Skip to Content

What an AI agent is, what it isn't and where a mid-sized company should start

An agent is a program that does a job inside your systems, within limits someone has written down beforehand. It's worth knowing what you can hand over to it and what you can't.
17 November 2026 by
What an AI agent is, what it isn't and where a mid-sized company should start
OdooBot

‘AI agent’ has become the label for almost anything with a language model inside. It's worth separating out what it really is, because that decides what you can hand over to it.

What an agent is

An agent is a program that is given a goal, decides what steps to take to reach it and carries them out using tools: reading an email, querying the ERP, creating a record, drafting a document. It uses a language model to understand the situation and choose the next step.

Four things set it apart:

  • It has tools. It reads and writes in your systems, as well as holding a conversation.
  • It works step by step. It looks at the result of what it has just done and decides what to do next.
  • It has a brief and limits. It knows what it's there for and what it mustn't touch.
  • It leaves a trail. What it does is logged and can be reviewed.

What it isn't

It isn't a chatbot

A chatbot answers; an agent acts. If you ask about an invoice, the chatbot explains where to look for it. The agent finds it, records it and tells you if anything is missing.

It isn't a traditional automation

An automation always follows the same steps: if an order comes in, create a task. It's the right choice when the process doesn't change, and it's cheaper and more predictable. An agent is useful when something has to be read, interpreted or decided between different cases.

Often what a company needs is good process automation, with no agent at all. And the two are frequently combined: the automation moves the data and the agent handles the cases that don't fit the rule.

It isn't an employee

An agent has no business judgement and isn't accountable for what it does. Someone still has to decide anything that involves money, people or clients.

It isn't always right

A language model can be wrong and sound very sure of itself. That's why an agent's work has to be tightly bounded and someone has to review it, especially at the start.

Where to start

In a mid-sized company, the first agent should meet these conditions:

  • A process that repeats every week and that a person does by hand today.
  • Data in a system. If the information is in someone's head or in scattered spreadsheets, the first project is to organise it. That's what an assistant with the company's knowledgeis for.
  • A cheap mistake. An error that can be spotted and corrected without harm: sorting emails, preparing a draft, matching a bank statement against invoices.
  • A measure. Hours saved, cases resolved, response time. Without one, you won't know whether it has worked.

And an order that works. First, the agent only reads and proposes. Then it acts, and a person approves. Once it has been getting things right for a while, it's left to handle the lowest-risk tasks on its own.

What has to be written down first

Every agent needs the same as a new starter: to know what its job is, what it has access to and what it must never do. If that isn't written down, the agent will do whatever seems reasonable to it, which isn't always what you wanted.

The company, for its part, needs rules on using AI: which tools are approved and what data can be given to them. We cover this in Safe AI and trainingis for.

How we do it in-house

At SolveIT we work with our own agents. One keeps the accounts up to date inside Odoo, another coordinates projects and another answers with what the company already knows.

None of them pays, deletes or makes decisions about anyone on the team. They get things ready and alert whoever decides. Before an agent is handed a task, it does it alongside the person for a while and we compare the results.

If you'd like to see which process in your company an agent could take on, start with AI agents or tell us about your caseis for.

Shall we look at your case?

Tell us what you have today and what you would like to have.