AI and automation / 2026
The owner was not mainly losing orders because replies were slow. People were using fake purchase intent to reach her directly, then turning a business conversation into harassment or unwanted personal contact.
The service verifies WhatsApp ingress, restores conversation state, loads the owner's published menu and operating rules, applies deterministic safety and order policy, and uses AI only inside those boundaries. A customer does not reach the owner just by saying they want to order. The owner console is reserved for qualified orders, legitimate support exceptions, ambiguous cases, and credible safety concerns.
Walkthrough
Recreated chat and owner-console screens follow a genuine order, an order outside the published rules, and an abusive thread that receives no reply.
Conversation replay
A customer asks for the menu and orders. The total comes from published prices and the order is not surfaced until it is complete.
My part
Decisions
The core invariant is that saying 'I want to order' is insufficient. A conversation reaches the owner only after deterministic qualification is complete or a legitimate support or safety exception requires human judgement.
The model can help interpret and phrase a response, but application code owns order completeness, owner surfacing, safety state, takeover state, and whether any outbound message is allowed.
Greetings, personal chat, and clearly off-purpose messages are handled by deterministic policy. This reduces needless engagement and avoids sending sensitive conversation to a model unless the configured policy explicitly allows it.
The customer-facing assistant cannot silently mutate published menu or operating rules. The owner publishes those facts and handles exceptional cases through an authenticated console.
Inbound and outbound processing uses duplicate-resistant records and bounded retry behaviour so webhook redelivery or a crash does not create duplicate replies or repeated owner work.
Screens
SYSTEM ARCHITECTURE
Genuine customer, off-purpose chat or abuse
Signature check, normalisation and deduplication
Per-contact ordering, stage and takeover
Owner's menu, windows, fulfilment and boundaries
Deterministic qualification and containment
Classification and phrasing inside policy
Duplicate-safe, stale-aware delivery
Qualified orders, exceptions and safety review
Built with
Work with me
If you are planning something similar, a 30-minute consultation covers feasibility, suitable tools and next steps. From $25.
More work
Sellavie AIAn independent project: multi-tenant conversational-commerce software that connects social messaging with business knowledge, guarded AI, human handoff, orders, billing, analytics and operations.MaintainedAI Automation Control PlaneA runnable multi-tenant control plane for AI-assisted business workflows, combining FastAPI, Next.js, PostgreSQL, Redis, n8n, provider routing, human approval, idempotency, replay, cancellation, audit traces, and deterministic local demonstrations.MaintainedWebhook RelayWebhook reliability tool for receiving, verifying, inspecting, replaying, and forwarding events with bounded retries and a live inspector UI.Maintained