AGRES AI Sales Assistant

Why it exists

In sales, a made-up price, spec, or discount is a liability, not a typo. So this assistant can talk naturally, but every price, spec, and policy comes from stored data. If the data does not have the answer, it says it has not found verified information instead of guessing.

What it does

A conversational sales assistant for the AGRES.ID laptop catalog, built on a snapshot of its public catalog and policy pages. A shopper narrows laptops down by budget and needs, compares two or three models, and asks about warranty, shipping, or returns. The conversation is also summarised for a human salesperson, with a score for how ready the shopper is to buy and a suggested next step. Not affiliated with AGRES.ID; prices and stock are snapshot data, not real-time.

How it works

The assistant reads your budget and needs, filters the catalog with fixed rules, and ranks what is left. Prices, specs, and policies are always read from the stored catalog, never written by the AI.

1. Understand: it picks out your budget and requirements from the chat and remembers them for the rest of the conversation.
2. Recommend: it removes laptops that break your limits (budget, RAM, graphics card, weight, out of stock), ranks the rest, and shows two or three with reasons and trade-offs. Comparisons re-read each product from the database.
3. Answer policy questions: it finds the relevant section of the store's policy pages and quotes it.
4. Hand off: it scores how ready the shopper is to buy (0–10), suggests a next step, and writes a summary for a human salesperson. Nothing is sent automatically.

A key decision

The AI is optional and is never the source of facts. It can help interpret an unclear message or reword a verified answer, but a checker compares every price, spec, discount, and link in the reworded text against the database and discards the text if anything does not match. The live demo runs without the AI rewording step.

How well it works

I tested it with full conversations against the real database. My main set of 53 test conversations all passed, but I also used that set while tuning, so it says little about new questions. Two fresh sets written afterwards passed 9 of 16 and 10 of 14 on the first try. On every set it never recommended a laptop that broke the shopper's limits and never stated an unsupported fact. The weak spot is unusual phrasing: on policy questions it had not seen, the right section was among its top three results only 50–71% of the time.