Chatbot or AI agent? 5 questions to find out what your business needs

A chatbot is enough when your problem is recurring questions: it answers, pre-sorts and passes things on to a person. You only need an AI agent once the same manual steps in other systems always follow the answer, such as entering data, creating appointments or kicking off a process. The difference in one sentence: the chatbot talks, the agent gets things done.

My view up front: most small businesses should start with a good chatbot, not an agent. If you jump straight to an agent, you often end up automating a process that nobody in your business has ever properly written down. The AI then acts fast, but on fuzzy rules. This guide helps you find the right level for your business.

What is the difference between a chatbot and an AI agent?

An AI chatbot holds a conversation. It understands your customer's question, looks for the answer in your content (website, FAQ, price lists or process descriptions) and puts it into plain words. When it gets stuck or the question becomes professional, it hands over to your team. At the end of the conversation nothing has changed in your systems, except maybe an enquiry sitting in your inbox.

An AI agent goes one step further. It gets a goal, plans the steps and carries them out itself: it reads an email, checks the CRM to see whether the customer is already known, creates a record, suggests an appointment and puts it in the calendar. So the agent has not only knowledge but also tools it can use to change things in your systems.

	AI chatbot	AI agent
Core job	Answer questions, pre-sort	Carry out tasks, work through processes
Access to systems	Mostly read-only (knowledge base)	Read and write (CRM, calendar, email)
Trigger	A user asks a question	A question, an email, a new record or a schedule
When something goes wrong	A wrong or vague answer	A wrong entry, a wrong email, a wrong appointment
Need for oversight	Maintain content, spot-check conversations	Approvals, permissions, logs, ongoing monitoring
Typical starting point	One website, one knowledge area	One clearly defined process

The “when something goes wrong” row matters most. A chatbot that gets it wrong gives a bad answer. An agent that gets it wrong writes the mistake into your system or sends it to a customer. That is why agents are not simply “better chatbots” but a different category with different requirements. You can read more about both on our pages for AI chatbots and AI agents.

Chatbot or AI agent: which 5 questions decide?

Answer the following questions honestly for the process you have in mind. The more often you say “yes” to questions 2 to 5, the more likely an agent is worth it.

1. Is it about answers or about the work afterwards?

If your staff mainly answer the same questions again and again (“Are you open on Saturday?”, “Which documents do I need?”), that is a classic chatbot case. If the real time gets lost after the question, when typing things up, forwarding and entering data, you are in agent territory.

2. Do the same steps always follow in other systems?

An agent needs a pattern. If every enquiry follows the same path (enquiry comes in, customer is created in the CRM, an appointment is suggested), an agent can take over that path. If every case is different and your gut decides, it cannot.

3. Can your systems be accessed via an API?

An agent can only act where it has access. Modern calendars, CRMs and mailboxes almost always have an interface (API). Old industry software without one, on the other hand, is often a dead end. That is not an AI question but a technical one, and it should be settled before anything else.

4. Can you write down the rules?

If you can explain the process to a new apprentice on one page, including “if X, then Y”, an agent can learn it too. If you cannot write the rules down, what you need first is a process analysis, not an agent.

5. Is the volume big enough for the action to pay off?

Quite often, by the way, the answer is neither a chatbot nor an agent but a fixed workflow without any AI decision. I explain when that is enough in the guide Workflow vs. AI automation.

What should an AI chatbot for a law firm or medical practice be able to do?

Law firms and medical practices are good chatbot candidates: lots of recurring questions, a phone that never stops ringing, and a team that should really be doing professional work. At the same time, the limits here are tighter than in a trade business. A good chatbot in these sectors should be able to:

• Answer organisational questions: opening hours, directions, how a first consultation or appointment works, which documents to bring.
• Pre-sort requests: by area of law in a law firm, by type of request in a practice (appointment, prescription, test results, new patient), and pass the request on in a structured way to the right person.
• Hand over clearly: refer anything professional, meaning any question about prospects of success or symptoms, straight to a person instead of attempting an assessment.
• Identify itself as AI: since 2 August 2026, the transparency obligation under Art. 50 of the EU AI Act applies. Users must be able to tell that they are talking to an AI. Details in the guide on the chatbot labelling obligation.
• Be sparing with data: only ask for what is needed to pass the request on. Health data and case details have no place in a chat window.

What such a chatbot should not do: answer legal questions or give medical assessments. Here a person is not just better but essential. A bot that writes “you are probably entitled to …” is a liability, not a service.

There are still agent tasks in both sectors, just more behind the scenes. In our project AI assistant for law firms, the system reads in uploaded documents, extracts deadlines automatically and classifies contract clauses by risk level. That is not customer communication but support for the lawyers, who review the results. You can see which processes suit each sector on our pages for law firms and healthcare, and specifically for chatbots under AI chatbots for law firms.

What are the risks and limits of AI agents?

An agent is only as good as the guardrails you give it. The three most common problems are not science fiction but very down to earth:

• Too many permissions: an agent that is allowed to do “everything in the CRM” can also get everything wrong. It only gets access to the actions it needs for its task.
• Missing approvals: anything that goes out (emails to customers, quotes, appointment cancellations) or is hard to undo should be approved by a person, at least in the first weeks.
• No traceability: every action the agent takes belongs in a log. If something goes wrong, you need to see what it did, when and why.

Our project Agentursysteme shows a good pattern for this: the AI lead search finds matching companies by industry and location, but the user reviews the results and decides which ones go into the CRM. The AI prepares, the person approves. That is exactly the principle I recommend for most agents in a small business.

And here is the honest limit: an AI agent does not pay off if the process is rare, looks different every time, or the systems have no API. It also does not pay off if nobody in the business has time to give approvals and report mistakes during the first weeks. An agent without supervision is like a new hire nobody bothers to train.

Which tools fit which level?

For a chatbot that knows your business, a language model such as the Claude API or Mistral is enough, combined with a database such as Supabase in which your content is stored in a searchable way. Mistral is interesting if a provider based in the EU matters to you. Supabase fits because it brings database, vector search and access control in one package and can be hosted in the EU.

For agents we like to use n8n: every step shows up as a box on a canvas, approvals can be built in as a separate step and every run can be traced. For business owners that is much easier to follow than pure code. When processes get more complex, with branches and follow-up questions, LangGraph is an option because it models such multi-step processes with fixed states. For most businesses, though, that is only the second or third stage.

Which factors determine the price?

Quoting fixed prices without looking at the process would not be serious, but the factors are always the same:

• Number of connected systems: every interface (CRM, calendar, practice or law firm software) adds effort for integration and testing.
• Quality of your content and data: well-kept FAQs and clean customer data save work. Knowledge scattered across people's heads and old PDFs costs time.
• Number of special cases: a process with three variants is built quickly, one with twenty exceptions is not.
• Oversight and data protection: approval steps, logs, EU hosting and AI labelling are mandatory for agents and belong in the calculation.
• Running costs: usage fees for the language model, hosting and content maintenance continue after launch.

The biggest lever is in your hands: start with a single, clearly defined process and only expand once it runs. The guides What does an AI chatbot cost? and What does an AI agent cost? go into more depth.

My take

Start with the question of where your time actually goes: on answering, or on the work afterwards. If it is the answering, a chatbot with a clean handover to your team is the right start. If it is the work afterwards, and you can write the rules down on one page, an agent is worth it, but with approvals and a log, not flying blind. Often the best route is both, one after the other: first the chatbot that collects enquiries cleanly, then an agent that takes over the processing.

If you are unsure which level fits your business, let's look at your specific process together. Book a free AI potential check or use the contact form to arrange an initial call.

Note: The remarks on the EU AI Act and data protection are a technical assessment, not legal advice.