When buying a chatbot, people ask about price, languages and data protection. Almost nobody asks about the language model. Yet that is what decides where a business stands in two years.
This article is for businesses buying a chatbot or already running one. If you are building one yourself and want to know how the testing is set up technically, that is in Evaluating AI chatbots with LangSmith.
Which language model is inside your chatbot?
Every chatbot that answers halfway usefully has a language model behind it. That model comes from a provider such as Mistral, OpenAI or Google, and the chatbot provider buys it there.
It rarely comes up in a sales conversation. People ask about the monthly price, the languages, the server location. All of it fair enough. Which model writes the answers, and what happens when that changes, is something hardly anyone wants to know.
That is understandable, because when buying you look at what can be compared. Even so, this question is where it is decided whether the chatbot still runs in two years, and what it costs by then.
What happens when your AI provider raises its prices?
A chatbot provider tied to a single AI provider has five things outside its control. In procurement this is called vendor lock-in, meaning being bound to one single supplier.
- Raising prices. If the AI gets more expensive to buy, the chatbot provider pays along. It can pass the increase on or absorb it. A third option only exists if it can switch. What a chatbot costs in the first place is covered in what an AI chatbot really costs.
- Discontinuing a model. Models get retired, often with short notice. Anyone with only one model wired in ends up rebuilding under time pressure.
- Changing the terms of use. What may happen to the data is set by the AI provider. Its customer gets to take note.
- Throttling performance. Response times and monthly quotas are guaranteed nowhere for the long run.
- Moving the server location. The most awkward one for data protection, because it undoes a promise you made to your own customers. Why location matters is covered in EU hosting vs. US cloud.
We do not exempt ourselves from this. We buy the language model in as well, and when it gets more expensive there, we feel it just the same. Anyone claiming to be exempt from this mechanism either has a clause they are not showing you, or a calculation that does not add up. The difference lies in whether you can move.
How does a chatbot have to be built so the model can be swapped?
The language model phrases the answer. But it only does part of the work.
Before it is even asked, this happens: the visitor's question is classified, the relevant passages in your documents are located, and the hits are ranked and assembled. Only from that does the model form a sentence. After that comes the course of the conversation, the handover to a human when things get tricky, and forwarding an enquiry by email.
That foundation is ours, and it does not depend on the model.
What a model change touches, and what it does not
The language model sits on top and can be replaced. Everything your chatbot is made of stays where it is.
So a change leaves most of it untouched. Your documents, the instructions on how the chatbot should answer, the settings on the chat window, the existing conversations and the integrations with other systems, as far as your plan includes them. Actions and CRM integrations are currently part of Managed Pro; in Self-Service they are in preparation.
A human taking over has nothing to do with the language model, and neither does the AI disclosure in the chat window. Everything that belongs to it is on the features page.
What does switching the language model cost?
The default at ServasBot is Mistral. If you explicitly want OpenAI, OpenAI gets wired in.
It is still work. Depending on the setup, the knowledge base has to be read in and rebuilt, because your documents are prepared for search and that preparation cannot always be carried over. After that the answers need checking before anything goes live.
With Managed we do that: switch over, read the documents in again, check the answers. In Self-Service the customer carries out the change themselves when it becomes necessary. Setup is available as an add-on package, in which case we take that part over. How the work divides up between the two paths otherwise is on Self-Service or Managed and in why ongoing chatbot maintenance matters.
With a chatbot hard-wired to one provider, this path does not exist at all. There the question of effort never comes up, because it already stops at whether.
How do you tell whether a new AI model really answers better?
A new model version means nothing at first. Newer is not automatically better, and certainly not better for your business.
The only way to test that is against your own questions, and those depend on your knowledge base. Whether a model correctly conveys how long a plumber's emergency call-out runs and what it costs appears in no general benchmark. That is why the question set is built per customer, from their documents, with the expected answers on record. The technical term for such a set is a golden dataset.
The procedure is then straightforward. The same question set, the same documents, the current configuration against the new one. Both run through, we evaluate them via LangSmith and put the results side by side. We switch when the answers come out as good or better. Otherwise everything stays as it is, and the new model waits for its next version.
The same applies to a version jump at the same provider. Especially then, because behaviour can change without anyone announcing it.
This customer-specific testing is part of Managed. In Self-Service it sits with the customer, because nobody else knows which answers are right for their business.
Why a European language model?
The default is Mistral, a provider from France. Your data and the chatbot itself sit on servers in Germany.
Those are two different things, and both were a decision. Hosting is about where your customer data sits and which law can reach it. The model is about who we depend on. More on the first point in EU hosting vs. US cloud and in GDPR and AI.
The fact that we wire in OpenAI on request does not contradict this. It shows the choice exists. And that closes the circle: choosing a European model is only worth something if it could be done differently. A chatbot without provider lock-in at the model layer is what makes that a choice at all. Anyone who cannot switch never made a choice. They took what was built in.
What should you ask a chatbot provider?
What to look for in a chatbot provider boils down to five questions. Put these to anyone who wants to sell you a chatbot. Us included.
- Which language model do you use, in which version, and where does it run?
- What happens to my price when your AI costs go up? Does the contract say anything about it?
- What happens if this model is discontinued? Do I have to start over?
- Do you test a new version against my own questions before switching? And do I get to see the result?
- Who owns my documents and conversations, and which data can I take with me when switching provider?
Question 5 comes out mixed for us, so here is the answer straight away. Your documents and content belong to you, that is in our terms. After you cancel, we delete everything from our servers completely. Exporting your conversations is not something you can trigger yourself in the dashboard today; that goes through a request to us. A self-service export is in preparation.
A question you ask yourself is not one you need to fear.
What you can do now
If you are comparing providers right now, take the five questions with you. The answers tell you more about the next two years than any list of features.
If you already have a chatbot and do not know what is inside it, we can look together. A short call is usually enough.
Rather try it yourself?
Set up a chatbot, upload your documents, try it free for 30 days. The language model behind it is the same one.
Start 30-day free trial