Does your Polish contact center need a custom model?

When a question is enough

An unclear statement is easy to take as proof that the model knows Polish poorly. Yet “I don't want this package” (“nie chcę tego pakietu”) leaves two possibilities for a human too. More training examples will not establish which one this particular customer meant. Only their answer adds the information.

So when you buy a solution, look at the whole short dialogue. Can the system leave the purpose of the call open, help ask a clear question and take the answer into account? Marking a statement as unrecognised does not yet provide that kind of handling. How the conversation continues is part of the app's design.

When the customer has already said they want to keep the internet and change the package, asking again about cancelling only makes the call longer. Good intent recognition should cut such needless repetition. When you assess a tool, look both at situations that need a question and at those where the customer gave enough information.

What the model really did not understand

A different problem appears when the model regularly reads clear requests to change a package as ending the service. Customers may speak informally, break off sentences or come back to something they said earlier. If an agent understands their purpose from the same words and context, there is a concrete mistake to fix.

With voice calls, first compare the transcript with the recording. Transcription turns speech into text, which the model then analyses. Dropping “nie” (“not”) from “Nie chcę rezygnować z internetu” (“I don't want to give up the internet”) reverses its meaning. Adapting intent recognition on such a transcript does not fix an error that arose while the speech was being read.

Context works the same way: if the app passes the model only the last “yes”, without the question the customer answered, you first have to fix how the conversation is passed on. Only with a correct transcript and enough context can you sensibly judge whether the model recognises the purpose of the contact.

Does a Polish contact centre need its own intent model?

Start with your current tool or a ready multilingual model. Microsoft lists Polish among the languages supported by Conversational Language Understanding, its service for recognising intents in statements. That gives grounds to consider such a tool, and its usefulness has to be tested on the company's own conversations. Source: Microsoft CLU language support documentation.

The trial should also show whether the company clearly separates a package change from ending the service. If agents assign the same kind of request to both categories, agree on their meaning first. With a small number of clearly defined cases, a simpler menu or an improved existing dialogue may be enough.

Adapting a model means training it on examples of statements with the correct purpose of contact. It is meant to help the model recognise the wording that the company's customers use. Such work makes sense when a ready model still confuses specific, understandable requests after the categories have been explained and the right context provided. Conversations in Polish do not, by themselves, settle the need for adaptation.

How to compare results on separate conversations

Compare the errors on package change requests and on cancellations separately. A good result on frequent questions about the offer can hide mistakes on ending the service. Microsoft describes evaluating each intent separately and a matrix showing which categories the model confuses.

The comparison of the ready and the adapted model should cover separate test conversations not used for training. In our dialogue, what matters is whether the system recognises a package change after the customer's answer and does not go back to asking about cancellation. Agents can also judge how many wrong suggestions they have to correct and whether the clarifying questions help move the call forward.

Assess phone and chat separately if both channels will use the solution. The phone adds a transcription step, and a chat result does not test it. When the offer or the meaning of a category changes, look at recognition again. A separate topic, assigning finished tickets to teams, is covered in the article on fine-tuning for service ticket classification.

A conversation with Syntalith about a specific mistake

Syntalith builds apps and adapts AI models. In a contact centre we suggest starting with one distinction your current tool often gets wrong. We check whether the fix is needed in reading the call, in access to earlier statements or in the recognition itself. A comparison with a ready model lets you decide whether a trial of adaptation is justified.

The chosen way of recognising intents can be connected to the agent's tool or to a dialogue that collects the purpose of contact. For the first call, an example of a confused request and a description of what an employee has to correct afterwards is enough. If call material turns out to be needed, we agree how it may be used. Information about the offer is on Syntalith's pricing page.

What an intent model for a contact centre costs and what it saves

Where to start

In the free process review we choose one process and 20 to 100 cases the team has already handled. We compare the simpler change with the adapted model on cases the model has not seen, and within 2 working days you get a written summary with a price. We adapt models as part of private LLMs, and build the app as an AI app.