Use a knowledge assistant for employee onboarding

Suitable questions. Suitable questions have an existing, approved answer. Good starting questions include where to request access, which form belongs to a travel approval, what an internal system is called and who supports a particular product group. The assistant can show the source and explain the next step.

Decisions to refer. Decisions that require missing personal context are less suitable. Examples include individual employment commitments, performance assessments and exceptions to an approval requirement. The assistant should explain the general process and refer the decision to the appropriate person.

Information versus action. Separate information from action. Explaining the access process does not create a user account. Summarizing a safety guide does not certify completion of training. The interface and the answer should preserve this distinction so that neither an employee nor a manager mistakes conversation for completion.

Begin with a small collection that has clear ownership. Every source needs an accountable business owner, a defined scope and an identifiable version. A practical starting set includes the team handbook, common procedures, a responsibility directory, a glossary and the policies actually required for the role.

Each source in an onboarding collection should pass a check before approval:

Avoid loading every HR document as a precaution. A small approved collection makes missing or incorrect answers easier for the team and new employees to identify. The knowledge-base guide explains the underlying content-maintenance work.

New employees often gain access gradually. An assistant must not bypass that process by collecting documents through a broadly privileged service account and exposing them to everyone. Answers need to respect the caller's reading rights, including subsequent changes.

Microsoft describes Microsoft 365 Copilot as respecting existing user permissions. That also means unnecessarily broad source permissions remain an organizational issue to correct before rollout. Do not assume every custom connector provides the same behavior. Microsoft: data, privacy and security.

Test meaningful roles: a new employee without project access, the same employee after approval, an external contractor and someone whose access has been withdrawn. Restricted information must not leak through filenames, suggested questions or shared conversation histories either.

A fictional new service employee asks how to handle a warranty request for Product A. The assistant retrieves the current procedure, identifies the required information and links to the responsible specialist group. It also explains that a goodwill exception needs approval from an authorized person.

On day three, the employee asks about a specific customer contract they cannot yet access. The correct response is a neutral explanation of the access limitation and the approved request route. The assistant should neither summarize contract details nor reconstruct them from earlier messages.

On day five, someone discovers that the product instructions are obsolete. The issue goes to the document owner. Changing the assistant's wording alone is insufficient: the source needs correction and the searchable collection needs updating. This is a proposed scenario, not a reported customer rollout.

New starters may not know internal terms or abbreviations. Retrieval should connect ordinary language with company vocabulary. A glossary could map “record my working hours” to the internal name of the time-recording system.

If a missing location or role changes which rule applies, ask a targeted follow-up question. There is no need to turn every simple request into a questionnaire. Clarification is useful when the answer changes the applicable source or next action.

For multilingual teams, a fluent English explanation must not change a policy's scope. A translated German policy still applies to the German organization it describes. Link to the authoritative original and label translations when they are not themselves approved versions.

Define a handoff for unresolved questions. A useful escalation includes the request, the approved sources already checked and the information still missing. It should not forward an entire personal conversation to a large distribution list.

A buddy can collect recurring comprehension problems while subject owners correct the content. Technical operations and business ownership are different responsibilities. When the owner changes, transfer the source responsibility as well. Atlassian, for example, documents explicit ownership transfer for Confluence content. Atlassian: content ownership.

This feedback loop can also reveal problems beyond the assistant. Several people asking the same question may indicate an unclear process rather than weak search. Fixing the underlying instruction benefits employees who never open the chat.

Do not measure only the number of chat messages. A high number can also mean that answers are confusing. Instead, check whether new employees complete a defined task using the correct source and less additional searching.

A proposed evaluation compares the existing document-search process with the assistant on comparable tasks. An onboarding evaluation should record these measures for each task:

• time
• incorrect actions
• follow-up questions
• the burden on supporting colleagues

Account for learning effects: someone who already solved a task has an advantage on the second attempt.

Include questions where the assistant should deliberately avoid making a decision. A correct referral to HR or a specialist is a successful outcome in those cases. Decide in advance which failures prevent wider use, and rerun the cases after source or model changes.

Do not use the evaluation as a hidden performance assessment of new employees. Its immediate purpose is to discover whether the information and workflow help them. Separate system-quality observations from individual employment decisions and agree who can access any recorded conversations.

A practical starting assignment for an onboarding assistant follows these steps:

1. Choose one role.
2. Name the content owner.
3. Collect common questions from recent onboarding rounds.
4. Clean up the few documents needed to answer them.
5. Then test answers, permissions and handoffs with both new and experienced employees.

The assistant should make its supporting source visible and provide a clear way to report wrong information. That creates another route into team knowledge. Whether it shortens onboarding or reduces repeated questions requires measurement in your organization; a chat interface alone does not establish either outcome.

These modernagents guides cover related decisions about AI knowledge assistants:

• AI knowledge assistants: sources, access and answers
• Build an AI knowledge base with owners and current sources
• Preserve SharePoint permissions in RAG answers
• AI literacy under Art. 4 AI Act: what SMEs must do after the omnibus

For a workflow review, start with the task, systems involved and typical exceptions. Discuss it with modernagents.