Problem
The restaurant chain operates 14 quick-service locations across three states and was struggling to keep up with the surge in digital ordering demand. During peak hours, front-line staff were overwhelmed by phone orders, which led to long hold times, misheard items, and frequent order errors. The chain had experimented with third-party delivery marketplaces, but the high commission fees and lack of access to customer data made those channels unsustainable for a business operating on thin margins.
The operations team also faced a visibility problem. Each location maintained its own spreadsheet for orders, inventory, and customer feedback, and there was no unified view of what was selling, what was running low, or which promotions were actually driving foot traffic. Managers spent hours each week reconciling numbers instead of coaching staff and improving service quality. The leadership team recognized that a patchwork of disconnected tools was the root cause, but they did not want to rip and replace their existing point-of-sale system.
They needed a solution that met customers where they already were — on their phones — without forcing a new app install, and that could feed clean, structured data back into the business so managers could make faster, better-informed decisions.
Solution
We designed a conversational ordering assistant accessible through WhatsApp, the channel the restaurant’s customers already used daily. A LangChain-powered agent handles the entire ordering flow: it greets the customer, walks them through the menu, answers questions about ingredients and allergens, suggests upsells based on the current cart, and confirms the order for pickup or delivery. The agent is grounded in a structured menu catalog stored in PostgreSQL, so prices and availability are always accurate and updates propagate instantly.
Behind the scenes, n8n orchestrates the business logic. When a customer confirms an order, an n8n workflow writes the order to PostgreSQL, pushes it to the kitchen display system via webhook, sends a payment link, and triggers a follow-up message 20 minutes after the promised pickup time to request feedback. The same workflow updates inventory counts and flags low-stock items to the shift manager’s dashboard. This orchestration layer is what turned the assistant from a chatbot into an operational system.
Because the entire flow runs through WhatsApp Business API, the restaurant owns the customer relationship. Every conversation is logged, consented, and available for segmentation, which has opened the door to targeted promotions and loyalty rewards that were previously impossible on marketplace platforms.
Architecture
The system is split into three layers: the conversational layer, the orchestration layer, and the data layer. The conversational layer is a FastAPI service hosting the LangChain agent, which uses a combination of tool-calling and retrieval-augmented generation to answer menu questions accurately. The agent maintains short-term memory of the in-progress order and long-term memory of the customer’s preferences, stored in PostgreSQL alongside the order history.
The orchestration layer is built entirely in n8n. Each business event — order placed, payment confirmed, order ready, feedback received — is a discrete workflow that can be versioned, tested, and rolled back independently. This gave the restaurant team a visual way to understand and modify business logic without touching code, which was a key requirement for their non-technical operations managers.
The data layer is a single PostgreSQL database that serves as the source of truth for orders, customers, inventory, and feedback. A read replica feeds the analytics dashboard, so reporting queries never impact the ordering flow. All customer-facing messages are templated and approved through the WhatsApp Business API to ensure compliance and deliverability.
Workflow Diagram
Customer (WhatsApp)
│
▼
WhatsApp Business API ──► Webhook ──► LangChain Agent (FastAPI)
│
▼
PostgreSQL (menu, orders, customers)
│
▼
n8n Orchestration
/ | \
▼ ▼ ▼
Kitchen Display Payment Feedback
Link (20 min later)The diagram above shows the end-to-end flow of a single order. The customer initiates the conversation in WhatsApp, the LangChain agent handles the dialogue and writes the structured order to PostgreSQL, and n8n fans out the downstream actions to the kitchen, payment, and feedback systems. Each arrow represents an asynchronous, retryable event, so a temporary outage in one downstream system never blocks the customer’s ordering experience.
Technologies
The conversational agent is built with LangChain, which provides the tool-calling framework and memory primitives that make multi-turn ordering natural. The agent is hosted in a FastAPI service that exposes a single webhook endpoint for incoming WhatsApp messages and scales horizontally behind a load balancer.
n8n serves as the orchestration backbone. Its visual workflow editor allowed the restaurant operations team to review and approve each business rule before it went live, and its built-in retry and error-handling features meant transient failures in third-party services were handled gracefully without custom code.
PostgreSQL is the persistent data store for all order, customer, and inventory data. The WhatsApp Business API provides the messaging transport layer and ensures all customer-facing messages comply with Meta’s template and consent policies.
Results
Within three months of launch, a substantial portion of all orders across the restaurant’s 14 locations were placed through the WhatsApp assistant, completely bypassing the phone line and reducing reliance on high-commission marketplaces. The conversational upsell prompts — which suggest complementary items based on the current cart — drove a meaningful increase in average order value compared to phone orders.
Phone call volume dropped significantly, which freed front-line staff to focus on in-store customers and food preparation rather than answering repetitive questions. Shift managers reported noticeably calmer kitchens during peak hours and a measurable improvement in order accuracy, since the structured conversational flow eliminated the misheard-item errors that plagued phone ordering.
The unified PostgreSQL data layer also delivered an unexpected benefit: for the first time, the restaurant’s leadership could compare performance across all 14 locations in real time and identify which promotions, menu items, and times of day were driving the most revenue, enabling data-driven decisions that were previously impossible.
Lessons Learned
The biggest lesson was that conversational design matters as much as the underlying AI. Early prototypes of the agent were technically accurate but felt robotic, and customers abandoned the flow. Once we invested in natural, context-aware dialogue — acknowledging previous orders, suggesting the customer’s “usual,” and handling corrections gracefully — completion rates climbed sharply. The lesson: treat the conversation as a product, not just a technical interface.
A second lesson was the value of decoupling orchestration from the agent itself. By keeping business logic in n8n rather than hard-coding it into the LangChain agent, the operations team could adjust workflows — adding a new payment provider, changing the feedback timing, or launching a promotion — without a developer deployment cycle. This separation of concerns is what made the system sustainable for a non-technical team to own long-term.
Finally, we learned that owning the customer channel is strategically critical. The first-party data the restaurant now collects through WhatsApp has become the foundation of their loyalty program and marketing strategy, something that was completely out of reach when they depended on third-party marketplaces.