An AI chatbot that answers like the best member of the team.
Around the clock.
Dometrics builds AI assistants for customer service that answer exclusively from approved company knowledge, know their sources and hand over to people as soon as needed. Hosted in the EU, measured by resolution rate and response time.
An AI assistant replies here. The team takes over when needed.
- Can I still return the shoes after 30 days?
- Yes. Unworn items can be returned within 60 days. The return label is available in the customer account under Orders. Returns policy, section 2
- And if they have already been worn?
- For worn items, the service team decides case by case. Should the request be passed on together with the order history? Hand over to the team →
Where customer service fails today.
Not for lack of knowledge. The knowledge exists, but it sits in people's heads, PDFs and old emails, and it is only available during office hours.
The same questions, every day
Delivery times, returns, opening hours, product details: a large share of enquiries repeats itself. Well-trained staff spend their time copying and pasting.
Answers arrive too late
A question asked in the evening gets an answer the next morning. By then, the prospect has long since bought or booked with a competitor.
Knowledge cannot be found
Manuals, data sheets and internal rules are scattered. New employees need months before they can answer questions with confidence.
Cheap chatbots do more harm than good
Rule-based bots do not understand questions, generic AI bots make up answers. In the end, the company is liable for false commitments.
From the company's own knowledge. With clear limits.
The assistant works on the principle of retrieval-augmented generation: it first searches the approved knowledge for the relevant passages and only then formulates an answer. Anything not in the knowledge base is not claimed.
- 01 · Knowledge
Consolidate sources
FAQs, manuals, product data, policies and proven answers from the inbox are reviewed and approved.
- 02 · Index
Make it searchable
The knowledge is split into sections and indexed in a vector database in the EU. Changes flow in automatically.
- 03 · Answer
Answers with evidence
The language model answers only on the basis of the passages found, in the customer's language, and cites the source where needed.
- 04 · Handover
A person takes over
For complaints, uncertainty or purchase intent, the conversation goes to the team or into the CRM with full context.
- 05 · Learning
Measure and improve
Resolution rate, escalations and open questions show where knowledge is missing. Gaps are closed, not glossed over.
Rule-based bot versus AI assistant.
Both are called chatbots. In practice, a decision tree and an assistant with retrieval-augmented generation are worlds apart.
| Criterion | Classic rule-based bot | AI assistant by Dometrics |
|---|---|---|
| Understanding | Only recognises predefined keywords | Understands questions in natural language, typos included |
| Knowledge | Every answer has to be created individually | Answers from manuals, FAQs and product data |
| Languages | A separate tree per language | Answers in the customer's language from the same knowledge |
| Wrong answers | Rare, but often no answer at all | Greatly reduced by source grounding and limits, continuously checked |
| Handover | Dead end or contact form | Handover to a person with conversation history and summary |
| Maintenance | Laborious, quickly outdated | Updating the knowledge is enough, the assistant follows |
What separates a good AI chatbot from a bad one.
An AI chatbot in customer service is good when customers experience it not as an obstacle but as the fastest route to the right answer. That requires less a particularly large language model than a clean architecture: approved knowledge, clear limits and a handover to people that works.
What an AI chatbot can deliver
The best-known figure comes from Klarna. In its first month, the payment provider’s AI assistant handled two thirds of all service chats, and according to the company the average resolution time fell from eleven to under two minutes [1]. Figures like these arise with very high enquiry volumes and well-documented processes. For a mid-sized company, the realistic question is a different one: what share of enquiries is recurring, well documented and answerable without discretionary judgement? That share is exactly the scope of the assistant.
Good to know: Around 30 percent of Austrian companies with at least ten employees used artificial intelligence in 2025 [4]. According to Statistik Austria, more than three quarters of companies without AI have not yet even assessed its use. The head start is still there to be gained.
Why many chatbots disappoint
Rule-based bots fail on language: they recognise keywords, not concerns. Generic AI bots fail in the opposite way: they write fluently even when they do not know the answer. A case from Canada shows how expensive that can become. Air Canada’s chatbot promised a customer a refund that did not exist under the fare conditions. The tribunal ordered the airline to pay and rejected the argument that the chatbot was a separate entity [2].
Retrieval-augmented generation in one paragraph
Retrieval-augmented generation, or RAG, separates knowledge from language. The company’s knowledge sits in a searchable database. For every question, the system first retrieves the relevant sections and gives the language model only these as its basis. The model formulates the answer but does not add invented facts. If the answer is not in the knowledge base, the assistant says so and hands over to a person. If a rule changes, updating the document is enough.
The legal framework
- AI Act, Art. 50: since 2 August 2026, users must be able to recognise that they are talking to an AI [3]. A visible notice in the chat window meets this requirement.
- GDPR: processing in the EU, a data processing agreement with the model provider, no use of conversations for third-party training, short retention periods.
- Security: protection against prompt injection, meaning attempts to push the assistant beyond its limits through manipulated input [5].
- Liability: the assistant does not make commitments on prices, deadlines or goodwill gestures itself. It refers to the binding source or hands over.
Calculate instead of believe
Whether an assistant pays off comes down to a simple calculation. An example: 1,500 enquiries per month, of which 40 percent can be resolved fully automatically, each costing the team six minutes on average. That frees up 60 working hours per month for the enquiries where advice makes the difference. Added to this are enquiries outside office hours that would otherwise go unanswered. After go-live, success is measured by the actual resolution rate, not the number of conversations.
Key point: An AI chatbot does not replace a service team. It takes over the recurring work so that the team has time for the exceptional. The handover to people is not an emergency exit but part of the design.
For hotels and tourism businesses, the page AI assistant for hotels describes the specifics. How AI tools can be evaluated before they are introduced is covered in the guide Evaluating AI tools.
What the assistant comes with.
- Knowledge inventory and approval process for all sources
- RAG architecture with a vector database in the EU
- Integration on the website, in the shop or in the customer portal
- Handover to the team, ticketing system or CRM with conversation context
- AI interaction notice under Art. 50 of the AI Act
- Protection against prompt injection and off-scope topics
- Dashboard with resolution rate, escalations and open questions
- Ongoing operation: monitoring, model updates, knowledge maintenance
How the project runs
Analysis
Evaluation of the most frequent enquiries and the available knowledge sources. Result: scope, limits and target metric.
Prototype
An assistant for a clearly defined topic area, tested with real questions from the inbox.
Go-live
Integration on the website, handover to the team, training staff to handle escalations.
Operation
Monthly evaluation, knowledge maintenance and expansion to further topics, measured by the resolution rate.
Frequently asked questions
How can an AI chatbot be prevented from giving wrong answers?
Through three layers. The assistant answers only on the basis of passages it has found in the approved knowledge. It has clear instructions not to answer but to hand over when there is no basis. And every week, real conversations are spot-checked. Errors cannot be ruled out entirely with language models, but they can be reduced to a controllable level.
Does an AI chatbot have to be labelled as AI?
Yes. Under Art. 50 (1) of the EU AI Act, people must be informed that they are interacting with an AI system unless this is obvious. This obligation has applied since 2 August 2026. Dometrics builds the notice into every assistant by default.
Can an AI chatbot be GDPR-compliant?
Yes, with the right architecture: processing by providers with data centres in the EU, data processing agreements, no use of conversation data to train third-party models, short retention periods and transparent information in the privacy policy. The assistant only asks for personal data when it is needed for the handover.
Is a company liable for what its chatbot says?
That should be assumed. In 2024, Air Canada had to pay a customer a refund that the airline's chatbot had wrongly promised. The Canadian tribunal rejected the argument that the chatbot was responsible for its own statements. That is why clear limits and a handover to people belong in every serious setup.
Which language models does Dometrics use?
That depends on the requirements for quality, cost and data location. Options include commercial models via EU endpoints as well as open-weight models on dedicated infrastructure. The architecture is built so that the model can be switched later without rebuilding the knowledge base.
How long does implementation take?
A first assistant for a defined topic area is usually live after four to eight weeks. Most of the effort lies not in the technology but in preparing and approving the knowledge.
What does an AI chatbot cost to run?
Ongoing costs arise for model usage, hosting and maintenance. Model costs depend on conversation volume and, for small and medium-sized businesses, are usually well below the personnel costs of an additional service employee. Before the project, Dometrics calculates the benefit using the actual enquiry volume.
- [1] Klarna: AI assistant handles two-thirds of customer service chats in its first month (February 2024)
- [2] Civil Resolution Tribunal British Columbia: Moffatt v. Air Canada, 2024 BCCRT 149
- [3] EUR-Lex: Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (AI Act)
- [4] Statistik Austria: IKT-Einsatz in Unternehmen 2025 (ICT usage in enterprises)
- [5] OWASP: Top 10 for Large Language Model Applications
