If you are considering an AI agent for a Norwegian customer base, this is usually the second question after price. It is a fair one. English is what these models were mostly trained on, and Norwegian is a small language.
The short answer: modern models write correct, natural Norwegian, and the remaining problems are not the ones people expect.
Where it is genuinely fine
Grammar, word order and vocabulary in bokmål are not the weak point any more. A customer reading a reply will not think "this was written by a machine" because of the language itself. If they think it, it will be because of the shape of the answer, not the sentences.
It also switches languages without being asked. A customer writes in English, it answers in English. Writes in Norwegian, it answers in Norwegian. For a business with any international customers, that alone removes a real cost.
Where it slips
The failures are subtler than bad grammar, and they matter more:
- Register. It tends to write more formally than Norwegian customer service usually does. Left alone it can sound like a bank letter when your customers expect a colleague.
- Translated idiom. Sentences that are technically correct but land like English wearing a Norwegian coat. A native reader notices immediately and cannot always say why.
- Your own vocabulary. Industry terms, product names, the words your customers actually use for your things. It will guess, and its guess will be the general one rather than yours.
- Compound words. Norwegian glues words together and the model occasionally splits them the English way. Small, but it is the kind of mistake that reads as careless.
All four are addressed the same way: by giving the agent real examples of how your business writes. Twenty of your own past replies do more than any amount of instruction about tone.
Bokmål, nynorsk and dialect
Bokmål is solid. Nynorsk is noticeably weaker, and if your customers write to you in nynorsk and expect nynorsk back, that needs testing specifically rather than assuming.
Dialect is a different matter. Customers write in dialect all the time, and the agent understands it better than it produces it. Reading a message in dialect and answering in clean bokmål is usually the right behaviour anyway, and it is what a person in a service role would often do too.
What happens when it gets it wrong
This is the part that deserves more attention than the grammar question, and it is the same in any language.
A language mistake is embarrassing. A factual mistake, delivered in perfect Norwegian, is expensive. The fluency is precisely what makes a wrong answer convincing.
So the protections are not linguistic. They are the same three as always: a clear list of what the agent must never answer, a person reading the conversations for the first weeks, and one click for the customer to reach a human.
How to check before you go live
A test that takes an afternoon and tells you most of what you need:
- Take thirty real customer messages, in the words your customers actually used.
- Run them through the agent and print the answers.
- Have a native speaker mark each one: correct, awkward, or wrong.
- Look only at the awkward and wrong ones. Awkward is a tone problem and is cheap to fix. Wrong is a knowledge problem and matters more.
Do this before launch, not after. Nobody regrets an afternoon spent this way.
See it answering in Norwegian first
We build a working demo on your own material before there is any commitment, so you can read what it actually says.