An internal AI knowledge assistant for marketing teams

The key idea

A useful assistant should retrieve relevant material, cite it and acknowledge when the available knowledge does not answer the question.

Our method

Start with a bounded collection of approved briefs, guidelines and process documents. Define owners and update dates. Test retrieval on realistic questions, keep document permissions intact and show the source alongside the answer. Escalate ambiguous or unsupported questions.

How to assess progress

Measure source relevance, unsupported answers and stale information. A fluent response is not proof that the system found the right document or that the requester may access it.

Decision

Decide which recurring internal questions the assistant should answer and which document collection can answer them authoritatively. Begin with a narrow subject, such as campaign approval or product messaging. Assign an owner who can resolve conflicting documents rather than asking the model to choose an organisational policy.

Exercise

Collect fifteen real questions from colleagues and identify the source passage a knowledgeable person would use for each. Add five questions with no approved answer and several that refer to an outdated process. Test from two permission profiles, checking both the generated answer and whether each linked document is accessible to that requester.

Deliverable

Produce an answerability map linking question, approved source, document owner, freshness requirement and expected response when information is absent. Add a feedback route that captures the question and cited source without requiring the colleague to diagnose the technical failure. Prioritise document repairs from repeated gaps.

Pitfall

An assistant may make an outdated document easier to discover. Treat source maintenance as part of the operating workload. When two approved files disagree, expose the conflict and send it to the owner; a fluent compromise can create a policy that nobody approved.