The challenge
Al Marwa runs 80 retail locations across Dubai, Abu Dhabi, Sharjah, and the Northern Emirates. Their contact center fielded 14,000 calls a week — store hours, stock availability, return policy, loyalty points, click-and-collect status — split roughly 60/40 Arabic to English. Hiring fully bilingual agents at scale was getting expensive, and Khaleeji dialect callers were routinely transferred two or three times before reaching someone who understood them.
What we built
A single voice agent fronting both languages, deployed on Twilio Voice with a custom orchestration layer:
- Speech-to-text via a fine-tuned Whisper model adapted for Khaleeji and Levantine Arabic dialects, with code-switching detection (callers often mix English brand names mid-Arabic sentence).
- Reasoning and intent handling powered by Anthropic Claude with a tool registry wired into Al Marwa's SAP retail backend, Salesforce loyalty platform, and store-level inventory feeds.
- Text-to-speech using ElevenLabs voices cloned with explicit licensing — one male and one female persona per language — to match the brand's existing IVR identity.
- A compliance layer that masks PII before any text ever leaves UAE-resident infrastructure (Azure UAE North), satisfying TDRA data residency requirements.
Results
92% of calls now resolve without human handover. Average handle time fell from 4 minutes 10 seconds to 1 minute 45 seconds. Arabic-language CSAT climbed from 3.9 to 4.7 out of 5 — finally matching the English experience. The contact center reduced agent headcount by 34 roles through natural attrition while extending service hours from 12 to 24 per day.