Build a trustworthy dealer AI knowledge base

A dealer AI assistant is only as useful as the information it can retrieve and the boundaries it respects. A folder full of old offers, conflicting policies, and unlabeled documents can produce confident but unreliable answers. Build the knowledge base as an operating resource with owners, versions, and clear limits.

Separate stable facts from changing facts

Store identity, department contacts, approved processes, and general service information change differently from inventory, incentives, and appointment availability. Label those categories so the assistant knows which facts require a current system check.

Do not treat a historical offer PDF as today’s pricing authority. Include effective dates, expiration dates, and the source owner where relevant. If a document is retained for history, mark it as archived rather than leaving it beside current material without distinction.

Give each source an owner

Assign responsibility for sales policies, service procedures, marketing offers, and technical integrations. The owner should know how to correct a source and how that correction reaches the assistant. A knowledge base without an update process becomes stale even if it starts accurate.

Keep the original source and context available. A short extracted sentence can lose an important qualification. Answers should be able to point a reviewer back to the material that supports them, especially when the information affects a customer decision.

Write for retrieval and understanding

Use descriptive headings, concise sections, and explicit definitions. State which department or situation a procedure applies to. Avoid a single enormous document that mixes unrelated policies with no useful structure.

Include common customer wording where it helps retrieval, but do not create dozens of duplicate entries. A clear authoritative answer with recognizable synonyms is easier to maintain than several versions that slowly contradict one another.

Protect sensitive information

Separate public information, staff procedures, and restricted records. Apply access controls based on the user and task. An assistant that can help a salesperson find a service phone number does not necessarily need access to finance applications or private employee records.

Treat retrieved documents and external text as information, not as permission to change the assistant’s rules. A pasted instruction inside a document should not authorize data sharing, messaging, or system changes. Test that boundary with controlled examples before wider use.

Teach the assistant to stop at uncertainty

Define what happens when sources disagree, an offer is expired, or the answer depends on a live record. The correct result may be a request for a human check. Do not reward the system for producing a complete answer when the evidence is incomplete.

For an illustrative warranty question, the assistant can explain where to find the relevant documentation and route the specific vehicle question appropriately. It should not infer coverage from the model alone or promise a repair outcome without verification.

Evaluate with real questions

Build a small set of representative questions from staff, stripped of private customer details. Include straightforward answers, ambiguous requests, expired information, and questions the system should decline to answer definitively. Have the responsible subject owner review the results.

After updates, rerun the questions affected by the change. Track wrong answers by their cause: missing source, stale source, retrieval failure, or unsupported inference. Fixing the knowledge base may be more useful than changing the model.

Knowledge-base checklist

• Separate stable information from live or expiring facts.
• Assign a source owner and correction process.
• Preserve context and make supporting material traceable.
• Apply access rules by role and purpose.
• Define uncertainty and escalation behavior.
• Test expired, conflicting, and out-of-scope questions.

Can we upload every dealership file?

That is rarely a good starting point. Curate the material needed for the approved use case, remove duplicates, and apply access controls. More documents do not automatically create better answers; they can create more contradictions.