GenAI Cuts RFP Assessment from Weeks to Days, Handles 400+ Proposals/Year

Context

A large procurement organization supporting a digital technology and construction programme in Saudi Arabia receives and evaluates more than 400 RFP responses annually. The procurement team was tasked with improving assessment quality and reducing the rate of project failure by making more consistent, defensible decisions during vendor selection. To achieve this at scale, the organization adopted a generative AI–driven assessment solution to help parse, score and surface the critical evidence within each proposal so assessors could focus attention on high-potential bids and reduce time spent on low-performing submissions.

Challenges

The engagement faced a combination of technical, regulatory and operational constraints. Saudi Arabia’s strict data protection and sovereignty requirements prevented use of public cloud LLM services and required all processing to remain on-premises or in tightly controlled local infrastructure. At the same time, no mature cloud-based LLM met the locality and compliance requirements, so an alternative approach was necessary. Incoming RFPs arrived in mixed formats — PDF, DOCX and PPTX — often containing embedded images, tables and diagrams that needed reliable extraction and interpretation. Finally, the procurement team needed an auditable decision trail to explain and defend why a particular bid was awarded, including clear evidence for reject/pass decisions for compliance and governance reviews.

Implementation

The project team deployed a tailored generative AI application into an air-gapped environment running a locally deployed, open-weight large language model. Multiple extraction pipelines were implemented to handle heterogeneous input: text parsers for DOCX and PDF, slide and image OCR for PPTX and embedded figures, and table extraction modules to capture structured cost and schedule data. The AI was used to generate an assessment framework composed of hundreds of binary questions. Each question mapped to a pass/fail outcome, enabling deterministic scoring rules that aggregate into an overall RFP score and ranking.

For every binary question, the interface displayed the AI’s supporting reasoning and the exact evidence extracted from the original documents. The workflow enforced a human-in-the-loop review: procurement analysts could accept, annotate or override each question’s result. Overrides, comments and the original evidence were recorded in an immutable audit trail so every decision carried a full lineage of why it was made. The system also provided configurable thresholds and business rules to automatically filter out low-scoring proposals and surface the most promising bids for expedited review. This combination of automated triage, transparent rationale and reviewer control reduced manual review overhead while preserving governance.

The project’s AI lead coordinated model tuning, question-template design, and validation rounds with domain SMEs from procurement and construction to ensure the binary questions aligned with technical and contractual risk factors. Model outputs were continuously validated against historical RFP outcomes to refine question weighting and scoring logic.

Results

The air-gapped, locally hosted GenAI solution transformed the procurement workflow. Average time to assess an RFP and reach an award decision fell from several weeks to a few days, enabling faster project mobilization and procurement cycles while still meeting compliance requirements. The automated binary-question scoring and triage reduced the volume of proposals requiring deep manual review, allowing the procurement team to concentrate on a smaller subset of high-potential vendors.

Full end-to-end traceability was provided for every decision: extracted evidence, AI reasoning, reviewer actions and override justifications are stored and accessible, enabling robust auditability and a defensible posture for contract awards. The combination of improved assessment quality and more consistent, evidence-based decisions contributed to a measurable reduction in project failure risk and greater confidence in vendor selection. The solution demonstrated that, even under strict data sovereignty constraints, a locally deployed generative AI pipeline can scale to handle 400+ proposals per year while shortening lead times and strengthening governance.