AI agent vs chatbot: how they differ and which to choose for your company

Take a hypothetical wholesaler with 70 employees. A customer writes in the website chat: “Where is my order 4512, and can the delivery be moved to Friday?” Let us see what three different solutions would do.

A classic chatbot will show a menu: “Order status”, “Complaints”, “Contact”. After the customer picks an option, it will give a link to the customer portal or a phone number. It will not answer the second question, because it is not in the script.

An AI assistant with a knowledge base will understand the question and reply according to the terms: “Delivery can be rescheduled up to 24 hours before dispatch; please contact customer service”. That is better, but the customer still has to write a second time.

An AI agent will check the order in the system, reply that it is still in the warehouse and offer delivery on Friday. Once the customer confirms, it will change the date in the system and send a confirmation email. If the change is not possible, it will hand the case to an employee with a full description.

Vendors use these words interchangeably, so it is worth distinguishing three levels instead of two. They differ in where they get their answers from and whether they change anything in systems. We compare them in the table below.

The line between an assistant and an agent does not lie in the artificial intelligence itself, but in the connections to systems. The same AI can be a simple assistant or an agent, depending on what it has access to and what it is allowed to do.

An agent has three features a chatbot lacks. First, it uses tools, i.e. connections to company systems: it can read an order from the ERP, log a ticket or start an approval in Power Automate. Second, it works out by itself the order of the steps needed to resolve the matter.

Third, a good agent knows when to stop. It hands the case to a person when an amount exceeds a threshold, when the customer is unhappy, or when it is not sure of the answer. These rules are set by the company, not the model – and they largely determine the quality of the rollout.

In practice, an agent can work in Teams for employees or on the website for customers. In the Microsoft ecosystem, such agents are usually built in Copilot Studio and connected to systems through Power Automate or ready-made connectors for popular software.

In many companies, an agent would be overkill. If most questions concern opening hours, prices, complaint terms or instructions, an assistant that answers from documents is enough. It is cheaper, quicker to launch and needs no access to systems. The list below sets out typical matters it can handle without an agent.

A simple test: go through the last 100 enquiries from customers or employees. Count how many end with just an answer and how many require someone to check or change something in a system. If the vast majority are just answers, start with an assistant.

An agent makes sense when, after the conversation, someone in the company has to perform an action in a system anyway, and that action is repetitive. Typical examples are checking an order status, changing a date, requesting equipment or logging a service ticket.

The second condition is access to systems. An agent needs a way to read and write data, usually through an interface for exchanging data (API). If the system has none, the connection can be expensive or unreliable – in that case it is better to stick with an assistant that collects the details and passes the case to a person.

The third condition is clear rules. The agent must know up to what amount it can act alone, when it asks a person, and what it never does. If these rules are not written down for employees, they will not exist for the agent either.

The cost has two parts: the platform licence and the implementation work. In Copilot Studio, conversations are billed in credits. According to Microsoft, a ready-made answer costs 1 credit, an AI-generated answer 2 credits and each agent action 5 credits.

Let us make a rough estimate. A conversation with two AI answers and two agent actions comes to 2 × 2 + 2 × 5 = 14 credits. A pack of 25,000 credits for USD 200 a month is therefore enough for about 1,785 such conversations. The same pool is enough for about 6,250 conversations with an assistant that gives two AI answers and changes nothing in systems (2 × 2 = 4 credits).

An important exception: employees with a Microsoft 365 Copilot licence use internal agents without consuming credits. By contrast, conversations with customer-facing agents on the website always consume Copilot Studio credits. When usage exceeds 125% of the prepaid pool, Microsoft disables the agents, so the limit has to be monitored.

The larger part of the cost is usually the work: preparing the knowledge, connecting to systems, testing and maintenance. It grows with every system the agent has to reach. In this respect, an assistant with a knowledge base is clearly cheaper than an agent with three integrations.

A chatbot that makes a mistake gives a wrong answer. An agent that makes a mistake can change data in a system. That is why an agent needs extra safeguards that you do not have to design for a simple chatbot.

The AI Act, the EU regulation on artificial intelligence, introduces a transparency obligation. From 2 August 2026, a person talking to an AI system must be aware of it, unless this is obvious. This applies both to an assistant on the website and to an agent in Teams.

In practice, a clear notice at the start of the conversation is usually enough, for example: “You are talking to Company X’s AI assistant. You can ask to speak to an employee at any time.” Most customer service or employee agents are not high-risk systems. The exceptions include uses in recruitment and employee assessment, or in creditworthiness assessment.

Before you talk to a vendor, answer the questions below based on real enquiries from recent weeks. If the answer to the last three is “no” or “we don’t know”, start with a knowledge-base assistant and come back to the idea of an agent in a few months.