ProcurePilot Autonomous AI Agent for Procurement

An AI-powered procurement assistant that autonomously extracts RFQ requirements, researches vendors via web browsing, negotiates prices, delivers optimized quote comparisons

Comment

Inspiration

Procurement teams waste countless hours on repetitive tasks: parsing RFQs, researching vendors, making phone calls, and comparing quotes. We envisioned an AI agent that could handle the entire procurement workflow autonomously, from document intake to vendor negotiation, freeing humans to focus on strategic decisions.

What it does

ProcurePilot is a fully autonomous procurement agent that:

  1. Extracts Requirements: Parses uploaded RFQ documents to identify line items, quantities, specifications, and target prices
  2. Researches Vendors: Launches an AI agent that browses Google in real-time to find B2B suppliers, visiting websites to collect pricing, availability, and contact information
  3. Negotiates via Voice: Automatically calls vendors using AI voice agents, discusses requirements, negotiates pricing, and handles objections—all in natural conversation
  4. Delivers Insights: Presents a comparison matrix with vendor quotes, savings calculations, lead times, and availability, complete with call recordings and transcripts

The entire process runs autonomously in a live dashboard where users can watch the AI agent work in real-time.

How we built it

Frontend: Built with React + TypeScript + Vite, using shadcn-ui components for a polished interface. The live agent workspace displays real-time activity feeds, extracted requirements, vendor quotes, and voice call logs.

Backend: Supabase Edge Functions (Deno) orchestrate the workflow:

  • process-rfq: Extracts line items locally via regex, then launches AGI.tech browsing sessions
  • poll-agi-session: Continuously polls AGI API for vendor research results
  • make-call: Initiates AI voice calls via Telnyx with Minimax TTS for natural speech

AI Browsing: AGI.tech's agi-0 agent autonomously navigates Google, clicks search results, extracts supplier information, and structures data into JSON

Voice AI: Telnyx Call Control API with programmable SIP for voice calls, integrated with AI models for conversational negotiation

Database: PostgreSQL (via Supabase) stores RFQs, line items, vendors, quotes, and voice call logs with real-time subscriptions

Challenges we ran into

  1. AGI API Authentication: Initially used the wrong API key format, causing sessions to fail silently with Google OAuth errors. Fixed by switching to UUID-format key.
  2. Session Lifecycle Management: The Edge Function was immediately deleting AGI sessions in a finally block, causing browsers to disappear instantly. Removed auto-cleanup to let sessions complete naturally.
  3. Polling Complexity: Frontend needed to poll multiple data sources (RFQ status, activity logs, AGI sessions, vendor quotes) without overwhelming the database. Implemented staggered polling intervals (2-3 seconds).
  4. Voice Call Integration: Telnyx requires verified phone numbers for production. Built demo mode fallback to simulate calls during development while keeping the full workflow intact.
  5. Real-time Coordination: Synchronizing three async processes (extraction → browsing → calling) required careful status management and activity logging to provide accurate progress updates.

Accomplishments that we're proud of

  • Truly Autonomous: The agent handles end-to-end procurement without human intervention—from document upload to vendor negotiation
  • Real Web Browsing: Actually browses Google and supplier websites in real-time, not just API calls
  • Voice Negotiation: AI agents conduct natural phone conversations with vendors, handling objections and confirming details
  • Production-Ready Architecture: Uses serverless Edge Functions, real-time database subscriptions, and external AI services for scalability
  • Live Transparency: Users watch the AI work in real-time through activity feeds and browser visualization

What we learned

  • AI Agent Orchestration: Coordinating multiple AI services (browsing, voice, LLMs) requires careful error handling and fallback strategies
  • Session Management: Cloud-based browser sessions need proper lifecycle management—create, monitor, collect results, and let them expire naturally
  • Async Workflow Design: Edge Functions with long-running AI tasks need fire-and-forget patterns with frontend polling rather than blocking requests
  • Voice AI UX: Real-time voice calls need immediate feedback (ringing status, duration tracking) even when actual transcripts arrive later

What's next for ProcurePilot - AI-Powered Procurement Agent

  1. Multi-Item Intelligence: Bulk negotiation strategies when purchasing multiple items from the same vendor
  2. Learning from History: Analyze past negotiations to improve pricing strategies and vendor selection
  3. Contract Generation: Automatically draft purchase orders and contracts based on accepted quotes
  4. Approval Workflows: Route high-value quotes through human approval chains before commitment
  5. Vendor Relationship Management: Track vendor performance, delivery reliability, and price trends over time
  6. Multi-Language Support: Enable global procurement with automatic translation in voice calls and document parsing

Built With

Share this project:

Updates