We Built an AI Agent That Ate 200-Page RFQs for Breakfast: Here's What Happened

"AI isn't the future of engineering: it's our present." Mark Zuckerberg, Meta Connect 2025

Introduction
The Problem Nobody Wants to Admit Is That Bad
First Contact: What the Client Actually Needed
The Architecture: Building Agents That Actually Work
The Human-in-the-Loop Layer
The Integration Layer: Theory Meets Chaos
What Happened When We Went Live
The Honest Lessons: What Was Harder Than Expected
What This Means for Procurement's Future

1. Introduction

There's a moment every procurement manager knows too well. It's 11:47 PM. There's a 400-page RFQ sitting on the desk: or more accurately, open in six browser tabs, printed in three different versions, with sticky notes creeping up the monitor like ivy.

Somewhere in those pages is a compliance clause that will disqualify a vendor. Somewhere else is a line item that doesn't match the scope. And the bid submission window closes in 13 hours.

This is not an edge case. This is Tuesday.

We just shipped a full end-to-end AI-powered procurement platform for a public works client: and it changed everything about how that 11:47 PM moment plays out.

This is the story of how we built it, what broke along the way, and what it means for every procurement team still fighting documents with highlighters and spreadsheets.

2. The Problem Nobody Wants to Admit Is That Bad

Let's be honest about what traditional procurement actually looks like before we talk about fixing it. A single public works RFQ can run anywhere from 80 to 600 pages. It includes technical specifications, legal compliance requirements, environmental clauses, insurance minimums, subcontractor eligibility rules, bonding thresholds, and bid form instructions: all written in the specific brand of dense bureaucratic English that makes lawyers comfortable and humans miserable.

A procurement team receives dozens of vendor bids in response. Each bid is structured differently. Each vendor interprets the scope differently. Some miss line items. Some substitute materials without flagging it. Some technically comply but practically don't.

The team's job? Read everything. Compare everything. Flag everything. Miss nothing.

In practice, teams miss things all the time: not because they're incompetent, but because this is an inhuman volume of detail to track manually. The consequences range from contract disputes to project delays to regulatory violations that cost governments millions.

The real cost of slow, error-prone procurement isn't just wasted time. It's awarded contracts that shouldn't have been awarded, missed savings, and compliance failures that hit hard.

3. First Contact: What the Client Actually Needed

When we first got on a call, the client said they needed "something to help with document review." After three hours of discovery, we understood the real scope:

Concurrent Projects: They were managing multiple concurrent public works projects simultaneously.
Document Overload: Their team was four people. The document load was a full-time job for forty.
Compliance Risk: They were one missed compliance flag away from a serious regulatory problem.

What they actually needed wasn't a tool to help read documents faster. They needed an AI agent system that could ingest procurement documents, extract structured requirements, and evaluate vendor responses against those requirements automatically.

4. The Architecture: Building Agents That Actually Work

Here's where most AI procurement demos fall apart: they show you a chatbot that can summarize a document. That's not agentic AI. That's a party trick.

Real agentic AI in procurement means building systems where the AI can take sequences of actions autonomously.

Our stack was built in Python and TypeScript, with purpose-built agents for each stage of the procurement pipeline.

Agent 1: The Intake Agent

The first problem is that documents arrive in every format imaginable: PDFs, Word docs, Excel bid forms, email attachments, and even scanned paper documents. Our Intake Agent handled document ingestion, normalization, and classification.

Agent 2: The Requirements Extraction Agent

This was the hardest agent to build. Every RFQ contains explicit requirements and implicit specifications buried in dense language. This agent produced a structured requirements matrix: a complete inventory of every obligation that a compliant bid needed to address.

Agent 3: The Bid Comparison Agent

When you have twelve vendors submitting responses to a 300-item matrix, the comparison problem is structurally complex. Our Bid Comparison Agent located corresponding responses, extracted positions, and flagged any gaps or substitutions.

5. The Human-in-the-Loop Layer

We designed the entire platform around the principle that humans must remain accountable for consequential decisions. The AI handles the extraction, comparison, and flagging. Humans handle the judgment.

Every workflow had clearly defined approval gates: moments where the system paused, surfaced its findings and reasoning to a human reviewer, and required explicit sign-off before proceeding.

6. The Integration Layer: Theory Meets Chaos

The systems our agents needed to talk to were not clean APIs. They were spreadsheets, shared drives, email threads, and legacy tools. We built a robust integration layer that connected the AI agent system to:

Document repositories (SharePoint, Google Drive)
Procurement tracking spreadsheets
Email systems for vendor communications
External vendor databases for cross-referencing

7. What Happened When We Went Live

The first real test was a live RFQ evaluation. Fourteen vendors. Responses to a 280-item requirements matrix.

Historical Timeline: Eleven working days.
With AI Agent System: Four hours for the initial analysis pass.

Total time to reach a recommendation: two days, including human review time at every approval gate. That's not an incremental improvement: that's a structural change in what's possible.

8. The Honest Lessons: What Was Harder Than Expected

Document quality was the biggest enemy. Scanned PDFs with poor OCR and handwritten bid forms required significant engineering time to handle robustly.

Prompt reliability also required obsessive iteration. "Usually gets it right" isn't good enough when the output feeds a procurement decision. We ran extensive testing against historical documents until we hit the necessary reliability thresholds.

9. What This Means for Procurement's Future

We're at an inflection point. Agentic AI in procurement is not a future technology. It's here, it works, and it's being deployed on real projects with real consequences.

The path isn't to replace your procurement team with AI. The path is to give your team AI agents that handle the volume, so the humans can focus on the judgment that only humans should make.

If your team is fighting procurement documents with highlighters and prayer, we've been there: and we know the way out. Reach out to talk about what agentic AI could do for your procurement pipeline.