Intent detection works by mapping the meaning of an incoming message to one of a defined set of categories the system knows how to act on - order tracking, refund request, appointment booking, and so on - rather than matching literal keywords. This requires the underlying language model to recognize paraphrase, sentence structure variation, and even spelling inconsistencies as expressions of the same underlying need. Accuracy is typically measured by how often the detected intent matches what a human reviewer would classify the message as.
In MENA markets, intent detection has to work across at least three simultaneous variations - Gulf Arabic, Modern Standard Arabic, and English - plus Arabizi and code-switched messages, since the same customer might type different phrasings of the same question depending on mood or habit. If the system only recognizes one register, it will fail to detect intent on a large share of real messages, driving unnecessary escalation and undermining first contact resolution across WhatsApp and other channels.
In Eshal: Eshal's intent detection is trained specifically on Gulf and Levantine Arabic alongside English and Arabizi, so it can map differently phrased or differently spelled messages to the same underlying workflow. This is the layer that lets the AI concierge route order-status style questions correctly regardless of dialect, which in turn supports Eshal's reported first contact resolution and average resolution rate performance.