AI customer service chatbot: what to demand in 2026
By Jacob Slot · Published
An AI customer service chatbot built on your own knowledge can, in 2026, close the enquiries where the answer is already written down: order status, delivery times, returns, opening hours and prices. It cannot make the judgement calls that need a person, and a vendor who says otherwise is selling you something you will have to clean up after.
The difference between those two groups of enquiries is not technology. It is whether the answer sits somewhere the bot can reach, and whether anyone has decided what happens when it does not.
Danish companies ask us every month: "Can it handle Danish?" Yes. But language is the smallest of the five problems, and the other four are the ones no vendor brings up.
What an AI customer service chatbot can do in Danish in 2026
The line is sharper than the marketing suggests. A bot resolves an enquiry when the answer is written down somewhere it can access, and when no decision has to be made.
That covers more than most people expect:
- Order status and shipping, when it can look the order up in your system.
- Returns and exchanges, including whether the deadline has passed.
- Product questions that can be answered from specifications.
- Invoices, subscriptions and "when will I be charged".
- Opening hours, shipping costs and delivery countries.
It does not cover cases where several things have gone wrong at once, where the customer is angry, where an exception to the rule is needed, or where the answer costs money. That is not a technical limit, it is a choice. A bot that is allowed to give discounts is a bot that will give discounts.
The progress is no longer in the language. It is in access: a bot that can read your order system resolves enquiries, while a bot without that access can only describe. Use of AI in Danish companies with more than ten employees rose from 15% in 2023 to 28% in 2024, and 70% of users point to more efficient workflows as the positive effect, according to Statistics Denmark.
The three kinds of bots, and why the difference decides the result
Three different products are sold under the same name. They cost different amounts, work differently, and resolve very different shares of your inbox.
| Type | Draws its answers from | Can close | Typical pitfall |
|---|---|---|---|
| Rule-based flow | Buttons and fixed paths you built | Simple, predictable questions | The customer falls out of the flow and starts over |
| Bot on your own knowledge | Your own texts, policies and FAQ | Anything written down somewhere | Answers confidently from an outdated document |
| Agent with system access | Your texts plus lookups in orders, stock and CRM | Specific cases using the customer's own data | Needs permissions, logging and clear limits |
Most Danish vendors sell the middle one: a bot trained on your company's knowledge. It is a good place to start, because it can go live quickly and cannot break anything.
But it cannot answer "where is my parcel". To do that, it needs access to the system where the answer lives, and then you are in AI agents territory, not a chatbot. Ask the vendor which of the three rows the proposal falls into before you talk price.
Demand 1: it speaks Danish the way your customers do
Danish is a small language area for the large language models. That does not mean Danish does not work, but it does mean quality varies most exactly where your customers write: technical terms, inflections, abbreviations and typos.
You do not have to take anyone's word for it. Take the 20 most recent enquiries from your own inbox, paste them verbatim into the vendor's demo, and judge the answers yourself. Verbatim is the point: not neatly rewritten questions, but the customer's own, with typos, abbreviations and all.
Four things to watch for:
- Does it understand the question when it is misspelled, or written as one run-on line with no punctuation?
- Does it use your words for your things, or translate to a generic term?
- Does it reply in Danish that matches your tone, with no English phrases dropped into the middle of a sentence?
- What does it do when it does not know? Does it guess, or say so?
The fourth point is the most important, and the one demos are worst at. A bot that guesses convincingly is more expensive than no bot, because the customer acts on the answer.
Demand 2: the knowledge base is in order before the bot goes live
A bot is not a source. It is an interface to what you have already written, and it inherits every error in it.
If two documents give two different return deadlines, it picks one, with the same calm confidence whichever it is. That is the most expensive failure in the category because it is invisible: no customer calls to say the answer was wrong, they just do not order again.
So the first piece of work is not technical. It is a review of the five to ten documents the bot will answer from: terms and conditions, returns policy, shipping overview, price list, and the emails your best employee keeps writing over and over.
Demand three things from the vendor here: that you can see exactly which sources the bot has access to, that you can update them yourself without calling anyone, and that each answer shows where it came from. The last one is what makes errors findable. The method is the same one we use in all workflow automation: clean up the source before you build on top of it.
Demand 3: escalation is defined, not just mentioned
Every vendor says the bot hands over to a human. Almost none say when, and that is the whole difference.
Write four escalation rules into the agreement: when the customer asks for a person, when the bot is unsure of its own answer, when the tone is angry, and when the case involves money, a complaint or an exception.
And demand that context travels with the handover. An escalation where the customer has to start over and explain everything again is worse than no bot. Whoever takes over should get the whole conversation, the order number and whatever the bot has already looked up.
Demand 4: you know where the answers are processed
Customer service enquiries contain personal data: names, addresses, orders, sometimes health or finances. The moment a vendor processes them for you, there are requirements you must be able to document, and they are yours, not the vendor's.
Four questions you need answered in writing before the bot goes live:
- Which model is used, and where in the world are conversations processed and stored?
- Is there a data processing agreement, and which sub-processors does it list?
- How long are conversations kept, and are they used to train models?
- Does the chat say clearly that the customer is talking to an AI, and how does the customer reach a human?
The last point is both an EU requirement and good business: a bot that passes itself off as an employee gets found out, and on that day the trust costs more to rebuild than the bot cost to build. How we set this up is described on our AI chatbot page.
Demand 5: it is measured on three numbers
"Better customer experience" is not a goal, it is a feeling. Three numbers decide whether the bot works, and they must be agreed before it goes live, not afterwards.
| Number | What it is | Our recommended threshold |
|---|---|---|
| Resolution rate | Share of conversations closed without a human | At least 50% of the repetitive enquiries after two months |
| Escalation rate | Share handed over, and how fast | Under a minute to a human, context included |
| Share of correct answers | Sample of 20 conversations a week, judged by you | Over 95%, otherwise back to the knowledge base |
The third row is the one no vendor suggests on their own, and it is the most important. Resolution rate without a quality measure rewards a bot that answers everything, including what it does not know.
Set a baseline before you start: enquiries per week, split by type, and the time you spend on them. Without it, what the bot delivered becomes a matter of opinion three months later.
How to test a Danish customer service chatbot in a week
You can settle most of it before you sign, and it takes five working days.
Monday: pull the last 100 enquiries and sort them by type. Tuesday: mark which of them have a written answer, and where it lives. Wednesday: run the 20 most common ones, verbatim, through the vendor's demo. Thursday: fix the documents that contradict each other. Friday: decide based on the share that could be answered correctly, not on how impressive the demo was.
If that share falls below a third, the problem is not the bot. It is the knowledge base, and that has to be solved first, whichever vendor you choose.
The build itself is not the long part. Plan for about 3 weeks from mapping to go-live for a scoped bot, and do not launch on every channel at once. Start in the chat on one page where you can read along yourself, and expand when the three numbers hold. The timeline is laid out in our process.
What an AI customer service chatbot in Danish costs
The price comes in two parts, and a vendor who gives you only one has not given you a quote.
The first is the build: mapping, setup, integration with the systems the bot looks things up in, and testing. With us, a scoped build starts from $1,500 as a one-time payment. The second is running it: model usage, maintenance of the knowledge base, and the fixes that come when one of your systems changes. A recurring agreement starts from $750/month. We have set out the ranges for both in our breakdown of what AI automation costs.
Ask for the third number too, the one that rarely appears in a quote: how many hours your own people are expected to put in. That is typically where a cheap offer gets expensive.
And count the grant. The SMV:Digital advisory grant covers 50% of your spend on external advice, and the grant can be at most 150,000 DKK. The pools open in rounds and close when they run dry, so the date matters as much as the application. We have gathered the schemes in our guide to grants for your AI project.
Among companies with more than 250 employees, 96% use AI, against 57% of those with fewer than ten, according to Dansk Erhverv's analysis of AI use in 2025. The gap is rarely technology. It is that nobody has had time to clean up the knowledge base.
Key numbers
15 hours
freed up per week at a typical customer
90 days
to measurable ROI, written into the contract
$1,500
starting price for a scoped build, one-time
With us, measurable ROI within 90 days is a contract term, not a slide: if we cannot show it, we keep working at no extra charge. Why we dare is explained under our 90-day ROI guarantee.
FAQ about AI customer service chatbots in Danish
What does an AI customer service chatbot in Danish cost?
A scoped build starts with us from $1,500 as a one-time payment, and running it from $750/month. The price depends almost entirely on how many systems the bot has to look things up in, and how messy your knowledge base is.
How good is AI at Danish customer service in 2026?
Good enough to write and understand ordinary Danish, including typos and abbreviations. Weaker on your technical terms and on saying "I don't know".
Test it in a morning with your own 20 most recent enquiries, pasted verbatim. That is worth more than any demo the vendor has set up themselves.
How many enquiries can a chatbot take over?
It depends on what share of your enquiries has a written answer. Sort the latest 100 by type, and count the share that can be answered from a document or a lookup in a system.
That share is your realistic ceiling, not a number from a vendor's website. If it is under a third, clean up the knowledge base first.
Do I have to tell customers they are talking to an AI?
Yes. It is an EU requirement, and it is the right call regardless of the rules, because a hidden bot gets found out and costs you trust.
Say it in the first line of the chat, and keep the route to a human visible at all times. In practice it sends fewer customers elsewhere, not more.
How long does it take to get a Danish customer service chatbot live?
Plan for about 3 weeks from mapping to go-live for a scoped bot on one channel. Cleaning up the knowledge base is the part that takes time, and it sits with you.
Start with one type of enquiry. Expand when the numbers have held for two months.
The order is everything: clean up the answers, put the demands in writing, and measure the three numbers from day one. Do that, and an AI customer service chatbot in Danish is not an experiment, but the cheapest employee in the building.