The problem
A Jeddah food distributor’s vans visit hundreds of baqalas a week and management learns about stock-outs when the shop calls angry. Reps won’t fill CRM forms on their phones between stops — the data simply never gets captured, and the distributor flies blind on its own shelves.
The build
No app, no training, no hardware: one WhatsApp number, transcription tuned with the distributor’s brand and shop names, a structuring prompt that outputs visit/order/complaint records, and a dashboard the sales manager already checks. The rep’s whole job is a 30-second voice note per stop — the lowest-friction reporting interface that exists, because it is the app they already live in. Two people build it from APIs in under two months.
Why now
Three years ago dialect speech recognition meant either MSA-only models that mangled Saudi speech or a custom data-collection project. Audio input in the Chat Completions API (Oct 2024) made async voice-note understanding a single call, and Azure ships Saudi Arabic (ar-SA) recognition plus neural voices for the confirmation replies — so the whole loop is bought infrastructure.
First customer
A Riyadh or Jeddah FMCG/building-materials distributor with 5-50 van salesmen: the sales manager is a named payer with a daily reporting headache, reps already use WhatsApp voice notes socially, and per-rep-per-month pricing is trivially below one lost shelf-week. Pilot: one 10-rep team, prepaid, measured on % of visits reported.
Challenges
Transcription accuracy on brand names shouted over a van engine is make-or-break — launch with a per-customer vocabulary list, not generic models. Adoption is behavioral: if the confirmation reply is slow or useless, reps stop sending notes within days. And unit economics need discipline: outbound reminders outside WhatsApp’s service window are charged per template message, so nudges must ride inside free reply windows or the margin leaks.