Production and hardening · Dubai, UAE corporate-services provider
I led a production WhatsApp system for a Dubai service line that combined approved-knowledge retrieval, structured qualification, document handling, booking, CRM submission, follow-up and human escalation.
The working context
The problem
The system had to be helpful without turning a conversational answer into a guarantee. It also had to survive ordinary messaging problems such as repeated events, media attachments, interrupted conversations and uncertain downstream responses.
My contribution
I shaped the product behavior, qualification structure, knowledge boundaries, CRM handoff, follow-up rules, booking flow and operational safeguards. The work continued through production versions because real conversations exposed edge cases that a demo could not.
I have kept the organisation, product and people private. I also left out exact figures and internal details that could identify the work. The status above tells you how far the implementation actually went.
How I worked through it
If you are working on a similar system, these are the decisions I would examine before choosing tools or adding more automation.
I kept service answers tied to approved material and separated possible eligibility from a confirmed outcome. The model needed a route to say that it did not have enough evidence.
Sessions, qualification fields, follow-ups and CRM submissions lived as structured records. That made retries and human review possible without asking the chat transcript to be the database.
A person taking over needed the reason for escalation, the relevant answers and the latest expected action. Passing a raw transcript would have shifted the work rather than completed the handoff.
Expert insights from the work
These are the lessons I took from the work. I would use them as questions for your own implementation, not as a universal recipe.
Messaging providers retry events. If your workflow cannot recognize work it has already accepted, the customer may receive duplicate replies or the CRM may receive duplicate records.
A fluent answer can still be wrong. Decide when a person must step in by looking at the consequence of the answer, not how certain the wording sounds.
A reminder is only useful while the conversation is genuinely abandoned. Cancel it when the person replies, books, reaches a human or reaches a terminal state.
What production changed
The first working conversation proved the concept. Production work was different. I spent more time on duplicate events, queues, retry behavior, media, follow-up timing and human ownership because those details decide whether your system can be trusted.
Connected disciplines
Continue through the work
If this overlaps with a system you are thinking through, start a relevant conversation.
