An internal knowledge base with private AI: policies, IT help and onboarding, in one place

The problem

In every company with a few dozen employees, the same questions go round in circles. How many days of leave carry over into next year? How do I set up the VPN on a new laptop? Who approves a software purchase? Where is the travel expense form? The answers exist, written down in policies, guides and IT support notes. People still cannot find them, so they write to HR, call IT or ask a colleague.

For a new employee it is harder still. The first weeks go on looking for things everyone else already knows, and on questions they would rather not ask twice.

General-purpose chat assistants cannot answer these questions, because they do not know your policies. If you upload the policies, internal documents leave the company. SharePoint or Drive search returns twenty documents that mention "leave" and leaves the reading to you.

What the private AI does

We install a language model on your server and connect it through RAG search to the documents the company already has, in SharePoint, Google Drive or folders on a file server. An employee asks in ordinary words. The system finds the relevant paragraphs in documents that employee can access, writes a short answer and cites the source.

A sample question and answer (a made-up policy, shown for format):

Question: Can I carry unused leave over into next year, and by when do I have to use it?

Answer: Yes. Under the internal policy, the unused part of annual leave carries over into the next year and must be used by 30 June. You do not need to request the carry-over; your manager simply includes it in the leave plan. If you could not use your leave because of sick leave, a separate procedure in Article 7 applies.

Source: HR/Working_time_policy_2025.pdf, Articles 5 and 7

Typical pilot questions:

• How do I request access to the project X folder?
• What should I do in my first week as a new team lead?
• Which printer is on the third floor, and how do I add it?
• What are the rules for working from home, and how do I report it?

When the documents hold no answer, the system says so and suggests whom to ask. That matters, because employees quickly learn whether they can trust a tool.

Data sources

The internal knowledge base is the broadest scenario, so choosing the pilot scope matters most here. We recommend starting with the one area that generates the most repeat questions:

• HR policies and guides: working time, leave, travel expenses, remote work.
• IT help: setup guides, FAQs, notes from the ticketing system.
• Onboarding: checklists, process descriptions, the org chart, who is responsible for what.
• Business processes: purchasing, approvals, travel orders, forms.

Connectors are available for SharePoint and OneDrive, Google Drive and file server folders. We support Word, PDF, PowerPoint, Excel and wiki pages, if you have a wiki.

Architecture

1. Connectors: the system reads documents from SharePoint, Drive or folders and takes over the access rights set there.
2. Indexing: documents are split into paragraphs and stored in a local search index; changes are detected automatically.
3. Permission-aware retrieval: each question only considers documents the user can open.
4. Model: an open language model (we choose between the Llama, Mistral, Qwen and Gemma families depending on language needs and hardware) writes the answer from the retrieved paragraphs only.
5. Interface: a web chat behind the company's single sign-on, optionally an OpenAI-compatible API to connect Teams or another channel.
6. Oversight: the list of questions the system could not answer is the best list of documents someone needs to write.

The company buys the hardware; we size it to the headcount and do not resell it. The options are on the private AI page.

Security and regulation

Internal documents contain employees' personal data, salaries, appraisals and business plans. In a private deployment they stay on your server and do not go to an outside model provider, and carrying the permissions over makes sure the system reveals nothing a user cannot already open today.

The General Data Protection Regulation (GDPR) applies to processing employees' personal data just as it does to your document system. What the EU AI Act means for a company that gives employees AI tools is covered in our article The EU AI Act for SMEs.

During setup we agree which sources are included, how long question logs are kept and who may review them. The question log is useful for improvements, but we do not use it to monitor individual employees, and that is written into the agreement.

Timeline and price

• Pilot (from 2,900 €, 2–3 weeks): one area, such as HR policies or IT help, a connection to one source and an evaluation on 30 to 50 questions employees actually ask.
• Production "Company knowledge" (6,000–15,000 €): several areas and sources, single sign-on, permissions carried over, monitoring and admin guidance.
• Support (290–790 € per month): new sources, updates and tracking of unanswered questions.

We start the pilot with a review of permissions, because in an internal knowledge base that is the most common source of surprises. In the first week we choose the area and the source, check who can see which folders, and collect the questions HR or IT support receive most often. In the second week we connect the system to the source and test it on those questions, including with accounts of employees who hold different permissions, to make sure nobody sees too much. In the third week a small group of employees uses it for a few days, and then you get a report with accuracy, unanswered questions and a proposal for extending it.

Prices are final; the provider is not VAT-registered. What each tier covers is on the private AI for business page. If your company also has extensive technical documentation, see private AI for manufacturing.

What we do not promise

• Better answers than the documents. If a policy is out of date or unclear, so is the answer. The system shows which questions stay unanswered; rewriting the documents is up to you.
• Perfect accuracy. The model can be wrong, which is why it always cites the source.
• Working without sound permissions. If permissions in SharePoint or on Drive are set too broadly, the system takes them over as they are. We review them together before the pilot.
• The quality of the largest cloud models on every task. For answers from internal documents open models are very good; we show the difference on harder tasks during the pilot.