AI knowledge assistants for internal company information

Help employees find the policy, handbook or process they need. AURUM designs internal AI knowledge assistants that retrieve from approved sources, provide references where supported and give users a way to handle missing or uncertain answers.

Microsoft Solutions Partner for Modern Work · Based in Nuremberg since 2014 · About AURUM

What your knowledge assistant should prove

  • People can verify answers against approved sources
  • Restricted information stays restricted
  • The assistant knows when to send someone to a person or resource

Build an internal AI assistant people can verify

We start with a clearly defined audience and topic, such as onboarding, policy or helpdesk information. Then we test the assistant on real questions.

Choose the sources

Inventory SharePoint sites, handbooks and other approved repositories. Identify which documents are current and who owns them.

Prepare content for retrieval

Remove or label outdated copies, organise the source material and make document titles and update dates clear.

Design access and permissions

Check how the chosen platform applies source permissions. Test the effective access with users who have different rights.

Design answers and references

Show supporting sources where the platform supports it. Define what the assistant should do when a source is missing or conflicting.

Test real questions

Gather representative questions, expected sources and situations where an answer should be withheld. Measure how often people need to correct outputs.

Plan ongoing maintenance

Define who updates sources, monitors issues and reviews answer quality after changes.

Microsoft 365 Copilot, Copilot Studio or custom retrieval?

Choose based on your sources, permissions, user interface and how much control you need over retrieval and evaluation.

Microsoft 365 Copilot

Assess this first if people mainly need help finding information in Microsoft 365. Check the available licences and how existing permissions and labels affect access.

Copilot Studio

Consider a scoped assistant for particular sources and teams. Check supported source types, access and citation behaviour in the planned configuration.

Custom retrieval-augmented generation

Consider a bespoke approach when you need a particular search index, application interface or evaluation workflow. Plan source synchronisation and enforce access in the design.

Microsoft documents Copilot Studio knowledge source controls and Azure AI Search as a knowledge source. For source organisation, see our SharePoint consulting service.

Three acceptance tests for a knowledge assistant

Check answer quality, permissions and fallback before letting more employees use the assistant.

The reference supports the answer

Check that the cited material is current and actually supports the answer. Test where similar policies differ by location or team.

Information stays within its audience

Ask the same question using accounts with different access. Check the answer, source links and any associated documents.

Missing information is handled clearly

Test questions with no answer or conflicting sources. Give people a link, a contact or an escalation route to follow.

From source review to rollout

Start with a manageable collection and questions your team can evaluate. Expand once the checks pass.

Discuss your requirements
  1. Agree the first topic

    Identify the audience, questions and approved sources.

  2. Set up the assistant

    Configure retrieval, access, references and fallback behaviour.

  3. Test real questions and roles

    Review answers against the source and test restricted or missing information.

  4. Launch and maintain

    Train users, assign source owners and agree how to handle corrections and monitor quality.

AI knowledge assistant FAQ

Answers on company sources, hallucinations, access and operating costs.

What is an AI knowledge assistant?

It is a tool that helps people find answers in approved company sources, such as policies, handbooks or helpdesk articles. It should point people to sources they can check.

Can it still generate incorrect answers?

Yes. Retrieval and citations help a user verify an answer, but errors are possible. Test with known questions and design a fallback for uncertain results.

Can everyone see every document?

The answer depends on how the assistant connects to sources and applies access. We test the design with users who have different rights before rollout.

Which documents should we start with?

Choose up-to-date documents with clear owners, consistent versions and high demand. Onboarding, IT help and internal policies often offer useful starting points.

Do we need a custom solution?

It depends on the sources, permissions, interface and evaluation needs. We compare Microsoft 365 Copilot and Copilot Studio with a custom approach before deciding.

What affects the cost?

The work depends on how many sources are involved, how they are maintained, who needs access and what testing and ongoing support you need. Licence or usage charges depend on the platform.

Discuss your project in English

Bring a few questions employees ask repeatedly and the documents that should answer them. We’ll discuss the source, access and quality checks for a first pilot.

Mustafa Yaylak
Managing Director, AURUM GmbH

+49 911 25361972