One of the interesting technical challenges in voice AI is connecting natural language phone conversations to structured ordering systems. We recently built an integration between our AI phone agent and Flipdish (a restaurant POS/ordering platform popular in Ireland).
The Architecture
Inbound call → STT → LLM (conversation management)
→ Structured order extraction → Flipdish API
→ Order confirmation → TTS → Caller
The tricky parts:
1. Menu item disambiguation. "I'll have the chicken" — which chicken? The AI needs to handle clarification naturally: "We have Chicken Tikka Masala and Chicken Korma — which would you prefer?"
2. Real-time pricing. Order totals need to be accurate as items are added. The AI maintains a running cart synced with Flipdish's pricing.
3. Delivery zone validation. Before completing an order, verify the address is within the restaurant's delivery zone via the Flipdish API.
4. Concurrent modifications. Menu availability can change mid-call (item sells out). The system needs to handle this gracefully.
Key Technical Decisions
- Cart state management: Maintained in-memory during the call, committed to Flipdish only on confirmation
- Allergen handling: Pre-loaded from Flipdish menu data, queried by the LLM when relevant dietary questions arise
- Payment: PCI-compliant tokenized card capture, or cash-on-delivery flag
Results
For restaurants handling 120+ calls/week, the integration captures 20+ orders that would have been missed during peak hours. Average order value €38, so roughly €760/week in recovered revenue.
The Flipdish API is well-documented and Irish-focused (Eircode support, local address formats), which made the integration smoother than US-centric alternatives.
We're building this at VoiceFleet. Happy to answer technical questions about voice AI → POS integrations.
Top comments (0)