Smarter Purchase Order Review with Agentic AI

To modernize and scale purchase order review, the DxHub team designed and implemented an agentic, generative AI–driven workflow that ingests procurement documents and references a structured set of compliance and risk checks to support buyer decision-making. Rather than relying on a single model pass, the solution orchestrates multiple specialized AI agents, each responsible for a distinct aspect of the review process, enabling greater accuracy, transparency, and extensibility. At the front of the workflow, a Document Parser Agent classifies incoming files by document type and dynamically selects the appropriate parsing strategy and prompt. Parsed outputs are written to a shared state that downstream agents can reliably consume. A Data Security Classification Agent then evaluates service and vendor information to determine data sensitivity levels (P1–P4), using LLM reasoning to detect indicators of sensitive or regulated data and returning structured JSON with both classification and justification. In parallel, a Purchasing Categories Mapping Agent applies contextual reasoning to map each purchase order to the appropriate procurement category, supporting downstream policy evaluation and reporting. Additional agents perform targeted compliance checks, including a PHI Agreement Check Agent that scans documents for Protected Health Information (PHI), HIPAA-related clauses, and required agreement language. Once all checks are complete, a summarization agent consolidates outputs from each agent into a single, human-readable summary that clearly surfaces risks, missing information, and review signals for procurement professionals. The solution ensures scalability, security, and auditability while allowing UC San Diego to evolve rules, agents, and checks over time—demonstrating how agentic AI can augment complex administrative workflows without replacing human oversight.