Preparing business data for an AI agent is not glamorous work. It is also where many projects quietly succeed or fail. A capable model cannot compensate for documents that contradict each other, permissions nobody understands, or a process that exists only in one person’s head.

Start with an inventory, not an upload button

List the sources the agent may use: policies, product notes, help articles, tickets, spreadsheets, contracts, and internal procedures. For each source, record an owner, last review date, audience, sensitivity, and what question it is meant to answer.

This quickly exposes a common problem: a folder called “final” beside another folder called “final-new.” Do not make the agent guess which one is authoritative. Archive duplicates, label drafts, and decide what should happen when two sources disagree.

Give information a useful shape

Break long documents into sections that make sense on their own. Keep headings, definitions, tables, version dates, and links. Put the answer near the question it addresses. If a rule has exceptions, say so next to the rule rather than hiding the exception in an unrelated appendix.

Structured records need the same care. Use stable field names, consistent dates, and explicit empty values. A customer status of “unknown” is different from a blank cell, and an effective date is different from an upload date.

Separate access from relevance

A document can be relevant and still be forbidden to a particular user. Retrieval should respect the same roles and tenants as the source system. Test with two users who should see different answers. If the results are identical, the permission boundary deserves attention before launch.

Test with real questions

Collect the questions people actually ask, including shorthand, typos, follow-up questions, and questions with missing context. For each one, write the expected source and the acceptable answer. Include “I cannot find that” as a valid result when the source is absent.

Refresh rules matter too. Decide who reviews changed documents, how quickly the agent should reflect an update, and what happens during the gap. Freshness is a product decision, not just an indexing setting.

A sensible first collection

Start with a small set of owned, current, high-value sources. In ChatNexus, the getting-started flow covers uploading data sources and then configuring the behaviour around them. Treat the first upload as a curated knowledge base, not a dump of every file the business has ever created.

Read the ChatNexus documentation before connecting sources, then keep a simple source register outside the agent so someone can maintain it.