A Hybrid AI Framework for Automating High-Stakes Procurement Workflows

Priyanka Choudhary, Ratinder Paul Singh Ahuja

Keywords: LLM Application, Enterprise AI, AI for finance, AI Automation, Generative AI Applications

Abstract: The automation of complex, human-centric workflows remains a central challenge in enterprise artificial intelligence. We address this through the Purchase Requisition (PR) approval process, a time-intensive task where manual reviews can span from 15 minutes to 2 hours, creating significant operational bottlenecks across thousands of annual submissions. To overcome these inefficiencies, we propose an Augmented Intelligence framework, an automated PR Review Assistant engineered in Python, that integrates deterministic logic with probabilistic large language model (LLM)–driven validation. Our system employs a multi-agent architecture: an LLM-EXTRACT agent transforms unstructured attachments into machine-readable data, an LLM-VAL agent performs contextual validation against complex business rules, and an LLM-REPORT-GEN agent generates natural-language rationales for review outcomes. Validation is governed by a Hybrid Confidence Score, combining deterministic gates with LLM confidence outputs to ensure both rigor and interpretability. Benchmarking across leading LLMs reveals complementary strengths, with OpenAI’s models demonstrating the highest consistency, Meta’s excelling in tax reasoning, and others offering nuanced trade-offs. Deployed in production for Marketing and G&A departments, the framework processes over 4,000 PRs annually, achieving a 62% reduction in manual review time, a 37% faster procurement cycle, and a 66% decrease in submission errors. User feedback highlights faster approvals and enhanced compliance, underscoring the system’s capacity to reallocate human effort toward higher-value strategic assessments. This work establishes a generalizable model for human-in-the-loop automation, bridging deterministic and probabilistic methods to deliver scalable, explainable, and trustworthy enterprise AI.

Submission Number: 403