WhatsApp Sales Agent

Context

The store already lived in Shopify. Buyers were already on WhatsApp. The failure mode of a naive bot is well known: it invents a price, confirms stock it does not have, and a human has to unwind the mess.

My ownership

• Agent on WhatsApp wired to Shopify as the source of truth.
• Tools for live catalog, price and inventory.
• Draft orders instead of a hallucinated checkout.
• OpenAI / DeepSeek failover.
• Human handoff when the tools cannot answer or the buyer needs a person.

Constraints

• Never invent price or availability.
• WhatsApp is the interface. Answers about price and stock come from Shopify tools.
• If the model provider fails, fail over or hand off. Sales does not stop on one API.

Architecture

WhatsApp message
  → agent
  → tools: Shopify catalog / price / stock
  → tool: create draft order
  → OpenAI, fail over to DeepSeek
  → reply on WhatsApp
  → human handoff on tool failure, unsupported intent, or an explicit escalation rule

AI architecture

• The model is not the catalog. Price and stock come from Shopify tool calls.
• Vision when the buyer sends a photo of a product.
• Failover across providers.
• Handoff runs on a tool failure, an unsupported intent, or an explicit escalation rule.

Tradeoffs

The model does not create the charge. It opens a Shopify draft order. Payment stays with a human or the existing checkout.

Impact

The agent reads Shopify for catalog, price and stock, opens a draft order, and hands off to a human when the tools cannot answer.

What I learned

• Catalog, price and stock are tool results. The model does not invent them.
• OpenAI and DeepSeek are both in the request path, with failover.
• Human handoff is a branch in the same flow as the draft order.

Example exchange

Illustration of the tool path. Not a customer transcript.

Buyer: Do you have the 2kg bag?
Agent: Shopify shows 14 in stock. I can open a draft order.
Buyer: Yes.
Agent: Draft order is open. A person on the team confirms payment.