Types of AI agents: what kinds there are and which one fits the process in your company

An AI chat, such as ChatGPT or Microsoft Copilot, answers questions and writes texts. An AI agent goes a step further: it uses the company’s systems itself and carries out tasks. It checks an order’s status in the ERP, opens a ticket in the service system or enters data from a document into a spreadsheet.

An agent works according to instructions and has access to selected sources of knowledge and to specific actions in systems. It is the list of these actions and their limits that determines whether an agent is a simple helper or carries out part of a process on its own. You will find a more detailed comparison of agents and chatbots in a separate article.

Vendors describe agents in many ways, often in jargon. It is simpler to ask three questions about each agent. The answers will show you what type you are dealing with and what risk comes with it.

The table below lists the types of agents that most often make sense in a company with anywhere from a few dozen to several hundred employees. The classification is practical, not scientific – one agent may combine features of two types. The “Implementation difficulty” column is our assessment of the implementation effort, not market data.

A knowledge assistant answers employees’ questions based on procedures, instructions, price lists and contracts. It works best where people often look for the same information: in customer service, in HR, in production. Instead of searching through folders, an employee asks in Teams and gets an answer with a link to the document.

There is one condition, but it is a strict one: the documents must be up to date and in one place. If there are three versions of the same procedure in SharePoint, the assistant may answer based on the oldest. That is why the project often starts with tidying up documents, not with configuring the agent.

In Microsoft 365, such an agent can be built in Copilot Studio. If it is used by employees with a Microsoft 365 Copilot licence, their conversations with an internal agent do not consume additional credits. Our case study from a factory shows what a knowledge assistant for production engineers looks like.

This type of agent talks to a customer or employee and does not stop at an answer. Take a hypothetical parts distributor whose customers phone and email to ask where their order is. The agent checks the status in the ERP, gives the tracking number and, when goods are out of stock, opens a ticket for a salesperson.

It is worth setting a clear limit on its independence. The agent can check statuses and open tickets on its own. A change of delivery date, a discount or accepting a complaint should only be prepared by the agent and approved by a person.

This type of agent is subject to a transparency obligation. From 2 August 2026, customers must know they are talking to AI, not to an employee. On top of that, there should be an easy way to reach a person when a case goes beyond the script.

Many agents never talk to anyone. They are triggered by an event: a new email in the orders mailbox, a new file in a folder, a new entry in a system. The agent reads the document, extracts the data it needs, checks it against the systems and passes it on, or asks a person for a decision.

In the Copilot Studio documentation, Microsoft gives the example of an order-handling agent. For each new order the agent carries out four actions: it checks availability, checks the shipping date, approves the order and emails the customer. That is 4 × 5 = 20 credits per order.

Let us apply this to a hypothetical wholesaler that receives 40 orders a day, or about 840 a month with 21 working days. That gives 840 × 20 = 16,800 credits, which fits within one pack of 25,000 credits for USD 200 a month. Actual usage depends on how the agent is built, so work it out in Microsoft’s calculator before you decide.

A note on purchase invoices from Polish companies. Since 1 February 2026 every company in Poland has received them through KSeF, Poland’s National e-Invoicing System, and since 1 April 2026 most companies have also issued them there. So they arrive as structured data, and the agent does not need to read them from PDFs. A document agent is more useful for orders, specifications, contracts and foreign invoices. Our case study shows how such a system works at a plant importer.

An analytics agent answers questions such as “how do sales in the Silesia region compare with last year?”. It builds the data query itself, calculates the result and shows a table or chart. For the management board, this means fewer requests to the finance department for one-off reports.

In Microsoft 365 Copilot, this role is played by the built-in Analyst agent, which analyses data from files such as Excel spreadsheets. Data from an ERP or a data warehouse needs more work. Someone has to describe what the tables and columns mean; otherwise the agent will confuse revenue with cash receipts or include cancelled orders.

The most common mistake is trusting a figure without checking where it came from. A good analytics agent shows how it calculated the result, and it is worth spending the first few weeks comparing its answers with reports the company already knows.

A screen-operating agent looks at a program’s window and clicks through it like a person. It makes sense where an old system has no connection at all for other applications and someone spends several hours a day re-keying data. However, it is slower than connecting the systems directly and sensitive to every change in how the screen looks.

In Copilot Studio, each step of such an agent is an action costing 5 credits. This type of use is not covered by the Microsoft 365 Copilot licence, so it always consumes credits from the company’s pool.

A team of agents is several specialised agents that pass a task between them. For example, one receives a request for quotation, a second selects products and calculates the price, and a third checks the customer’s commercial terms. Such a solution only makes sense once a single agent can no longer cope with the number of tasks, because every additional agent makes testing and tracking down errors harder.

The type of agent is one thing; how it is built is another. You have three routes to choose from, which differ in cost, flexibility and who maintains the solution afterwards. We compare them in the table below.

The market is changing quickly, and it is worth taking this into account when choosing. For example, on 26 August 2026 OpenAI shut down the Assistants API, on which many early agents were built, and replaced it with the Responses API. Companies that had built agents on it had to rebuild them. So ask your vendor what will happen to the agent when the model provider changes its offering.

A safe rule is: the agent prepares, a person approves. Leave independent action to tasks that are easy to undo and have little impact, such as opening a ticket or sending a confirmation. Once the agent has made no mistakes in a given step for several weeks, you can consider removing mandatory approval for it.

The law also sets limits. From 2 August 2026, an agent that talks to people must say that it is AI. Under the AI Act, an agent that assesses job candidates or the creditworthiness of individual customers is a high-risk system. The full obligations for such systems apply from 2 December 2027.

Every agent with access to company systems should also run under a separate account with minimal permissions. Its actions must be recorded in a log. That way, if something goes wrong, you know what the agent did and when.

Your first agent should show results within a few weeks and teach the company how to work with this technology. So choose a process that is narrow, frequent and well documented. The steps below will help you go from an idea to choosing the type of agent.