AI agent vs chatbot: how to tell what you're actually buying

What is a chatbot?

A chatbot is software that holds a conversation. Older chatbots followed fixed scripts and buttons. Modern ones use a large language model to understand free text and write natural answers, often grounded in your own documents so they can answer questions about prices, opening hours or policies.

What a chatbot does not do on its own is change anything in your systems. It answers, and if the customer needs something done, it hands over to a person or links to a form.

What makes something an AI agent?

An agent combines a language model with tools it may call. In technical terms this is tool use or function calling: the model decides that it needs, for example, to check free slots, calls the calendar tool with the right parameters, reads the result and continues. A booking agent can therefore find a time, create the appointment and send a confirmation in one conversation.

Agents can also run without a chat window, for example reading incoming emails, classifying them and drafting replies. The defining feature is not the interface but the ability to take actions through tools.

Autonomy levels and human approval

It helps to describe any system by how much it may do on its own:

• Answer only: replies from approved content, never acts. This is a chatbot.
• Draft: prepares an email, quote or record, and a person approves it before it goes out.
• Act with limits: performs low-risk actions alone (booking a slot, logging a lead) and asks for approval on anything involving money, deletion or sensitive messages.
• Act freely: performs most actions without review. For most small businesses this is not where to start.

How should an agent be tested?

Ask how the provider evaluates the system before and after launch. A sound approach includes a set of realistic test questions and tasks with expected outcomes, checks for what happens when the agent does not know the answer, and logs of every answer and action so mistakes can be traced.

Testing should continue after go-live. Prices change, documents go out of date and new questions appear, so someone needs to review a sample of conversations and update the sources the agent relies on.

What drives the cost?

Price differences between offers usually come from four places:

• Model usage: language models are billed per token, so long documents, long conversations and high volume cost more. Choosing a smaller model for simple steps keeps this down.
• Integrations: every system the agent connects to (CRM, calendar, ERP, WhatsApp) needs building, permissions and error handling.
• Safeguards: approval steps, logging and fallbacks to a human take work, but they are what make an agent safe to run.
• Maintenance: content updates, monitoring and model changes. Ask whether this is included or billed monthly.

Red flags in vendor pitches

Be cautious if a provider:

• Calls a scripted button bot an AI agent.
• Cannot say which model is used or where your data is processed.
• Promises the agent will never make mistakes, instead of explaining how mistakes are caught.
• Offers no logs, no approval steps and no way to hand a conversation to a person.
• Keeps the accounts, prompts and data on their side so you cannot leave without starting over.

The EU AI Act: what applies to chatbots

Under Article 50 of the EU AI Act, people must be told when they are interacting with an AI system, unless this is obvious from the context. As of November 2026 this transparency obligation has applied since 2 August 2026 and was not postponed by the 2026 Digital Omnibus changes. Rules on marking AI-generated content follow their own timeline, so check the current text for details that affect you.

In practice this means your chatbot or agent should introduce itself as an AI assistant. GDPR also applies whenever personal data is processed.

Questions to ask a provider, and how Coban Labs works

Before signing, ask: which actions will it take alone, and which need approval? What happens when it does not know? Where is our data processed? Who owns the accounts and prompts? How will we see what it did?

Coban Labs, based in Brussels, builds both chatbots and agents. Mehmet Emin Çoban sets them up in the client's own accounts, with limited tool permissions, approval steps for important actions and logs of every answer. The same approach runs through Digiens ERP, which his R&D team built and which 40+ employees use every day. A first automation starts from €1,500; larger agent projects get a fixed quote after a 15-minute call.