AI-Powered Procurement Automation

Context

Zetwerk, a manufacturing marketplace processing $3B+ in annual transactions, relied on manual procurement workflows. Buyers uploaded 3D CAD files and specifications, then human operators manually matched them with vendors — a process taking days per RFQ and costing the platform millions in operational overhead. I was engaged as a product consultant to lead the AI automation initiative.

The Problem

The manual buyer-vendor matching process was the single largest bottleneck in the procurement pipeline. With thousands of RFQs per month and a growing vendor network, the platform couldn't scale without fundamentally rethinking how parts were analyzed and matched. Key pain points included: inconsistent matching quality across operators, 3-5 day average cycle time per RFQ, and inability to leverage the rich geometric data embedded in CAD files.

Discovery & Research

I led 30+ stakeholder interviews across procurement teams, engineering managers, and vendor partners to map the end-to-end workflow. The critical insight was that 80% of matching decisions could be predicted from CAD geometry + historical transaction data alone. I conducted a competitive analysis of existing solutions and identified that no platform was using 3D geometry understanding at the feature extraction level — everyone was relying on text metadata.

• ·30+ stakeholder interviews across buyers, vendors, and internal ops
• ·Mapped 47 decision variables in the manual matching workflow
• ·Identified 3D geometry as the untapped data source for automated matching
• ·Evaluated 5 technical approaches with the ML engineering team

Solution

I built the business case and product strategy for a multimodal AI engine that combines B-Rep Transformer-based 3D CAD parsing with historical transaction data to automate buyer-vendor matching. The system extracts manufacturing features (tolerances, materials, complexity) directly from CAD geometry, then matches against vendor capabilities using a learned similarity model. I authored the PRD, defined success metrics, and managed the phased rollout starting with the highest-volume part categories.

Results & Impact

In the pilot cohort, the AI engine delivered a 32% reduction in procurement cycle time and improved matching accuracy by 15% over human operators on standardized part categories. Average RFQ turnaround dropped from 3 days to 1. Based on pilot performance, full-scale rollout was modeled to unlock $80M+ in annual efficiency gains. The system now processes the majority of incoming RFQs with minimal human review.

Key Learnings

The biggest lesson was that domain expertise matters more than model architecture. The breakthrough came from understanding how procurement engineers think about parts — not from a better transformer. Spending weeks on stakeholder interviews before writing a single spec saved months of engineering rework.

• ·Domain depth > model sophistication for enterprise AI products
• ·Phased rollout by part category reduced risk and built internal confidence
• ·Defining clear fallback to human review was critical for stakeholder buy-in