AI agents in business – how they automate processes and where a mid-sized company should start

A chatbot answers questions. Rule-based automation, for example in Power Automate, follows a fixed pattern: if an invoice above a set amount arrives, send it to the director. An AI agent combines both: it understands an instruction in plain language and carries out the steps in company systems itself.

For example, an employee writes in Microsoft Teams that they need a laptop for a new person in the sales department. A chatbot will give a link to a form. An agent will check the equipment stock, fill in the request, send it to the manager for approval and notify the IT department.

Agents differ in how independently they work. The table below shows three levels you will come across in practice. In mid-sized companies, most agents work at the first two levels, and fully independent action is reserved for low-risk matters.

An agent works well where the work is repetitive but requires reading and understanding text. The list below shows typical uses in mid-sized companies. For each one, we say what stays with a person.

The common thread is this: the agent takes over gathering data and drafting the first version of a document, and a person makes the decision. Where the decision has financial or legal consequences, or concerns assessing people, it should stay that way.

The most common mistake is starting with the tool: the company buys licences and then looks for something to use them on. It is better to reverse the order and assess several candidates against the criteria in the list below. A process that meets most of them is a good place to start.

Before you decide, estimate the gain using your own data. Suppose the IT department in a 150-person company receives 400 requests a month. If the agent resolves one in four on its own, and each took 10 minutes, that is 100 × 10 min = 1,000 minutes, or about 17 hours a month.

A result like this does not justify a large project, but it may justify a simple agent on an off-the-shelf platform. When the same calculation gives several hundred hours, it is worth considering a deeper connection to the company’s systems.

The idea of one assistant that knows everything about the company is tempting, but it rarely works. Such an assistant has too many sources, overly broad permissions and nobody responsible for the quality of its answers. After a few months, it is hard to say what it is for and whether it helps.

Several narrow agents work better: one for HR, one for IT support, one for invoices. Each has a limited set of documents, only the permissions it needs and an owner in the department. Employees do not need to know this, because they can find all of them in one place, for example in Microsoft Teams.

Once you have three or four agents, you need a simple register: who is responsible for each one, what data it can access and when it was last checked. Without it, after a year nobody knows how many agents are running in the company or what they can do.

If the company works in Microsoft 365, the shortest route is through Microsoft Copilot and Copilot Studio. Copilot Studio is Microsoft’s tool for building your own agents without programming. The agent then uses documents in SharePoint and works in Teams. Microsoft states that company data is not used to train models, and that the agent sees only what the user has permission to access.

Companies also use the business plans of ChatGPT and Claude, or their own installation of a model on company servers or in a dedicated cloud. Your own installation gives the most control over data, but it costs more and needs people to maintain it. It makes sense for data that, for various reasons, cannot be processed in a cloud service.

Make the decision separately for each process, starting with the question of what data the agent will process. For most HR, IT and sales matters, a cloud service with properly configured permissions is enough.

Employees are already using AI tools, often on personal accounts and with company data. An official agent with clear rules is safer than a ban that nobody follows. That is why it is worth writing the rules down before the first rollout.

You do not need a detailed policy straight away. To begin with, the five points in the list below are enough, written on a single page and discussed with managers.

The AI Act is the EU regulation on artificial intelligence, and it comes into force in stages. Since 2 February 2025, certain practices have been banned and companies have had to build their employees’ knowledge of AI. From 2 August 2026, an agent talking to a customer must say that it is not a human.

Most obligations concern high-risk systems, such as agents that assess job candidates. For those listed in Annex III, the obligations apply from 2 December 2027, and for AI in regulated products from 2 August 2028. An agent that answers questions about leave or suggests how to code an invoice does not fall into this group.

In Poland, the AI Act is supplemented by the Act on Artificial Intelligence Systems of 3 July 2026 (Journal of Laws (Dz.U.) 2026, item 1003). Supervision lies with KRiBSI, which has been carrying out inspections and imposing fines since 28 October 2026. Separately from the AI Act, an agent that processes personal data is subject to GDPR, so you need a data processing agreement with the provider and a legal basis for the processing.

Measurement starts before implementation. Write down how many cases the department handles a month, how long each one takes and what share comes back for correction. Without this, after a few months all you will have are impressions.

After launch, measure the same values and add two new ones: the share of cases closed by the agent without human involvement, and the share of suggestions accepted without corrections. In the costs, include licences, setup, connections to systems and the time of the person who checks the agent every month.

An honest calculation looks like this: number of cases × time saved per case × hourly labour cost, minus running costs. If the result is close to zero, the agent may still make sense, for example through faster customer service. But you need to be able to name that benefit and measure it.

An agent makes no sense for a few dozen cases a month that an employee handles in a few minutes. Nor will it help in a process where every case is different and requires experience, such as negotiating unusual contracts. There, it is better to stick with an ordinary AI assistant for writing and summarising.

The mistakes in the list below come up most often. All of them stem from treating the agent as an IT project rather than a change in how the department works.