Discover
We learn the business problem, the tools and data you have, the constraints and the outcome you want.
You get: a discovery summary
AI Knowledge Assistants
Turn your company knowledge into answers people can use.
We build AI knowledge assistants that search approved company information and help employees or customers find useful answers, supported by references to their sources.
Support assistant Internal knowledge search Document Q&A Private set-up
What it is
A knowledge assistant answers questions from your own documents: policies, manuals, product information, contracts or a help centre. It first finds the relevant passages, then writes an answer from them and refers to where it came from. It suits organisations where the information exists but takes too long to find, or where the same questions keep coming in.
Problems it solves
Information is scattered
Answers live in folders, wikis, mailboxes and PDFs, and finding the right version takes time.
The same questions keep coming
Customers and colleagues ask what is already written down somewhere.
New people depend on experienced ones
Knowledge sits with a few people who are asked again and again.
Search does not understand the question
Keyword search finds documents, not the answer inside them.
What's included
AI customer support assistant
Customer answers from your approved support knowledge, with a hand-off to a person.
Internal knowledge search
One place to ask questions across the company knowledge your team may access.
Document Q&A with sources
Questions about your documents, answered with the passages that support them.
Private knowledge AI
Knowledge assistants designed around your data location, access and provider requirements.
An example, step by step
Input
A test set with known answers
Questions the service desk of a webshop really gets, each with its answer and the passage it comes from in the returns policy and product manuals, plus questions the documents do not answer, such as whether competitor prices are matched.
- Each question is put to the assistant; it retrieves passages and writes an answer that refers to them.
- Each answer is compared with the known answer and its passage: right, wrong, or answered when it should have declined.
- Wrong answers are traced to their cause, such as a missing or outdated document or a badly split passage, and fixed.
- The whole set is run again, so a fix for one question does not quietly break another.
Output
An evaluation sheet
Per question: the answer, the passage it used and the verdict, including whether it declined the questions it cannot answer. The basis for deciding whether the pilot goes live.
How it works
Choose the sources
We agree which documents and systems the assistant may use, and who may ask what.
Prepare and index
Documents are split into passages and indexed so the right passage can be found for a question.
Answer with references
The assistant writes its answer from the passages it found and refers to them.
Evaluate before launch
A set of real questions with known answers shows where the assistant is right, wrong or should say it does not know.
Delivery and handover
Sources and access written down
Which documents the assistant may use, who may ask what and where the data is stored, agreed before anything is indexed.
A pilot with an agreed test set
A limited set of documents and users, judged against questions with known answers before it is widened.
Handover for your team
How to add or remove documents, read the logs and re-run the test set when your documents change.
What we measure
Answer quality
How many test questions are answered correctly and supported by the right source.
Abstention
How often the assistant says it does not know instead of guessing.
Time to an answer
How long it takes to find an answer before and after.
Escalations
Which questions still need a person, and why.
What to keep in mind
- Retrieval reduces unsupported answers, but it does not rule them out; answers stay checkable through their references.
- An assistant is only as good as its documents: outdated or conflicting sources give outdated or conflicting answers.
- Confidential documents need agreed access rules; we do not collect them through a public form.
HOW WE WORK
From your first conversation to a working solution
Most projects start with a free 30-minute scan and a fixed-scope pilot of about four weeks.
Design
We define the scope, the approach, the timeline, who does what and how we will judge the result.
You get: a solution plan and proposal
Build
We implement what was agreed, connect the tools it needs and test it.
You get: a tested implementation
Measure & improve
We review how it performs, write down its limits and agree on maintenance or next improvements.
You get: an evaluation and handover notes
Ways to start
Free first meeting
We look at your questions and documents and see whether an assistant fits.
Pilot
A limited set of documents and users, evaluated against agreed questions, scoped and priced in a proposal.
Implementation and support
A wider roll-out, connections to your systems and maintenance, if the pilot works.
What shapes the scope and quote
- How many documents there are, in which formats and in what condition.
- How many user groups need different access.
- Where the data must be hosted and which model providers are acceptable.
- Which systems the assistant connects to, and whether we support it after launch.
Results
Results from client projects
Published only with each client's approval.
Frequently asked questions
What is a RAG assistant?
Retrieval-augmented generation: the assistant first finds the relevant passages in your documents, then writes its answer from them and refers to where it came from. This reduces unsupported answers but does not rule them out, so the assistant is evaluated before launch and can say it does not know.
What happens to our documents?
We sign a data processing agreement under the GDPR before any document is indexed. By default documents and the index are stored and processed in the EU/EEA, nothing is used to train AI models, and everything is returned or deleted at the end. Our security and data page has the details.
How do you know the answers are right?
We test the assistant on a set of real questions with known answers before launch, and keep testing as documents change. Every answer refers to its sources so people can check it.
Can we start with a pilot?
Yes. A pilot on a limited set of documents and users is the best way to see whether an assistant answers your questions well enough.
What does the assistant do when the answer is not in the documents?
It says it cannot find the answer in the sources and, where you want it, hands the question to a person. Questions without an answer are part of the test set before launch, so we check that it declines instead of guessing.