On-Premise Field Sales Intelligence Agent
Built an ultra-lightweight AI agent running on sales rep laptops that answers product and customer questions in under 100ms—even offline. Trained on 200K+ products and 1M+ customers, it eliminated hours of manual research per sales visit.
The Challenge
Field sales reps were losing deals because they couldn't answer client questions on the spot.
- Sales reps carried a catalog of 200,000+ products with 100+ attributes each—nobody could remember it all
- Customer-specific pricing, past orders, and return history were locked in the ERP system back at the office
- Reps spent 2-3 hours after each client visit researching answers and sending follow-up emails
- By the time answers arrived, competitors had already closed the deal 35% of the time
- Warehouse locations, product availability, and shipping times changed daily—printed materials were always outdated
- New reps took 8-12 months to become productive because of the learning curve on products and customers
The Solution
We developed an on-premise AI assistant that runs entirely on sales rep laptops with sub-100ms response times.
- Lightweight local LLM (Phi-3 Mini quantized to 4-bit) running on standard laptops without GPU requirements
- SQLite database with sqlite-vec extension for vector search across all product embeddings
- Complete product catalog with semantic search understanding natural language queries like "waterproof connectors for outdoor use"
- Customer 360 view showing past orders, return patterns, shipping preferences, and payment history
- Intelligent pricing engine calculating customer-specific discounts based on volume and relationship tier
- Offline-first architecture syncing delta updates when connected—works in warehouses, factories, remote sites
- Voice input support for hands-free queries during warehouse tours
- Integration with mobile barcode scanner for instant product lookups
Implementation
Week 1-2: Data extraction and normalization from ERP, CRM, and product databases
Week 3-4: Embedding generation for 200K products using sentence-transformers
Week 5-6: Local LLM selection, quantization, and optimization for CPU inference
Week 7-8: Query orchestration system with intent classification and entity extraction
Week 9-10: Customer history engine with order pattern analysis
Week 11-12: Delta sync protocol for efficient data updates over limited bandwidth
Week 13-14: Desktop application with voice input and barcode scanner integration
Week 15-16: Pilot with 20 sales reps, gathering feedback and refining responses
Week 17-18: Full rollout with training sessions and ongoing model improvements
The Results
Sales reps transformed from catalog carriers to trusted consultants who answer any question instantly.
"Last week, a client asked about compatibility between three products I'd never heard of. I typed the question, got the answer in a second, and closed a $50K order on the spot. Before this, I would have said "let me get back to you" and lost the deal to our competitor who happened to know."