Danish Language Models in Sales: How to Test Whether an AI Understands Your Conversations
Does an AI understand Danish sales talk? Test transcription and language models on your own calls: names, jargon, dialects, numbers and indirect refusals.
Table of contents
Which AI model is best for Danish?
No one can answer that for you without trying. The large language models from, among others, OpenAI, Anthropic and Google write and understand Danish well for most tasks. But the quality of spoken Danish in a busy sales conversation, with dialects, company names, technical terms and background noise, is a different and harder discipline, and it varies between models, suppliers and versions. The right answer is therefore not a ranking, but a test on your own data.
TL;DR:
- General benchmarks rarely measure what matters to you: your customers, your names and your industry.
- The most expensive errors are wrong names, numbers and agreements. Look at those, not only at an overall percentage.
- You can test a solution in a week with 20 to 30 of your own recordings. Ask every supplier, us included, to show the result on them.
Why is Danish in sales conversations hard for an AI?
Danish is a small language, and there is far less speech and text material to train on than in English. In addition, sales conversations are harder than read-aloud text:
- Telephone quality. Compressed audio, noise and people talking over each other.
- Dialects and pace. Words run together, and local expressions recur.
- Names. Place and company names, especially for smaller regional companies, may be outside what the model has seen.
- Technical terms and abbreviations. Every industry has its own.
- Mixed language. Danish sentences with English terms in between, such as "pipeline", "forecast" or product names.
- Numbers. Amounts, dates and quantities, where one wrong digit changes an agreement.
At the same time, sales is a place where an error is not merely embarrassing. A wrong name in a follow-up email, a wrong amount in a note or a misunderstood deadline can cost trust and money.
Which errors cost the most?
Not all errors are equally serious. We divide them into three.
1. Names and numbers
Place names, company names and personal names are often written wrongly in an automatic transcription, and a person who only reads the minutes will not notice. Numbers are even more critical. If a model hears "fifteen thousand" as "fifty thousand", the CRM is suddenly wrong. Always test names and numbers separately.
2. Technical terms and industry language
Technical terms in insurance, finance, leasing, construction or healthcare are rarely included in general datasets. A model may guess a word that sounds the same but means something quite different. This is especially a problem if the minutes are used as documentation.
3. Tone and indirect language
Danish business communication is often indirect. "We need to look at that internally" can be a polite refusal, and "I think we will agree" can be a positive signal. A model that translates mentally into English and reads the sentences literally may read a no as interest. This applies both to summaries that must say something about the mood, and to suggestions for next steps.
What does "benchmark" actually mean?
There are different kinds of measurement, and they should not be confused.
Speech understanding (speech to text). Here the word error rate, WER, is often used: the share of words that are replaced, omitted or added compared with a human transcript. Lower is better. It is a useful number, but it treats all words equally. A wrong "and" counts as much as a wrong amount.
Text understanding and writing. Here you measure how well a model answers questions, summarises or classifies in Danish. There are public comparisons of language models for Nordic languages, for example EuroEval (formerly ScandEval), and open Danish speech datasets, for example CoRal. They are good for a first orientation. They do not replace a test on your own conversations, because they are built on other topics, other voices and other acoustic conditions.
Task measurements. What matters to you: Is the CRM field that was filled in correct? Is the minutes correct? Is the proposed next action reasonable? This requires that you define for yourself what "correct" means.
Be therefore sceptical of figures without explanation, such as "97 percent accuracy in Danish". What was measured, on which conversations, by whom, and how was it transcribed?
How do you test a solution in a week?
It requires no data scientist. It requires a few hours and a person who knows your conversations.
Day 1: Choose the examples. Take 20 to 30 conversations or excerpts from different salespeople, customers and channels (telephone, meeting). Make sure you include the difficult ones: dialect, noise, many numbers, technical terms. Do not use only the tidy ones.
Day 1: Remember the rules. Use only recordings you are allowed to use, and do not send them to a test supplier without a data processing agreement. Alternatively, use conversations where the participants have given consent, or excerpts without personal data. See the guide to call recording in Denmark.
Day 2: Make a reference list. Have a person transcribe 10 to 15 of the excerpts word for word. This is your reference. In the reference, mark all names, numbers, dates and technical terms.
Day 3: Run the candidates. Send the same audio files through the solutions you want to compare, and save the results.
Day 4: Measure. Count three things separately: total word errors, errors on names and numbers, and errors on technical terms. Also look at the individual errors, and at how serious they would be in your work.
Day 5: Test the task. Let the solution do what it is supposed to do: write minutes, fill in CRM fields, suggest next steps. Have two people assess each result independently of one another: correct, partly correct or wrong. If you disagree, talk about it. That is where you learn the most.
Day 5: Set a limit. How many errors can you live with, and which types can you not accept? If names and numbers must be right, the solution can be designed so that a person approves them before anything is written into the CRM.
Repeat the test when the supplier changes the model, or when you change supplier. Models are updated all the time, and a test from last year says little about today.
Which questions should you ask a supplier?
- Which speech recognition and which language model are used, and which versions? If the answer is "our own", ask for documentation.
- How is Danish measured, and on what data? Ask to see the measurement on conversations similar to yours.
- Can we test on our own recordings before we commit?
- How are names, product names and technical terms handled? Can you add a glossary?
- Where is data processed, and which sub-processors are there? Who sees it, and is it used to train models?
- What happens when the model is uncertain? Is that marked, or does it guess?
- Can a person correct, and how are corrections used?
How does language quality relate to CRM and compliance?
When minutes are used to update the CRM, an error propagates. A wrong amount ends up in the forecast, a wrong name in the follow-up email. Therefore, fields with high risk (amounts, dates, names, commitments) should either be verified by a person or marked as "uncertain" when the solution is not sure.
If conversations are recorded, language quality is also a compliance matter. Minutes that must serve as documentation must be able to stand next to the recording. That is a reason to store the recording and the minutes together, so that an unclear point can be settled by listening.
Is a specialised Danish model needed?
It can make sense, but it is not always necessary. Specialised models may handle particular tasks better, especially speech recognition in Danish. At the same time, general models are improving quickly, and for many tasks, such as writing a draft or summarising, the difference is smaller than expected. The right answer is found by testing, not by choosing in advance. And you do not have to commit: if the solution is built so that the language model can be replaced, you can switch when a better one arrives.
What can Agent360 help with?
We build AI solutions in the customer's own systems. When a solution is to understand your conversations, you should ask to see the result on your own examples. That applies to us too. On the use cases page you can see examples of tasks we can help with, for example "The meeting is over. The CRM work begins", where the agent writes minutes and updates CRM fields, but asks first before anything goes out to the customer. The price depends on the task.
We cooperate with JesperAI on voice agents. What a voice agent may do and must do in your context must be assessed case by case.
Frequently asked questions about Danish AI and language models
Q: Which AI is best for Danish? It depends on the task. For writing and summarising, the large models are good. For spoken Danish on the phone, there is a bigger difference between suppliers. Test on your own data.
Q: Can an AI understand dialects? Some handle it better than others, but it is one of the things that most often goes wrong. Test with the dialects your customers actually speak.
Q: How do you measure whether an AI understands Danish? Use a reference list of human transcripts, measure word errors, errors on names and numbers, and have people assess the result of the task you want to use the AI for.
Q: How many conversations do you need to test? 20 to 30 excerpts that represent your variation is a sensible starting point. More is better, but most error patterns show up early.
Q: May we send customer conversations to a test supplier? Only if you are allowed to record and process them and there is a data processing agreement. Otherwise use consent-based examples or excerpts without personal data.
Q: How often should we test again? When the supplier changes its model, when you change supplier, and at least once a year.
Conclusion: Test on your own conversations
Language is the foundation for everything an AI is to do with your conversations. No one can say in advance how a solution will handle your customers. Choose examples, make a reference list, measure what matters, and set a limit for what a person must approve.
Book 20 minutes of free advice if you want to hear how a test on your own conversations can be arranged. See also examples of tasks we can help with.
Sources
- EuroEval, benchmark for language models in European languages
- CoRal, open Danish speech dataset (Alexandra Institute)
- The Danish Data Protection Agency (Datatilsynet), recordings and surveillance
#danish language#language models#benchmark#transcription#sales
Contents
Should your conversations write themselves up afterwards?
Minutes, follow-up and a CRM note without anyone sitting down to do it. We look at what you use today.
Book a reviewShould your conversations write themselves up afterwards?
Minutes, follow-up and a CRM note without anyone sitting down to do it. We look at what you use today.
How we built it · Sales & follow-up
Sales cockpit with AI
A call list for salespeople. The call is transcribed in Danish, AI suggests the outcome (e.g. “Call again”) and drafts the email based on what the customer said.
- The salesperson calls
- AI writes transcript and email
- The salesperson sends
Related articles

Compliance-First Sales: When GDPR Becomes a Competitive Differentiator
Companies with documented GDPR compliance win more enterprise B2B deals. Five principles, a ready-to-use compliance pack and the sales mechanics behind it.

AI Ethics in the Nordics: Transparency Is Essential for Trust
Black box AI undermines trust in Nordic B2B. See what the EU AI Act and GDPR Article 22 require, and how transparent AI helps sales teams win trust.

The Agentic Workforce: When Software Gets Agency, and What It Means for Sales
What an agentic workforce is, how AI agents differ from automation, where they help in sales, and what must be in place before you build one.