An AI assistant that answers agents from the brokerage’s own knowledge

Context

The client is a residential real estate brokerage on the US coast with a large agent base. Its policies, forms, marketing material and local know-how lived in an intranet built on WordPress, and in documents attached to it, accumulated over years.

The problem

Agents asked the same questions over and over: which form for which situation, what the policy was on a particular case, where a template lived. The answers existed but were hard to find, so people asked colleagues or the admin team instead, and the knowledge base went further out of date because nobody trusted it.

Our approach

• Retrieval first. We indexed the intranet and its documents in Elasticsearch, with chunking and metadata (document type, date, audience) so that retrieval could prefer current, relevant material over old copies.
• AWS Bedrock for generation. The brokerage’s infrastructure is on AWS, so Bedrock kept the model inside the account boundary with no new vendor to onboard. Answers are generated only from retrieved passages and cite their sources.
• Web as a fallback, clearly labelled. For questions the company material cannot answer, the assistant can search the web, and tells the agent that the answer came from outside the company.
• Admin control. The admin team decides what the assistant knows: content is synced from WordPress on a schedule, and they can exclude or prioritise material without engineering help.
• Lives where agents are. The assistant is embedded in the WordPress intranet rather than a separate app, so there was nothing new to adopt.

What we built

• A retrieval-augmented assistant on AWS Bedrock, with Elasticsearch as the search and ranking layer.
• Content synchronisation from WordPress and attached documents, with admin controls over scope.
• Cited answers with a labelled web fallback.
• Usage reporting so the admin team can see which questions are asked most and fill gaps in the knowledge base.

The result

Agents get answers from company material in seconds instead of asking around, and the admin team controls what the assistant knows. The engagement runs as a dedicated resource over a year, covering build, rollout and ongoing tuning.