Three-Agent System for University Admissions Automation: Architecture, Costs, and Model Selection on Azure AI Foundry

The Case for Multi-Agent Automation in University Admissions As of May 23, 2026, university admissions teams face growing application volumes and increasing regulatory complexity. Manual processing leads to bottlenecks, scheduling conflicts, and compliance risks. A recent pilot across 10 universities demonstrated that a multi-agent system deployed on Azure AI Foundry—using Qwen 3.8 Max for application parsing, Llama 5 for compliance checking, and a fine-tuned scheduling agent—reduced admissions processing time by 30% and scheduling conflicts by 25%. This vendor-neutral guide walks through the architecture, model selection trade-offs (including Phi-4 vs Llama 5 for compliance), cost benchmarks, and a decision framework to help education IT leaders evaluate agentic automation versus traditional RPA. Admissions workflows remain largely manual in many institutions. Tasks such as parsing appl

ication documents, verifying regulatory compliance, and creating conflict-free class schedules are labor-intensive and error-prone. A typical mid-sized university processes 15,000–30,000 applications per cycle, each requiring multi-step review. Delays can cost enrollment revenue and frustrate applicants. The 10-university pilot, conducted over two admission cycles, deployed three specialized agents orchestrated on Azure AI Foundry. The system processed applications 30% faster than the previous manual-plus-RPA baseline, and scheduling conflicts (e.g., room double-bookings, instructor time overlaps) dropped by 25%. These results stem from intelligent automation that adapts to document variability and regulatory nuances—something traditional RPA struggles to do. Architecture Overview: A Three-Agent System on Azure AI Foundry The architecture runs entirely within Azure AI Foundry, leveraging

its agent orchestration, model hosting, and integration capabilities. The three agents are: 1. Application Parsing Agent – powered by Qwen 3.8 Max 2. Regulatory Compliance Agent – powered by Llama 5 (with Phi-4 as an alternative) 3. Scheduling Agent – fine-tuned on historical schedule data A central orchestrator manages inter-agent communication via Azure AI Agent Service. Each agent is deployed as a serverless endpoint with auto-scaling. Data flows through Azure API Management, which connects to the university's existing Student Information System (SIS). The orchestrator uses a state machine pattern: parsed documents are passed to the compliance agent, which outputs a pass/fail with annotations, and approved applications proceed to the scheduler. ![High-level architecture diagram: Three agents on Azure AI Foundry connected via orchestrator to SIS and databases] (placeholder description

) Agent 1: Application Parsing with Qwen 3.8 Max Qwen 3.8 Max, released by Alibaba Cloud in early 2026, excels at extracting structured data from unstructured documents (PDFs, scanned forms, emails). In the pilot, it processed an average of 850 documents per hour per endpoint, with a parsing accuracy of 98.2% for standard fields (name, GPA, essays). Cost per document: $0.012 using on-demand pay-as-you-go pricing (Azure AI Foundry pricing page, retrieved May 23, 2026). Key features used: Document Intelligence integration for OCR and table extraction Context window of 128K tokens to handle multi-page application packets Fine-tuning hooks – not used in pilot but available for institution-specific formats Qwen 3.8 Max was chosen over smaller models due to its superior handling of diverse form layouts and ability to process handwritten annotations. Its cost per document remains under the typi

cal labor cost of $0.50 per manual review. Agent 2: Regulatory Compliance Checking – Phi-4 vs Llama 5 For compliance checking (e.g., verifying that application materials meet federal and state requirements, such as FERPA or state-specific data privacy laws), the pilot compared two models: Llama 5 (Meta, May 2025) – 70B parameter dense model, available on Azure AI Foundry as a managed endpoint. Focuses on instruction following and reasoning. Phi-4 (Microsoft, late 2024) – 14B parameter small language model, optimized for cost-efficiency, with strong performance on structured tasks. Accuracy results on compliance rules Llama 5 achieved 96.5% accuracy in flagging compliance violations across 2,000 test cases. Phi-4 achieved 93.7% accuracy, with the gap primarily in nuanced regulatory interpretations (e.g., residency proof requirements). Cost and latency Using Azure AI Foundry pay-as-you-go

pricing (retrieved May 23, 2026): Model Latency per document Cost per 1,000 documents :------ :------------------- :----------------------- Phi-4 2.1 seconds (avg) $0.85 Llama 5 4.8 seconds (avg) $1.40 Phi-4 processed documents 2.3x faster and cost 40% less per document than Llama 5. For institution