HumaiX
P1Packaging manufacturing · B2B wholesale

A 24/7 Multilingual WhatsApp Sales Agent

A packaging manufacturer could not answer questions and orders arriving outside business hours, and customers wrote in three different languages. We built a WhatsApp agent that knows the product procedures, takes pre-orders into CRM, and knows its own limits well enough to hand over to a human. After-hours losses went to zero.

24/7

Uninterrupted response

3+

Languages, one system

0

After-hours enquiries lost

The problem

Customers sent questions and orders over WhatsApp at any hour, but the sales team was only there during business hours. Every message arriving after hours waited until the next day, and some were never answered at all. Language compounded it: conversations came in a mix of Turkish, English, and Arabic, and staffing a separate agent per language was not viable at this scale.

How we approached it

We did not build a chatbot; we built a sales specialist that knows the company. It knows the product procedures, can take a pre-order, and can write it to CRM — but above all it knows its limits, and hands over to a human rather than inventing an answer it does not have. We added a layer that distinguishes incoming message types (text, image, voice, document), tested prompt-injection scenarios separately, and documented the edge cases we hit. When business hours start, the human team takes over the conversation without losing context.

Outcome

  • 24/7 responses across several languages including Turkish, English, and Arabic
  • Pre-orders captured automatically and written straight into CRM
  • After-hours enquiry losses reduced to zero
  • Image and voice enquiries handled too, thanks to message-type detection
  • Human handover preserves context, so the customer never repeats themselves

Stack

n8nWhatsApp Business APILLMCRM

What this project generalises to

The real lesson here is not technical, it is about drawing boundaries. The value of a customer-facing system is measured less by how many questions it answers than by how fast it hands over the one it cannot. An agent that refuses to invent is more useful than one that answers everything.

The service this case is evidence for

AI Agent Development

Other case studies

All case studies

Client names are withheld under confidentiality; the sector and region describe the real engagement.

Last updated: 7 August 2026

Talk About a System Like This