- 01
Market
Saudi Arabia
- 02
Written for
Head of CX, CTO
- 03
Decision
Deploy a conversational system that handles how customers actually speak
- 04
Languages
EN · AR
What is hard
Four things that decide whether it works
Dialect
A caller in Riyadh, a caller in Jeddah and a Levantine or Egyptian resident are not speaking the same variety, and vocabulary differs on exactly the words a service conversation turns on. Evaluate on recordings from the actual customer base, not on a general benchmark.
Code-switching
"أبي أغير الـ booking حقي لـ next week" is an ordinary sentence, not an edge case. Systems that route an utterance to one language model or the other break on it; the handling has to be within-utterance rather than per-call.
Names
Arabic names carry orthographic variation that is genuine rather than erroneous, and a system that matches strictly will fail to find a customer who exists. Matching has to tolerate the variation while confirming enough to be safe, which is a design decision rather than a threshold.
Numbers
Phone numbers, identity numbers and amounts get spoken in Arabic, in English, or with the digits alternating between them. This is where transcription errors concentrate, and it is why any flow that captures a number needs a read-back confirmation rather than a confidence score.
Interruption and silence
Callers interrupt, pause mid-sentence to check something, and answer a question two turns after it was asked. Handling this well is what makes a voice system feel usable, and it is invisible in a scripted demonstration.
Formality
Register carries meaning in Arabic in ways a direct translation of English copy loses. A system that addresses a customer too casually, or too formally for the context, reads as wrong even when every word is correct.
Escalation
Design the handover before the automation
Every conversational deployment has a failure path, and the quality of that path determines how customers experience the whole system. A caller who is understood immediately never thinks about the automation. A caller who is misunderstood twice and then reaches a human who asks them to start again has had a worse experience than if the automation had never existed.
So the handover is specified first: after how many failed attempts, to whom, and carrying what. Carrying the transcript and the intent the system had inferred is the part that is usually skipped and the part that matters — it is the difference between a handover and a restart. Where the automation runs outside business hours, what happens then is a decision to make rather than a gap to discover.
The second design decision is which intents to automate at all. A system that handles one intent completely is more valuable than one that handles six partially, because a caller learns quickly whether the automation can help them and adjusts. Breadth without depth trains customers to ask for a human immediately, which is the outcome the deployment was meant to avoid.
Pilot
How a pilot is run so it produces a number
1 · One intent, chosen for volume
The single most frequent reason customers make contact. Its volume is what makes the pilot measurable within weeks rather than months. A scope narrow enough to do completely.
2 · Evaluation from recordings
A hundred real interactions on that intent, transcribed and labelled with what the customer actually wanted. This is the benchmark; a scripted set measures the script. A target that reflects how customers speak here.
3 · Escalation path built first
Handover with transcript and inferred intent, working before a single call is automated, including out of hours. A floor under the customer experience.
4 · Live on a share of traffic
A proportion of real contacts, measured against the evaluation set and against what a human handling the same intent achieves. Increase the share when the numbers support it. A decision based on measurement rather than impression.
Scope boundary
How we decide what to automate
What we do
- Evaluate on recordings from your own customers, in the dialects they actually use.
- Build the escalation path with transcript and intent carried across, before any call is automated.
- Confirm captured numbers by reading them back, rather than relying on a confidence threshold.
Recording and processing customer conversations involves personal data, and the obligations on consent, retention and residency come from the data protection regulator. Settle the recording and retention position before building the evaluation set, because the evaluation set is itself a collection of customer data.
Primary sourceCommon questions
What buyers ask
Which dialects can you handle?+
The honest answer is: the ones we have evaluated on your customers' recordings. Gulf and Levantine registers are well represented in available models and Egyptian is generally strong, but performance varies by intent as much as by dialect — a system can handle a booking change in a variety and fail on a complaint in the same one, because the vocabulary is different. So the coverage question is answered by measurement on your traffic rather than by a list.
Should we start with voice or with chat?+
Chat, in most cases. It removes transcription from the problem, which is where the largest share of errors sits, so the first deployment measures whether the intent handling and the escalation design work. Voice then adds one variable to a system whose other parts are known good. Starting with voice means debugging transcription, intent handling and escalation simultaneously, and attributing a bad interaction to the right one is guesswork.
What proportion of contacts can realistically be automated?+
It depends entirely on the intent mix, and quoting a percentage without seeing yours would be meaningless. What is generally true: a small number of intents account for most of the volume, and those intents are where automation pays. The realistic first target is one intent handled well enough that customers stop asking for a human on it, and the number that follows from that is measurable rather than predicted.
In this cluster
Next step
Evaluate on your own recordings
A pilot on one intent, evaluated on recordings of real calls rather than a script, with the escalation path built before the automation.

