Building Conversational Commerce on WhatsApp: What We Learned at Production Scale
After 5,700+ live conversations across multiple business domains, the patterns are clear. Here is what works, what does not, and why the two-stage AI pipeline was the most important architectural decision we made.
WhatsApp has 2 billion users. Most businesses know they should be on it. Very few understand what it actually takes to run a commerce operation there at scale — not a FAQ bot, not a broadcast tool, but a full sales channel where customers browse, compare, add to cart, and check out, all in a chat window.
The Naive Architecture and Why It Fails
The obvious architecture — route every message to an LLM, let it handle everything — fails at scale for two reasons. Cost and latency. An LLM call costs money and takes time. At a few hundred conversations a day it is fine. At thousands of concurrent sessions it becomes both expensive and slow. Customers notice a two-second response time. They leave.
The Two-Stage Pipeline
The solution we settled on is a two-stage AI pipeline. The first stage is fast keyword and intent matching — a lightweight classifier that handles the 85% of queries that are predictable (show me products, check my order, what is the price of X). Only the 15% that are genuinely ambiguous or complex get routed to the LLM. The result: median response time under 800ms, and cost per conversation under $0.02 at production volume.
The Hardest Problem: Session State
A WhatsApp conversation is not a web session. It can span days. A customer can start a cart on Monday, go quiet, and come back on Wednesday expecting to find their items still there. We manage 60+ conversation states — from initial greeting through product browsing, cart management, checkout, payment confirmation, and post-purchase support — with session memory persisted across devices and sessions.
What the Numbers Actually Look Like
After 5,722 production conversations: 40–60% higher conversion than the same business running a traditional e-commerce channel. The reason is frictionless intent — customers describe what they want in natural language and get the right product. They do not need to navigate a catalogue designed by someone else.
The businesses that succeed on WhatsApp commerce are the ones that treat it as a sales channel that requires a commerce strategy — not a customer service channel with a bot bolted on. The architecture follows from that decision.