Chatbot or AI agent? Differences, costs and when to use which

Three different things under one name, “AI”

The most common mistake is to start with the most impressive option. An AI agent looks great in a demo, but if a rule can solve the problem, plain automation will be cheaper, faster and predictable.

When plain automation is enough

When the task always looks the same: form data goes to the CRM, an invoice goes to a spreadsheet, a report goes out on Monday at 8:00. I build this kind of work in n8n, Make, Zapier or my own code. There is no language model involved, so there is no risk of made-up answers either.

An AI model is worth adding only where a rule is not enough: the content of an email has to be understood, a cost category assigned or a document summarised.

When a chatbot

A chatbot makes sense when customers ask many similar questions and you answer them by hand, often after hours. A good business chatbot:

• answers only from your content (price list, FAQ, terms), not from the model’s general knowledge;
• says “I don’t know” and hands the conversation to a person when the answer is not in its sources;
• collects contact details and a description of the issue, so nobody has to ask again from scratch;
• clearly says that the conversation is with an AI (more on that obligation below).

Answering from company documents is a technique called RAG (retrieval-augmented generation): the system first finds the matching passages in your material, and the model writes the answer from them, ideally pointing to the source.

When an AI agent

An agent makes sense for multi-step tasks that today take clicking through several systems and some judgement. Example: a request for quotation arrives by email, the agent pulls out the requirements, checks availability in the stock system, prices it by your price list and prepares a draft quote. A person approves the sending.

Security: prompt injection and too many permissions

The 2025 OWASP Top 10 for LLM applications puts prompt injection first: an attacker can change how the model behaves with an instruction hidden in a message, a document or a web page the model reads. In a chatbot the effect may be harmless (an odd answer); in an agent with access to systems it can be serious. That is why I:

1. treat content from users and documents as data, not as instructions;
2. keep the agent’s tools and permissions to a minimum (OWASP LLM06: Excessive Agency);
3. check the model’s output before it reaches other systems (OWASP LLM05: Improper Output Handling);
4. set limits on requests and spending, so a single user cannot run up a bill worth months (OWASP LLM10: Unbounded Consumption);
5. log what the agent did and why.

Legal duties in 2026: the AI Act and the GDPR

From 2 August 2026, art. 50 of the EU regulation on artificial intelligence (the AI Act) applies. AI systems intended to interact directly with people must be designed so that users know they are talking to an AI, unless that is obvious to a reasonably well-informed, observant person. Breaching the transparency obligations can cost up to 15 million euros or 3% of worldwide annual turnover. The European Commission published guidelines on art. 50 in July 2026. A clear message at the start of the conversation is usually enough, e.g. “I’m the AI assistant of company X”.

If the chatbot or agent processes personal data (a name, an email address, the content of a request), the usual GDPR rules apply too: a legal basis, information for the user and a data processing agreement with the model provider. It is worth choosing a provider and a setup that do not use your data to train models.

What it costs

On top of the build comes the cost of the AI model, billed by the provider per processed token. It depends on the number of conversations and the model you choose, so for every deployment I set a spending cap and show what a month of running costs at your traffic. Worked examples on current price lists are in the guide how much AI implementation costs.

How to start without burning the budget

1. 01

   One process

   Pick the one task that eats the most time and count how many hours a month it takes.

2. 02

   The simplest solution

   Check whether automation is enough. Add an AI model only where a rule cannot cope.

3. 03

   Test on real data

   Run a few dozen real questions or documents through it and see where it goes wrong.

4. 04

   A person in the loop

   At the start, leave approval to a person. Automate further only once the results are stable.

More examples and prices are on the page AI automation for businesses. I compare tools for plain automation in the guide n8n, Make or Zapier (in Polish).

Prepared with the help of AI tools, checked and completed by the author. Numbers and regulations come with sources.