Conversational AI · Commerce & support
AI WhatsApp Commerce & Customer Support Agent
Project overview
Customers already message businesses on WhatsApp; most of what they ask is answerable from data the business already holds. We built an agent that lives in that thread — handling product discovery, order status, returns and refunds against live systems rather than canned replies — with an operations dashboard behind it for the queue, escalations and COD verification.
A full session: order lookups returning status, courier and tracking, a returns-policy answer, a refund and cancellation flow, and the operations dashboard behind it.
The challenge
Support queues fill with questions that have a definite answer sitting in an order system — where is it, can I return it, what did I pay. Answering them by hand is slow for the customer and expensive for the business, but answering them badly with a scripted bot is worse than not answering at all.
- Answer product, order and returns questions directly in WhatsApp
- Read live order state rather than replying from a script
- Escalate to a human with the conversation attached, not a ticket number
- Verify cash-on-delivery orders before anything ships
- Stay reliable when a user deliberately tries to break the agent
Our solution
A conversational agent on the WhatsApp Business API, grounded in the store’s own catalogue, policies and order data. It resolves what it can resolve, and routes the rest to a person with the full thread attached.
- Product discovery and support inside the customer’s existing WhatsApp thread
- Live order lookup returning status, items, total, courier and tracking number
- Returns and refund answers grounded in the store’s actual published policy
- Cancellation and refund flows handled in-conversation with confirmation steps
- Human hand-off carrying the full conversation context
- Operations dashboard: message queue, escalations, COD verification, security log
Technical implementation
The agent is grounded rather than free-associating: order and policy answers are retrieved from the client’s systems and passed to the model as context, so a wrong answer requires wrong data rather than a hallucination. Guardrails are enforced independently of the prompt, which is what holds up when the model is attacked directly.
- WhatsApp Business API for messaging and delivery receipts
- Retrieval grounded in the store catalogue, policy documents and live order records
- Order-system integration for status, courier and tracking lookups
- Guardrails constraining the agent to support scope, independent of user input
- Escalation routing with conversation context preserved
- Operations dashboard for queue state, escalations and COD verification
Technologies
- WhatsApp Business API
- Conversational AI
- RAG
- LLM Guardrails
- Order Management Integration
- Human-in-the-Loop
- Operations Dashboard
Key results
- Order, returns and refund questions resolve inside the chat, without a queue position or a ticket number
- Answers carry real detail — order number, items, total, courier and tracking — because they are read from the order system
- Escalation hands a human the whole conversation, so the customer never repeats themselves
- Prompt-injection resistance is demonstrable, not asserted: in the recorded demo the agent is told to adopt an “unrestricted” persona and ignore its rules, and it declines and returns the conversation to support
- COD orders are verified through a defined queue before dispatch rather than by ad-hoc phone calls
Why it matters
The interesting question about a support agent is not what it does on a cooperative customer — it is what it does on a hostile one, and on the edge cases where being confidently wrong is expensive. Grounding answers in live systems and enforcing scope outside the prompt is what makes the difference between a demo and something you can leave pointed at real customers.
Services behind this project
AI Chatbots & Customer Support
Support agents across chat, email and messaging
RAG & LLM Integrations
Retrieval grounding against real business data
Workflow & Process Automation
Escalation, verification and ops tooling