RAG: make the AI read your price list before it answers

A customer asks on WhatsApp what discount applies at 30 days. The assistant answers with confidence: 15%. It sounds reasonable. It does not match your price list, and the salesperson who has to honor it finds out on Friday.

The model does not remember your company. RAG — retrieval-augmented generation — adds a step before writing: search your price list, your credit policy, the contract, and the account statement. Only then does it draft, and it cites where the answer came from. That is how we build AI assistants for selling and collecting, wired through systems integration into what you already run.

What changes when there is a source

• The answer carries a version and a clause: “price list v.6, tier-A customers, 8% from $4,000, net 15.”
• With no evidence, the assistant does not close blind. It escalates to a person, with context.
• There is a log. You can check which document it read and what it said.

RAG does not fix bad documents. If three price lists are in force, the assistant will cite one of the three. Decide which one rules first.

An assistant with no source has opinions. One with a source quotes.

If sales and collections answer the same question all day and the answer lives in a PDF, tell us where that PDF is. That is the first flow.