Procurement Intelligence Agent

Pricing history is the most useful thing a procurement team owns and the hardest thing for them to reach. Answering "what did we pay for this last time?" meant asking someone to run a report.

The goal was to let the buyers ask directly, in the AI assistant they already use, without giving them a new tool to learn.

An agentic connector built on MCP — the open standard that lets an AI assistant call real tools. The assistant does not guess at prices; it calls a search tool that queries a live index of purchase orders and returns actual records, which the buyer can check.

Two tools do the work: one searches purchase prices with filters for supplier, material group, currency and date, and one reports how much coverage the data actually has for a given question. Asking for a specific purchase order number switches the call to an exact lookup instead of a relevance search, so you get all of that order's lines rather than a sample.

A nightly loader keeps the index current against the company's data lake, rebuilding changed months and resuming where it left off if it runs out of time. Failures raise an alarm rather than going quiet.

Procurement data is messy in specific ways: fifteen currencies, future-dated orders, cancelled lines that are not marked as cancelled, and material codes that only make sense with a lookup table.

A system that hides that returns confident, wrong answers. The coverage tool exists precisely so a buyer can ask how much history actually backs a number before they rely on it.