WhatsApp Shipment Exception Desk
Representative architecture: the bilingual Arabic/English WhatsApp support agent we delivered for an e-commerce retailer, adapted into a shipment-exception desk for delay, damage and wrong-item cases with a clean human handoff. Live tracking lookups are not included.
- Logistics/3PL
- n8n
- Exception Management
- Arabic/English
Overview
The project at a glance
"Where's my order?" and "it arrived damaged" messages flood support. This representative architecture adapts a self-hosted, bilingual (Arabic/English) WhatsApp agent we delivered for an e-commerce retailer. It has per-customer memory, explicit flows for delay, damage, wrong-item and missing-item cases, and a clean handoff to a human.
The original client was a retailer, not a 3PL. Live tracking lookups (AfterShip, carrier APIs) and handoff into a WMS queue were not part of the build.
What the engagement had to achieve
- Recognize exception types (delay, damage, wrong item, missing item, return, complaint)
- Collect the right details once, with bilingual AR/EN responses
- Remember the customer across turns (per-customer memory)
- Hand off to a human with a structured summary — without inventing delivery dates
The story
Exceptions arrive on WhatsApp, in two languages
What was at risk
The Challenge
In the Gulf especially, customers and consignees don't email, they WhatsApp. A single late or damaged delivery can generate a dozen messages across two languages. The CS team either drowns or bolts on a generic chatbot that confidently invents delivery dates. A 3PL or last-mile operator needs something narrower. It should recognize the exception type, collect the right details once, remember the customer's history and hand off to a person cleanly, without making promises it can't keep.
Delay and damage threads arrive as mixed Arabic/English WhatsApp traffic. Generic bots either stall or invent ETAs.
Free-form chat loses the reference number, photo, and exception type — so the human agent re-asks everything on handoff.
The original build did not include live order-status or tracking lookups. Connecting carrier or WMS tracking APIs would be new work.
How we responded
The Solution
n8n + WAHA keeps conversation data on infrastructure you control.
Responses in Arabic and English, with per-customer memory in Postgres so the agent does not re-ask known facts.
Rather than free-form chat: return/refund, exchange, wrong / missing item, damaged item, delivery delay, complaint. Flow-by-flow regression testing on real customer messages before go-live.
The agent summarizes the case so the human agent doesn't re-ask. For a 3PL, handoff routes to ops or claims.
Tracking data would typically come from Aramex, SMSA, Naqel, Shipox, Quiqup, AfterShip or carrier APIs. Live tracking lookups were not part of the original build.
What we built
The concrete capabilities designed, built, and shipped in this engagement.
Self-hosted WhatsApp stack on n8n + WAHA
Conversation data stays on infrastructure you control. "Self-hosted" refers to this stack — not an official Meta partnership.
Bilingual AR/EN responses
The agent replies in Arabic and English so Gulf WhatsApp-first exception traffic is handled in the customer's language.
Per-customer memory in Postgres
Known facts persist across turns so the agent does not re-ask for reference numbers, photos or prior exception context.
Explicit exception decision-tree flows
Rather than free-form chat: return/refund, exchange, wrong / missing item, damaged item, delivery delay, complaint — mapped here to shipment-exception handling.
Structured human handoff
The agent summarizes the case so the human agent does not re-ask. For a 3PL, handoff routes to ops or claims.
Flow-by-flow regression testing before go-live
Each exception flow was regression-tested on real customer messages before go-live. Live tracking lookup was not included.
The stack
The tools behind the build, and the role each one played.
n8n
WhatsApp Business Platform
PostgreSQL
OpenAI
What changed
Representative architecture: there are no shipment-exception results to report yet. The design guarantees below come from the original retail support agent.
Beyond the launch
Lasting improvements
The changes that keep paying off after the engagement ended.
- Explicit decision-tree exception flows instead of free-form chat that invents ETAs
- Structured human-handoff summary so ops/claims do not re-ask known facts
- Customers are not asked twice for reference numbers or photos they already shared
Swipe to see more
All case studies
Want an outcome like this one?
A short introductory call covers what's holding you up, whether we can help, and the first thing worth building.