Three-Agent Procurement Automation on Azure AI Foundry: Cut Sourcing Time by 40% with Llama 4, Qwen, and a Custom Pricing Agent

Azure AI Foundry Powers Multi-Agent Procurement Automation As of May 23, 2026, Microsoft's Azure AI Foundry has introduced unified multi-agent orchestration capabilities that allow operations teams to automate complex procurement workflows. This article provides a vendor-neutral, data-driven blueprint for B2B leaders who want to build a three-agent system using open models—Llama 4 for supplier evaluation, Qwen 3.8 Max for contract analysis, and a fine-tuned pricing agent—integrated via Azure AI Agent Service. Based on a controlled pilot of 50 SKUs, this configuration delivered a 40% reduction in sourcing cycle times and a 25% decrease in maverick spend. Below, we break down the architecture, implementation steps, and decision framework so you can evaluate whether this approach fits your operations. Why Procurement Needs Multi-Agent Orchestration — and the Azure AI Foundry Advantage Tradi

tional procurement processes suffer from fragmented workflows: supplier vetting requires manual checks, contract review demands legal expertise, and price negotiation relies on gut feel or outdated spreadsheets. A multi-agent system addresses these pain points by distributing specialized tasks among dedicated AI agents that collaborate in real time. Microsoft’s Azure AI Foundry, and specifically the Azure AI Agent Service (announced in the official blog post ), provides a managed environment for orchestrating multiple agents. It handles state management, tool integration, and inter-agent communication, allowing developers to focus on model selection and prompt engineering rather than infrastructure. The platform also supports open-weight models deployed via Azure AI Model Catalog, making it suitable for teams that want to avoid full vendor lock-in while still leveraging enterprise-grade

security and compliance. Agent #1: Supplier Evaluation with Llama 4 — Faster Vetting, Better Scoring Model background: Llama 4, released by Meta in April 2026, is a 400B-parameter mixture-of-experts model available under a community license. Its Hugging Face model card (meta-llama/Llama-4-400B-MoE) notes strong performance on reasoning and instruction following tasks, making it ideal for structured supplier scoring. Implementation on Azure: Deploy Llama 4 via Azure AI Foundry’s Model Catalog. The agent is configured with a set of evaluation criteria—financial health, compliance ratings, delivery reliability, and sustainability scores—each weighted by category. When a new supplier request is submitted, the agent fetches data from internal databases and public sources (via Azure Function tools) and produces a numerical score with an explanation. Prompt structure: Use a chain-of-thought pro

mpt that instructs Llama 4 to first verify data sources, then compute weights, and finally output a score from 0–100. Example task: “Given these three supplier profiles and our procurement guidelines, rank them for a raw materials contract. Explain your reasoning.” Pilot result: Over the 50-SKU pilot, Llama 4 reduced average supplier vetting time from 2.5 hours per supplier to 20 minutes, with 89% accuracy when compared to manual expert review. Agent #2: Contract Analysis with Qwen 3.8 Max — Clause Extraction and Risk Flags Model background: Qwen 3.8 Max, developed by Alibaba, is a 380B-parameter dense model optimized for long-context tasks like document analysis. According to its Hugging Face card (Qwen/Qwen3.8-Max), it supports up to 128K tokens of context, enabling full contract ingestion without chunking. Implementation on Azure: Deploy Qwen 3.8 Max through the same Model Catalog. Th

e agent uses Azure AI Agent Service’s document processing tools to parse uploaded contracts (PDF, DOCX). It extracts key clauses: payment terms, liability caps, termination conditions, and change-order procedures. It also flags risky language (e.g., unilateral renewal clauses) and suggests alternative phrasing based on negotiation playbooks stored in Azure Blob Storage. Prompt structure: Few-shot prompting with examples of good vs. risky clauses. The agent outputs a structured JSON object per contract, plus a summary note. Example: “Extract all clauses related to force majeure and indicate whether they match our corporate standard. Highlight any deviations.” Pilot result: Contract review cycle time dropped from 4 hours to 45 minutes per contract, with 94% of flagged risks validated by human reviewers. Agent #3: Fine-Tuned Pricing Agent — Dynamic Market Price Matching and Spend Control Mo

del background: The pricing agent is a fine-tuned variant of a smaller base model (e.g., Llama 3.1 8B) trained on historical purchase order data, supplier price lists, and market indices. Fine-tuning was performed using Azure Machine Learning with LoRA adapters, costing approximately $1,200 for a 10