AI Agents, Chatbots & Voice AI

AI assistants for sales, support, phone and documents that answer from your own data and hand over to a person when they are unsure.

Most of what we know about AI assistants comes from running our own. The chat on this site answers visitors through a workflow on our own n8n server, with a system prompt that describes our products. LynBro PBX answers phone calls on an Asterisk exchange we operate ourselves. ai.lynbro.dk runs image and text pipelines for clients behind a separate key for each client, and Danmark Data MCP lets ChatGPT and Claude look up Danish companies, addresses and vehicles.

That work has also shown us where AI does damage. A model will state something wrong with full confidence when its instructions are out of date: our own chat once told a visitor that a product had no trial, although it had one, because the prompt had not been updated together with the site. So every assistant we build gets a narrow job, a source it must answer from, a rule for when to stay silent and a log a person can read. We also write down which AI provider sees which data, so it can go straight into your GDPR records.

What we build with AI

AI sales assistant that knows when to stay quiet

A sales assistant reads incoming messages from the website, Messenger or Instagram, works out what the person wants, asks for the missing details and hands a clean lead to your CRM. It pays off when enquiries arrive in the evening or when staff paste the same answers all day. For a Danish construction company we tested the triage step on 55 real message threads: the model agreed with what staff had done in 46, would have asked a clarifying question in 3 and held back in 6, with no unsafe replies. A fixed policy, not the model, decides whether to reply, use a company template or wait for a person, and a filter stops other people's phone numbers, CPR numbers and bank details from appearing in a reply.

Support agent that learns from what it cannot answer

A support agent answers the recurring questions, such as opening hours, prices, delivery and rebooking, from a knowledge base you control, and passes everything else to a person with a short summary. What makes it better over time is the handling of unanswered questions. In our phone agent each unknown question lands in a Questions tab. A member of staff answers it once, the answer is turned into one to five short facts, and the next customer gets it. Before anything is added, the question is checked against the price list and the website: if the answer is already there, the agent is pointed to it instead of getting a second copy that can drift out of date.

AI voice receptionist

A voice agent that picks up the phone in Danish, English or Ukrainian, answers from your knowledge base, books or takes a message and transfers the caller to a person on request. It runs on our own Asterisk 22 and FreePBX 17 exchange with OpenAI Realtime, and we build custom call flows on the same stack.

Learn more →

Website chatbots that stay in step with the site

A website chatbot is the cheapest way to try AI with customers, and also the easiest to get wrong. The chat on lynbro.dk runs as a workflow on our own n8n server and answers in the visitor's language in about three and a half seconds. Running it taught us two things. The chat widget weighed 2 MB, so it now loads only after the first interaction or when the browser is idle. And facts about our products lived in four places, namely the footer, the portfolio, the chat prompt and llms.txt, so the chat fell out of date until updating all four became part of every release. If you want a ready chatbot rather than a custom build, cbot.dk is our own product for that.

Telegram and WhatsApp bots

Bots in Telegram and WhatsApp for bookings, confirmations, community moderation and staff alerts. We have built a Telegram Mini App used by around 300 members of a Ukrainian community, aiogram bots backed by Google Sheets, and WhatsApp booking confirmations on Meta's Cloud API wired through n8n.

Learn more →

Internal knowledge AI

An assistant that answers staff from your own documents, shows the source and stays silent when the answer is not there. Our cross-language search finds an article written in Ukrainian from a question asked in Danish, and our MCP server gives ChatGPT and Claude read-only access to Danish company and address data.

Learn more →

Document assistant for quotes, jobs and texts

A document assistant reads what people send you, such as job descriptions, quotes, PDFs and spreadsheets, and turns it into structured data your systems can use. For a Danish construction company we built a step that splits one job description into work packages for subcontractors. Every package carries a reference to the sentence it came from and a confidence value, and the model is not allowed to guess durations. For the same client a translation step turns app texts into Danish, English and Ukrainian with a construction glossary, then checks that the numbers are unchanged and that no new e-mail addresses, links or amounts have appeared. One text costs about 0.0003 USD. Malformed input is rejected before it reaches the model, so a test run costs nothing.

Call, meeting and e-mail summaries

These tools take the reading off your plate: a summary after each call, a morning list of what needs an answer, a draft reply that waits for approval. Our own setup is the test bench. After every call our phone agent posts a card in Telegram with who called, a summary, any request and open questions. The owner's private assistant sends a brief at 07:30, sorts the inbox by 62 tested rules and never sends or changes anything without an explicit yes; its model spending is capped at 2 USD a day. We have not built a bot that joins video meetings for a client, and if that is what you need, we will tell you so in the first conversation.

AI we run ourselves

AI chat on lynbro.dk

The chat bubble on this site. It runs on our own n8n instance with a system prompt about our products, answers in Danish, English or Ukrainian and loads lazily, so it does not slow down the first page view. A recurring audit compares the prompt with the site after it once gave an outdated answer about a trial.

LynBro PBX

Our own AI phone agent on an Asterisk 22 and FreePBX 17 exchange connected to OpenAI Realtime. It answers from a knowledge base, transfers to a person and sends a call card to Telegram. Public sign-up with Stripe billing opened on 21 October 2026.

cbot.dk

Our own chatbot product for websites, which passes leads on to a CRM. Meta Lead Ads from 16 of our own Facebook pages flow into it, and an end-to-end test lead arrived in 7 seconds. Its autopilot for social messages runs in shadow mode next to staff before it may reply on its own.

LynBro AI (ai.lynbro.dk)

Our AI pipeline service. A client's website keeps only a thin proxy; prompts, model choice and the OpenAI key stay with us, and each client key has a daily call limit and a daily cost ceiling in DKK. The public demo Farv dit hus recolours a photo of a house for about 0.61 kr. per photo.

Tools we use for AI

OpenAI models through the API: Realtime for voice, smaller models for triage, summaries and translation. n8n for chat workflows that should change without a deploy, Python and FastAPI for anything that needs tests, Meilisearch for combined keyword and vector search over a knowledge base, MySQL or PostgreSQL for logs and Telegram for alerts. We pick the cheapest model that passes the test set and log the cost of every call.

Prices for AI solutions

AI chatbot for a website
from 1,500 DKK
AI chatbot in Telegram
from 1,000 DKK
AI call processing
from 5,000 DKK
Document processing (PDF/Excel)
from 1,000 DKK
Consultation
1,000 DKK/hour
Project support
600 DKK/month

All prices ex. VAT. The final price depends on the number of channels and languages, the systems the assistant must read from or write to and how much testing the risk calls for; we quote after a free scoping talk through the contact form.

How an AI assistant goes live

1

Start from real conversations

We begin with a sample of real messages, e-mails or call notes, not a feature list. They show what customers actually ask, which answers need a person and which data the assistant must be able to read.

2

Put the rules outside the model

What the assistant may say, when it must hand over and which data it must never repeat are written as code and settings. The model handles language; the rules handle risk and stay the same when the model is replaced.

3

Test against a fixed set

We run questions that must be answered and questions that must be refused, read the results and set thresholds from the scores. The same set is run again after every change to the prompt or the model.

4

Shadow mode, then live

First the assistant drafts and a person sends. When the drafts are right, it answers on its own within its limits. We review logs, costs and unanswered questions, and update the prompt whenever the website or prices change.

Frequently Asked Questions

Will the assistant make things up?

It can, if it is allowed to answer from general knowledge. We restrict it to your sources, set a score below which it must say it does not know, and log those questions for a person. In our cross-language search test that gave the right article for 13 of 13 questions and no invented answer to any of 5 off-topic ones. That lowers the risk; it does not remove it, which is why a person reviews the logs.

Is it allowed under GDPR to send customer messages to an AI provider?

It can be, but it has to be planned. We list which provider sees which data, keep transcripts as text rather than audio recordings, leave out details the model does not need and agree retention periods with you. Our phone agent tells callers at the start that it is an AI assistant and that the conversation is written down. You remain the data controller and decide the legal basis.

What does it cost to run each month?

Model usage is paid per use, so it depends on volume and on the model. We log the cost of every call and can put a daily ceiling on it; our house-recolouring demo, for example, costs about 0.61 kr. per photo. Hosting and project support from 600 DKK a month come on top.

Can one assistant handle Danish and Ukrainian at the same time?

Yes. All our own assistants run in Danish, English and Ukrainian. The knowledge base can be written in one language and searched in another, because for each article we store sample questions in all three languages and index them together with the text.

Should we start with a chatbot or a phone agent?

Start where the repeated work is. If most questions come by phone, a voice agent saves more time. If they come by chat or social messages, a text assistant is cheaper to build and easier to test. Many businesses begin with a website chatbot and later reuse the same knowledge base on the phone.

⚡

Get an Instant Price Estimate

Use our interactive calculator to get an approximate price for your project in 2 minutes.

⚡ Calculate Price

Ready to Cross the Bridge?

Let's build your lightning bridge to digital transformation. From first consultation to live deployment — we move fast.

Get Started Today