MIRKO

Procurement Copilot — LLM Agents for Supplier Documents and RFQs

An agentic procurement assistant for a European industrial group: it reads supplier offers, compares RFQ responses against specifications, drafts clarifications and routes approvals through SAP, cutting quote evaluation from days to hours.

Industry
Manufacturing
Country
Germany
Team Size
6
Services
LLM & AI Orchestration, AI Integration, Custom Software Development
Technologies
Claude, LangGraph, Python, SAP S/4HANA, PostgreSQL, pgvector, Azure
Procurement Copilot — LLM Agents for Supplier Documents and RFQs

Overview

About The Project

A manufacturing group with twelve plants processes thousands of supplier offers a year. Buyers spent most of their time extracting prices, lead times and compliance details from PDFs and comparing them by hand in spreadsheets.

We built a procurement copilot: a set of LLM agents that ingest offers, normalise them against the RFQ specification, flag deviations and prepare a comparison a buyer can approve, edit or reject inside the existing SAP workflow.

The Challenge

Thousands Of Offers,No Structured Data

Supplier documents arrived in every format imaginable: scanned PDFs, Excel sheets, email bodies and photos of price lists. Specifications lived in SAP, but matching an offer line to a specification line required engineering knowledge.

Any automation had to be auditable. Procurement decisions are reviewed by finance and compliance, so every extracted number needed a source, and every recommendation needed a human sign-off.

Thousands Of Offers, No Structured Data

Our Solution

Manager-Worker AgentsWith Approval Gates

A manager agent breaks each RFQ into tasks and delegates them to worker agents: document extraction with citations, specification matching through retrieval over SAP material data, deviation analysis and clarification drafting. The ledger-based control loop we describe on our YouTube channel keeps the manager from losing track of long RFQs.

Nothing leaves the system without a buyer's approval. The copilot proposes, the buyer decides, and every decision with its evidence is written back to SAP. An evaluation suite of 400 historical RFQs runs on every release to keep extraction accuracy above the agreed threshold.

Manager-Worker Agents With Approval Gates

Key Features

What WeDelivered

The Impact

Results ThatMatter

Quote evaluation became a review task instead of a data-entry task.

-78%

Time per RFQ evaluation

97.4%

Extraction accuracy

100%

Decisions with evidence trail

11 weeks

Pilot to production