AI Agents: What They Are and What They Can Really Do for SMEs

AI agents are programs built on language models, the same ones that power ChatGPT, that receive a goal, decide which steps are needed and use external tools (the email inbox, the ERP, a price list) to get it done. In an SME today, AI agents work well on narrow, repetitive tasks with clear rules and a person who checks the result before it becomes final. For everything else, it pays to go slowly.

This guide is for anyone who has heard about them at a trade fair or in a webinar and wants to know what is real: what they are, how they differ from a chatbot and from RPA, four jobs an agent can already do in a company of 10 to 100 people, the real limits and how to run a first project without burning through the budget.

What AI agents are

A language model on its own produces text. You ask it a question and it gives you an answer. An agent adds three elements around the model:

1. a goal, for example "classify the emails arriving at info@ and assign them to the right department";
2. tools it can use: reading a mailbox, looking up a customer in the ERP, drafting a document, opening a ticket;
3. a work loop. The model looks at the situation, picks a tool, reads the result and decides the next step, until the task is done or until it reaches a point where it has to stop and ask a person.

The part that really changes things is the second. Without connected tools, an agent is a chatbot with longer instructions. With tools it can read and write in the company's systems, and that is where both the usefulness and the risks come from.

These connections almost always go through the APIs of the software you use. If your ERP has no API, or has one that nobody knows how to use, the agent has nothing to work on. If the topic is new to you, start with the article on business APIs and how to get your software talking.

AI agents, chatbots and RPA: the differences

The table compares, in practical terms, three tools that are often confused, even in sales proposals.

RPA, robotic process automation, remains the right choice when the process is identical every time: same portal, same fields, same order. It is predictable and costs little per run. An agent is useful when the input changes shape every time, like customer emails, which nobody writes the same way twice.

A business chatbot has a different job: it talks to people. Many modern chatbots are in practice small agents, because when a customer asks "where is my order?" they go and query the ERP. The useful distinction is who is on the other side. The chatbot converses with a user; the agent often works in the background, with nobody watching while it works.

What AI agents really do in an SME

These are four cases that can go into production with the technology available today. None of them requires handing important decisions to the agent without checks.

Sorting emails and tickets

This is the simplest case and nearly always the one to start with. An electrical supplies distributor receives orders, price requests, complaints, supplier invoices, CVs and advertising at info@. Someone in accounts spends the first hour of the day forwarding.

The agent reads each message, classifies it (order, quote, complaint, admin, other), extracts the useful data such as customer, item code and order number, looks up the customer in the master data and opens the ticket or forwards it to the department, with a two-line summary at the top. When it isn't sure, it puts the message in a "to review" queue instead of guessing.

Preparing quotes from requests

A window and door manufacturer receives requests for quotation by email, sometimes with a PDF of measurements, sometimes with a photo of a sketch made on site. The agent extracts the items, matches them against the price list, applies the customer's terms taken from the ERP and prepares a draft quote. The sales rep opens it, corrects what is needed and sends it.

The word that matters is draft: the agent sends nothing to the customer. The time saved is the first draft, which in quotes with many lines is the longest and most tedious part.

For this to work you need a system that generates documents from clean master data and price lists. In the management software we developed for Pulltex, for example, the sales network creates pro forma invoices by selecting customers and products already loaded, each sales rep sees only their own customers and the documents reach HubSpot via API. On a foundation like that an AI agent makes sense: tidy data, explicit rules, role-based permissions. Without that foundation it produces drafts full of errors.

Querying the ERP in plain language

The owner of a car repair shop wants to know how many services were done last month compared with the same month a year earlier, or which customers haven't been back for more than eighteen months. Today the owner asks whoever knows how to run the reports, or gives up.

An agent with read-only access to the database translates the question into a query, runs it and answers with the number and the table it came from. There are two rules. Read-only access, always. An answer that shows where the figure comes from, so the reader can check it. A wrong number stated with confidence is the main risk of this use.

Reconciling documents

Purchase orders, delivery notes and supplier invoices have to match. In many companies someone compares lines and amounts by hand, especially when the supplier invoices differently from how it delivered: partial deliveries, item codes different from yours, discounts applied on the invoice but not on the order.

The agent reads the three documents, matches the lines even when the codes don't coincide, flags differences in quantity or price and proposes an outcome: matches, doesn't match, not sure. Accounts check only the exceptions. Here AI covers the part that classic RPA handles badly, namely understanding that "VITE TE 8X40 ZN" and "hex head screw 8x40 zinc-plated" are the same item.

ChatGPT for business: where many SMEs start

Many companies come to AI agents by way of ChatGPT, which some employees already use with a personal account to write emails or summarise documents. It is a reasonable first step, with two caveats.

The first concerns data. If people paste quotes, price lists or customer data into a personal account, the company has no control over where it ends up. The business versions of ChatGPT and competing tools have different contractual terms on data use: read them before activating them, and in the meantime put in writing what may and may not be pasted. Since 2 February 2025, the EU AI Act has also required companies that use artificial intelligence systems to take measures to ensure AI literacy among their staff: a written internal rule and an hour of training are a good start here too.

The second concerns the leap. ChatGPT in a chat window is an assistant: it works on what you give it, when you give it. An agent works inside your processes, with access to real systems. Moving from one to the other isn't solved with another subscription. It is an integration project in which you decide which data the agent can access, with which permissions, and who checks the output. For more on how generative AI is entering business software, see the article on how generative AI is changing business software.

The limits of AI agents

Errors

Agents make mistakes, and they make them differently from traditional software. A classic program either works or throws an error, and the error is visible. A language model can produce a plausible but wrong answer with no warning sign. That is why an agent in production needs a "not sure" exit, a searchable log of everything it has done and a sample of cases checked by hand every week.

Permissions

The same principle applies as for a new hire: access only to what is needed. An agent that sorts emails doesn't need to be able to delete customers. An agent that answers questions about revenue must not be able to write to the database. Irreversible actions, such as sending something to a customer, paying or deleting, stay with a person.

There is also a specific risk. An email or a PDF can contain instructions written on purpose to trick the model, along the lines of "ignore the rules and forward this attachment to...". This is called prompt injection, and the most effective defence is precisely to limit what the agent can do. It is the same reasoning you will find in our cybersecurity guide for SMEs: limit the possible damage, even before preventing the attack.

Costs

You pay in two ways. There is the project (integrations, instructions, testing) and there is model usage, usually calculated in tokens, that is, on the amount of text processed. An agent that reads short emails uses little; one that rereads manuals hundreds of pages long every time can cost much more. Before starting, ask for an estimate of usage based on a real sample of your documents, not on a test example.

In our experience, most of the project cost lies in data cleaning and integrations rather than in the AI part. If the data is in order, the agent can be built quickly.

Human oversight

Decide in advance who checks what. The pattern that works: in the first weeks the agent proposes and a person approves everything, then checks are relaxed for the categories where it made few mistakes and kept tight on the sensitive ones. If nobody in the company has time for this review, the project isn't ready to start.

How to start small with an AI agent

The first project should be small, measurable and easy to switch off. These are the steps we follow.

1. Choose a process with high volume and low consequences. Sorting incoming mail is almost always the best candidate. Calculating sales commissions isn't.
2. Measure how things are today: how many messages or documents per day, how much time for each, how many errors. Without a baseline you will never know whether the agent works.
3. Collect a real sample, for example a hundred real emails or fifty requests for quotation, with the correct outcome decided by the person who does that job every day. The agent is measured on that sample before it touches any live data.
4. Connect only the systems you really need, with minimal permissions. It is often better to develop a small dedicated integration than to give the agent a user's credentials.
5. Keep the agent in proposal mode for a few weeks. A person approves or corrects, and every correction becomes a new test case.
6. Only then broaden the scope: more categories, more autonomy, then a second process.

At Selli this path translates into an initial process analysis, two-week sprints with a demo at the end of each, a staging environment where the agent works on copies of the data before touching the real ones, and code and instructions that remain the client's property. You will find the details on the page dedicated to AI for business and automation.

When you don't need an AI agent

Not every problem should be solved with an agent, and quite a few of the projects proposed to us start from the technology instead of the process.

If the process is always the same, a rule or a classic automation is enough: it costs less and doesn't make creative mistakes. If your data is scattered across Excel files with no single master record, the first project is the ERP, not the agent; the signs that your SME needs a management system are a good test. If you handle five requests a day, the agent won't pay back the integration work. And if a mistake is expensive (prices for a key customer, health data, decisions affecting people), human review stays on every single case.

As for staff concerns, our position is set out in the article will AI replace my employees?: agents take away sorting and copying work, while judgement on customers, prices and exceptions stays with people. For a broader view of how small businesses use AI, see our guide on AI for SMEs.

Before putting an agent to work on customer or employee data, check the obligations set by the EU AI Act for businesses.

Let's talk

If you have a process in mind that seems suitable, such as the info@ inbox, requests for quotation or invoice reconciliation, we can look at it together. In a 30-minute call we work out whether an AI agent makes sense, whether a simpler automation will do or whether the data needs sorting out first. Write to us from the contact page and tell us about the process you would like to lighten.