Guide · AI in practice
An AI chatbot for your website: which kind, what it costs and how to handle GDPR
An off-the-shelf chatbot service typically costs from a few hundred to a few thousand kroner a month and suits standard questions about delivery, opening hours and returns. A custom chatbot typically costs DKK 40,000–150,000 to build and makes sense when it has to look things up in your own systems. Either way, quality is decided by the knowledge base you give it. Here is the choice, the price and the GDPR checklist.
11 min read · Updated 1 October 2026
It is Sunday evening. Three customers ask whether the parcel can arrive before Friday, one wants to know if an item can be returned after 30 days, and a business wants a quote for 200 units. On Monday morning customer service spends the first hour answering the same questions they answered last week. That is exactly the situation an AI chatbot is good at: many similar questions, answers that already exist somewhere, and customers who want the answer now.
What can an AI chatbot on a website do, and what should it leave alone?
Old chatbots ran on fixed decision trees: press 1 for delivery, press 2 for returns. A modern AI chatbot understands the question however it is phrased, finds the answer in your material and writes it in a natural tone. It can also ask follow-up questions and pass the conversation to a person with a summary. What it should be kept away from are decisions with financial or legal consequences, such as discounts, complaints and credit assessment. Those belong with a person.
- Answers to repeated questions about delivery, returns, opening hours and prices.
- Help finding the right product or service based on the customer’s needs.
- Order status and tracking, once the chatbot is connected to your systems.
- Lead qualification and meeting booking straight into the calendar.
- Internal questions from staff about procedures, the handbook and products.
- Handover to a person with a short summary, so the customer does not have to repeat themselves.
Off-the-shelf or custom chatbot: which should you choose?
There are many off-the-shelf chatbot services, often as part of a helpdesk or customer-service system, where you upload your FAQ and add a widget to your website. They are quick to get started with. A custom chatbot is built into your own solution and can talk to your own systems. The choice comes down to how much the chatbot needs to look up and do.
Off-the-shelf chatbot service
- Up and running in days by uploading your FAQ and pages
- Monthly subscription, often priced per conversation or resolved case
- Limited access to your own systems
- Design and data follow the provider’s frame
- Good for standard questions and lower volume
Custom chatbot
- Typically built in four to eight weeks
- One-off price plus hosting and usage with the AI provider
- Looks up orders, stock, bookings and CRM
- You decide the design, the model and where data lives
- Good at high volume and when the answer needs your data
Decision table. If two or more rows point to the right-hand column, a custom solution is usually the better choice.
| Question | Points to off-the-shelf | Points to custom |
|---|---|---|
| Where do the answers come from? | FAQ and static pages | Orders, stock, bookings, customer data |
| How many conversations a month? | Up to a few hundred | Many hundreds or several thousand |
| Does it need to take actions? | No, only answer | Yes, such as booking, opening a case or changing an order |
| How sensitive is the data? | Ordinary product questions | Personal data, health, finances |
| How important are design and tone? | A standard widget is fine | Must match your brand and product |
How does the chatbot know what to answer?
A language model knows a lot about the world and nothing about your delivery times. It is tied to your reality with a knowledge base: your FAQ, product data, terms, manuals and past good answers. When a customer asks a question, the system finds the most relevant passages in the knowledge base and hands them to the model with an instruction to answer only from them. Technically this is called retrieval-augmented generation, or RAG. In practice it means the chatbot is exactly as good as the material it is given.
Prompt
Menneske godkender før udsendelse
What goes into the knowledge base?
- The 30–50 questions customer service gets most often, with the answers you actually want to give.
- Terms of sale, returns policy, delivery times and shipping prices in the current version.
- Product data and specifications, ideally straight from your PIM or webshop so they stay current.
- Rules for when the chatbot must hand over to a person.
- A short style guide: form of address, tone and words you avoid.
- A list of topics the chatbot must never answer, such as legal assessments and individual pricing.
What does an AI chatbot for a website cost?
The price depends on the type, the number of conversations and especially how many systems the chatbot has to talk to. Off-the-shelf services are typically priced per month, per seat or per resolved conversation, and prices change often, so check current pricing with the provider and calculate on your expected volume.
Typical price levels in Denmark in 2026. The biggest cost driver for a custom chatbot is the integrations.
| Solution | Set-up | Ongoing | Suits |
|---|---|---|---|
| Off-the-shelf, basic | Your own time to set up | A few hundred to a few thousand DKK/month | FAQ questions, low volume |
| Off-the-shelf within a helpdesk | Set-up and tuning, often DKK 5,000–20,000 | Licence plus price per resolved conversation | Customer service with an existing helpdesk |
| Custom, knowledge base | DKK 40,000–80,000 | Hosting plus AI provider usage | Your design, your rules, EU hosting |
| Custom with integrations | DKK 80,000–150,000 | Hosting plus AI provider usage | Order status, booking, CRM, login |
A worked example: a webshop with 1,500 customer enquiries a month, half of them about delivery and returns. If the chatbot resolves two thirds of those without a person, that is 500 enquiries a month. At five minutes each, that is over 40 hours a month. On an off-the-shelf service priced per resolved conversation, the bill grows with success. On a custom solution, AI provider usage is typically a small item, and the cost sits in the build. At high volume the custom solution often pays for itself within the first year. Run both models on your own numbers.
AI chatbots and GDPR: what needs to be in place?
Customers type all sorts into a chat: names, addresses, order numbers and sometimes health information or CPR numbers. So the chatbot must be treated like any other system handling personal data. This list is a practical minimum. Our guide to GDPR for websites and apps goes into depth.
- Data processing agreements with both the chatbot provider and the AI provider behind it.
- A commitment that conversations are not used to train the provider’s models.
- Processing in the EU, or a valid transfer mechanism if data leaves the EU.
- A fixed retention period for conversations, such as 30 or 90 days.
- Clear text in the chat window saying it is an AI and how the customer reaches a person.
- An updated privacy policy that mentions the chat and the purpose of processing.
- Filtering or masking of CPR numbers and other sensitive data where relevant.
- Cookie consent if the widget sets non-necessary cookies. See cookie banner rules.
How do you launch a chatbot without disappointing customers?
1. Scope
Choose the topics the chatbot covers and the ones it hands over.
2. Build the knowledge base
Gather answers, terms and product data, and remove outdated versions.
3. Test with real questions
Run 100 real enquiries through it and grade every answer.
4. Launch to part of the traffic
Start on a few pages or at certain hours, and read the conversations daily.
5. Improve weekly
Fix the knowledge base based on answers that were wrong or missing.
Which chatbot mistakes do we see most often?
- The chatbot is set up with the entire website as its knowledge base, including old campaigns and outdated prices.
- There is no route to a person, or it is hidden behind five questions.
- The chatbot may answer anything, including questions about competitors, legal matters and discounts.
- Nobody reads the conversations after launch, so the same mistakes repeat for months.
- The widget slows the site down by loading heavy scripts on every page. Keep an eye on your website speed.
- The chat window cannot be used with a keyboard or screen reader, which is a problem under the Accessibility Act if you are covered.
How do you measure whether the chatbot works?
Four numbers that tell you whether the chatbot creates value. Measure them monthly.
| Metric | What it tells you | How to use it |
|---|---|---|
| Resolved without a person | Share of conversations not handed over | The main figure for time saved |
| Correct answers | Share of sampled answers that were right | Read 20 conversations a week and grade them |
| Unanswered topics | Questions the knowledge base does not cover | Your list of new answers to write |
| Customer satisfaction | Thumbs up or down after the conversation | Find the conversations to read first |
Once the chatbot answers well, the next step is often letting it take small actions, such as creating a return label or moving a booking. That takes you into AI agents, which need more rules about what they may do. If you are unsure where the chatbot fits in your overall AI plan, start with AI in your business.
We build custom chatbots into websites, webshops and customer portals, with your design, data in the EU and connections to the systems you use through an API. See more under AI solutions, or send us your 30 most frequent questions and get a fixed price within 24 hours.
Questions about AI chatbots for websites
Can an AI chatbot answer in Danish?
Do we have to tell customers they are talking to an AI?
Can the chatbot see order status and the customer’s own data?
What happens if the chatbot gives a wrong answer?
Does a chatbot need consent in the cookie banner?
Can the chatbot generate leads or book meetings?
Want us to build it for you?
You get a fixed-price proposal within 24 hours.
