in force since 27 July 2026).
• Article 4 asks companies to support the AI literacy of staff who use AI. The Omnibus watered the wording down; I’d still train people a little.

The GDPR. Your emails are full of personal data, so the GDPR applies too:

• only process what you need for the purpose;
• if a vendor processes data for you, get a contract in place; you’re on the hook for their compliance too;
• decisions made purely by automated processing, if they affect people legally or in a similarly significant way, are only allowed in limited cases, with safeguards such as human intervention.

How would I get started in one week?

I’d follow the NCSC here: go step by step, with small pilots and clearly defined tasks.

• Day 1: pick one process you repeat a lot and can check easily, like the morning inbox.
• Day 2: get the rules on paper. Categories, tone, a few good past replies as examples, and the things it must never do: send, pay, delete.
• Day 3: connect the tools with the least access possible. Reading and drafting, yes. Sending, no.
• Days 4 and 5: run it alongside your normal work and jot down every mistake.
• Day 6: rework the instructions, or add examples, wherever it got things wrong.
• Day 7: make the call. Keep it, change it or switch it off, depending on how long reviews took and how often it slipped.

This page is general information, not legal, tax or investment advice. Rules and terms change: check the official source or ask a professional before you act.