Customer Service Automation

Automation Process

1. Fuzzy Matching System: The biggest technical challenge was handling voice recognition with cannabis strain names. Customers say "Gorilla Glue" but products are listed as "GG#4." Built a custom algorithm that scores product similarity, handles abbreviations and voice errors, and returns ranked results. GPT generates voice aliases so the AI knows "weed," "bud," "flower," and "cannabis" all mean the same category.

2. Integration Architecture: System pulls Lightspeed inventory every 15 minutes, caches in n8n, and serves responses within 2-3 seconds. AI checks which location customer asks about, searches that location's inventory, and provides accurate availability. VAPI voice interface handles natural conversation flow.

3. Voice Personality & Training: Programmed AI with friendly budtender persona that explains strain effects, suggests alternatives when out of stock, and escalates complex medical questions to humans. Processed hundreds of real calls to refine fuzzy matching and improve slang understanding. SMS follow-up sends product menus after calls.

‍

‍

‍

Results

The AI voice receptionist now handles 90% of incoming calls automatically, saving each location approximately 150 minutes daily of staff time. Zero calls go to voicemail during business hours, and the system operates 24/7 including after-hours when customers can still get information and prepare their shopping lists for next-day visits.

Customer satisfaction increased significantly, reflected in improved Google review scores specifically mentioning "always able to get someone on the phone" and "quick answers to product questions." The dispensary recovered an estimated $180,000+ in annual revenue from customers who previously abandoned purchases due to unanswered calls or would have called competitors instead.

The system's accuracy improved over time as the fuzzy matching algorithm learned from real queries. Staff report that the AI often provides better product information than newer budtenders because it has instant access to complete inventory data, THC percentages, and product descriptions across all locations.

‍

‍

Conclusion

This project demonstrates that effective voice AI isn't just about speech recognition - it's about building intelligent systems that understand industry-specific language, handle real-world messiness like voice recognition errors and nickname variations, and integrate deeply with existing business systems.

By focusing on the fuzzy matching problem and creating a robust multi-tier architecture, we built a solution that doesn't just "work" in ideal conditions but thrives in the chaotic reality of customer service calls with background noise, accents, slang, and imprecise questions. The result is a system that genuinely augments human staff rather than attempting to replace them, allowing budtenders to focus on the high-touch, high-value in-store experience while AI handles repetitive phone inquiries efficiently.