Get in touch

AI assistants and company knowledge bases

An assistant is only as good as what it can look up. We connect yours to the documents, systems and rules your company already has, make it show where each answer came from, and make sure it only shows people what they are allowed to see. For staff in Teams or the browser, and for customers on your website.

Discuss an assistant Read the case study
Knowledge base
  • Procurement policy.pdf Everyone
  • Price list 2026.xlsx Sales
  • Salaries 2026.xlsx HR only

Up to what amount can a department head approve a purchase?

Up to 500,000 HUF the department head approves; above that, the CFO.

Procurement policy, section 4.2

Where it runs
Microsoft Teams, a web app, your website or your own product
What it knows
Documents, shared drives, wikis and tickets, plus live data from your systems through MCP
Answers
With sources; numbers checked against the data they came from
Access
Your single sign-on (e.g. Entra ID) and per-user permissions
Voice
Optional speech input and output

Typical assistants

  • Internal knowledge assistant. Policies, procedures, contracts and project files, searchable in plain language, with links to the source.
  • Operations assistant. Answers that need live data: a tenant’s open invoices, a project’s status in Jira, stock levels in the ERP.
  • Customer-facing assistant. Product questions on your website or inside your app, grounded in your data, with hand-over to a person.

If the answers to most questions are not written down anywhere, that comes first. An assistant will not invent your processes for you.

What sits behind the chat window

Most of the work is outside the model: getting the right content in, keeping permissions intact and checking answers before they go out.

  1. 1 Sources Documents and drives indexed with metadata; live systems through tools
  2. 2 Retrieval Hybrid search, with permission filters applied before results reach the model
  3. 3 Model + tools Current Claude, GPT or Gemini models, chosen per task and data rules
  4. 4 Checks Answer validation, refusal rules, personal-data scrubbing
  5. 5 Channels Teams, web or your product, with logs and feedback for every conversation

What you get

  • Knowledge base

    Your content cleaned up, structured and kept in sync, so the assistant answers from the current version.

  • Evaluation set

    Real questions from your team with expected answers, rerun on every change to catch regressions.

  • Permissions

    Single sign-on and access rules that follow your existing groups.

  • Guardrails

    What the assistant must not do, written as rules and tested, for example no legal advice and no personal data in answers.

  • Analytics

    Which questions get asked, which ones fail, and what to add to the knowledge base next.

  • Hosting choice

    EU-hosted models, your own cloud tenant, or self-hosted on your servers.

Numbers you can trust

An assistant that does not invent figures.

For a company-information platform we built an assistant in which every number in an answer must match a value returned by a tool, and two judge models check figures and compliance before the user sees anything. The case study explains how, including the trade-offs we measured.

Questions

How is this different from ChatGPT Enterprise or Copilot?

Those are general assistants with some access to your files. We build assistants for a specific job, with your systems connected through tools, your rules enforced and an evaluation set that shows it works on your questions.

What about hallucinations?

We reduce them by design: answers grounded in retrieved sources, citations, numeric checks against the data, and a safe fallback when the assistant is not sure. Then we measure the remaining error rate on your own questions.

Can it run without sending data outside the EU?

Yes. We can use EU-hosted endpoints of Claude, GPT or Gemini models, or run the whole system on your own servers.

Can customers talk to it by voice?

Yes. We have built Hungarian speech input and output into an assistant, with numbers and tables rewritten for listening.