ProcureAI: An AI-Assisted Procurement Platform with Multi-Criteria Vendor Intelligence and Agentic Workflow Automation
A human-in-the-loop AI procurement platform combining deterministic multi-criteria vendor evaluation, LLM-assisted recommendation and explanation, tool-augmented interaction, and approval-controlled agentic workflow automation.
Research Motivation
Procurement involves interconnected activities that often require manual comparison and fragmented communication. The system addresses the need to separate reproducible vendor evaluation from LLM-assisted interpretation and controlled workflow automation.
Problem Statement
How can we assist procurement decision-making across multiple criteria (price, delivery time, reliability, rating) using LLMs without granting generative AI direct authority over consequential purchasing actions?
Methodology & Architecture
Quotations are ranked using an explicit weighted model (Price: 35%, Delivery Time: 20%, Reliability: 30%, Vendor Rating: 15%). The numerical ranking is then provided as context to an LLM for natural-language interpretation, while a tool-augmented agent coordinates tasks and strictly requires human approval before purchase order generation.
System Architecture
A modular monolithic architecture (React, Node.js, Express, PostgreSQL, Prisma, Docker) featuring a Presentation Layer, Application & Workflow Layer, AI Intelligence Layer (Multi-Criteria Scoring + LLM Recommendation), and an Agentic Procurement Workflow with a firm human-approval boundary.
Experiments
End-to-end functional verification across 61 test cases, combined with a controlled vendor evaluation covering 20 procurement scenarios (111 vendor quotations). Weight-sensitivity analysis demonstrated the top-ranked vendor changing in 18 of 20 scenarios under alternative weighting configurations.
Results & Findings
Achieved 100% pass rate among 60 executed functional test cases. Proved that the deterministic multi-criteria model responds robustly to varying procurement priorities.
Discussion
Future Work
Exploring Retrieval-Augmented Generation (RAG) on historical procurement data, predictive vendor risk analysis, and larger controlled evaluations of recommendation quality and response time.