AI agents for business: what they are, with everyday examples

Here we explain, in plain terms, what an AI agent is and how one is used in a Portuguese company of 10 to 250 people, with five examples from ordinary work, among them a quote request and the monthly sales report. We also set out what, in the way we work, an agent never does without a person.

What an AI agent is, in plain terms

An AI agent is software that reads a request, uses some of the company's systems through connections it has been given permission to use, carries out a few steps and hands a draft or a prepared action to a person.

The request might be an email, a message from the website or a question from a colleague, and the systems are the ones the company already has, such as the mailbox, the document folders, the ERP and the CRM. Each connection is a specific, limited piece of access that the company authorises, for instance reading one inbox or looking up one price list. The steps are few and known in advance, and the result waits until someone reads it.

Behind the agent is a language model, the same kind of technology that runs chat assistants, and it is this model that understands the text of the request and drafts the reply. The agent connects it to the company's systems and has it follow a sequence of work.

How does it differ from a chatbot?

A chatbot answers inside the conversation. It can explain, summarise or draft a piece of text, and whatever happens next is left to the person who asked, who has to copy the answer, open the ERP, check the price and paste it all into an email.

An agent is given permission to use a few of the company's tools, which lets it fetch the information it needs and leave the work ready where the team will use it, for instance a draft in Outlook or a PDF proposal in the company's own template. With that access, an agent calls for more care than a chatbot, and we explain why further down.

AI agents for business: examples from everyday work

The examples below are illustrative, and we have kept them small on purpose: each deals with a task that comes up several times a week and ends with a person making the decision.

When an email arrives in the orders inbox asking for a quote on 400 cardboard boxes delivered to Braga, the agent works out the item, the quantity, the place and the date, checks the price list and the production plan in the ERP, drafts the reply and prepares the proposal in the company template. The draft then waits, and goes out only after someone on the team has read it and confirmed the terms.

Enquiries that come in by email and through the website are sorted by type and passed to the right person, with a first reply already drafted in Portuguese, English or Spanish, and whoever receives one starts by reading and adjusting that draft.

For follow-ups to clients, the agent checks the CRM for proposals that have had no reply within the agreed time and, on the right day, prepares follow-up emails for the salesperson to approve. Because it records each contact, the client list stays in order.

An agent can also find a procedure or a past answer. Someone asks, in ordinary words, how a return is handled or what price a client was given last year, and the agent searches the procedures, the price lists and old emails, then answers with the source shown so that the person can check it without interrupting the colleague who knows.

For the monthly sales report, the agent pulls the figures from the ERP at the start of the month, fills in the company's report template and writes a first comment on what has changed. Someone in finance checks the figures before the report goes round.

What an agent should not do on its own

There are two things we do not let an agent do without a person. The first is sending anything to clients: the agent prepares the message and leaves it for a person on the team to read, correct and send. This takes them a few minutes and means every reply has been seen by someone in the company before the client receives it.

The other is deciding about people, so we do not build tools that screen job applicants or assess the creditworthiness of individuals, which the AI Act, Regulation (EU) 2024/1689, treats as high-risk uses. Annex III, point 4(a), covers 'AI systems intended to be used for the recruitment or selection of natural persons, in particular to place targeted job advertisements, to analyse and filter job applications, and to evaluate candidates'. Point 5(b) covers 'AI systems intended to be used to evaluate the creditworthiness of natural persons or establish their credit score'.

What these agents are built with

Agents like the ones above are built with tools such as Microsoft Copilot Studio and Power Automate, Google Workspace with Gemini, n8n, or the Claude and GPT models used through an API, which is the technical connection that lets another program use the model. Where data should not leave the company, the agent can use an open model installed on the company's own server.

The choice depends on the task and the data involved, as well as on the systems the company already pays for. Our technology page explains which parts we use and where the data stays.

Where to start

The first agent should be modest, and four rules are worth following when choosing it and setting it up.

1. One task that comes up often, with a result someone on the team can judge in a minute, such as a quote request or sorting the orders inbox.
2. Limited permissions: the agent can reach only the inbox, folders and data that task needs, and it only reads where it has no need to write.
3. A person reviewing everything that leaves the company.
4. A record of what the agent did, showing what it read, what it prepared and who approved it, because that record is where you find out what happened when something goes wrong.

In the Start programme, over 30 days, the team takes the workshop, AI in daily work, and ends up with two tools working on the systems the company already uses, with support while it gets used to them. If you are still working out where your company should begin with AI, the page on AI adoption puts the first questions in order.