- 01
Market
Saudi Arabia
- 02
Written for
Head of CX, COO, Head of Digital
- 03
Decision
Decide whether a voice agent can carry real call volume
- 04
Languages
EN · AR
Architecture
What happens between the call and the action
Seven stages, and the escalation path is reachable from every one of them. The systems named under each stage are the ones a deployment actually integrates with.
01
Caller
Inbound on the number the business already publishes, in Arabic or English, switching between them mid-sentence.
- Telephony
- SIP
02
Voice layer
Recognition and synthesis tuned on your own call recordings, including the names and terms your callers actually use.
- ASR
- TTS
03
Orchestration
Identifies intent, holds the state of the conversation, and decides which action is permitted at this point.
- Policy
- State
04
Knowledge
Answers grounded in your published material and your records, cited internally so a wrong answer can be traced.
- Retrieval
- CMS
05
Business system
Reads the account, the order or the diary. Reads are open; writes are constrained by the next stage.
- CRM
- ERP
- PMS
06
Allowed action
A short list agreed with you — book, amend, cancel, send. Anything outside it is a handover, not an improvisation.
- Write API
07
Logging
Transcript, intent, action and outcome recorded against the customer record, which is what makes the deployment auditable.
- Audit log
Human escalation
Reachable from every stage — on request, on a second failed understanding, on any signal of distress, and on anything outside the allowed actions. The transfer carries the transcript, so the caller never repeats themselves.
The stages left of the allowed action are cheap to change. The two on the right are where an incident comes from, which is why they are the smallest part of the design and the most reviewed.
The problem
Most voice deployments fail at the handover, not the speech
Speech recognition and synthesis are largely solved for this use case. What still goes wrong is everything around them: the agent answers confidently from stale information, cannot see the caller's account, takes an action it should have confirmed first, or transfers to a human without any of the context — so the caller repeats the whole story and the experience is worse than the queue they started in.
The useful design question is therefore not "can it hold a conversation" but "what is it allowed to do, and what happens at the edge of that". An agent that answers ninety per cent of calls well and mishandles the last ten silently is worse for a business than one that answers seventy and escalates the rest cleanly, because the failures in the first case are invisible until they show up as complaints.
Arabic adds real engineering rather than a configuration flag. Callers move between Gulf dialects and Modern Standard Arabic, switch to English mid-sentence for technical terms, and give names whose transliteration has to reach the CRM in a form that matches the existing record. Testing this needs an Arabic evaluation set built from real calls — translating an English one measures a language nobody is speaking.
Failure handling
What the system does when something goes wrong
These are the questions worth asking any voice vendor, including us. The answers determine whether the deployment survives its first bad week.
The agent is not confident
Confidence below a tuned threshold routes to a human rather than guessing. The threshold is set per intent — a rate enquiry and a cancellation do not warrant the same caution.
What the caller experiences: A transfer with context, not a wrong answer delivered confidently
The business system is down
The agent stops offering actions that depend on it, says so plainly, and either takes a callback request or transfers. It does not accept a booking it cannot write.
What the caller experiences: An honest "I cannot do that right now" and a route forward
The caller asks something out of scope
Intents outside the permitted set are escalated by design rather than attempted. The scope is an allow-list, not a block-list.
What the caller experiences: A prompt handover instead of a plausible invention
The caller wants a person
Honoured immediately, on the first request, without a retention loop. Making this hard is the fastest way to make customers hate an otherwise good system.
What the caller experiences: A transfer, first time of asking
Something went wrong after the call
Every call has a transcript, the retrieved sources, the actions attempted and their results, retained per your policy and searchable.
What the caller experiences: A supervisor can reconstruct exactly what happened
Where it earns its place
Call types that suit a voice agent
High volume, narrow intent, and an action the system can complete. Where a call is rare, emotionally charged or commercially delicate, a person should take it.
Booking, changing and cancelling
Availability lookup, a held slot, a confirmation the caller receives in writing. The intent is narrow and the action is verifiable, which is exactly the shape that works.
Status and account enquiries
Where is my order, what is my balance, when is my appointment. Reads from the system of record after verifying identity to whatever standard your sector requires.
Qualification and routing
Establishing what the caller needs and getting them to the right team with the context attached, so the specialist starts informed rather than at the beginning.
Out-of-hours cover
The clearest case. The alternative is not a human answering — it is voicemail nobody returns, or a missed booking that goes to a competitor at nine in the evening.
Common questions
What buyers ask
How good is the Arabic, really?+
Good enough for narrow, well-defined intents in Gulf dialects and Modern Standard Arabic, and it degrades on unusual names, heavy background noise and long unstructured explanations. We would rather show you that on your own calls than claim a percentage. The evaluation set we build is in Arabic from your recordings, so what you see in testing is what production looks like.
Will callers know they are talking to a system?+
They should, and we build it that way. Disclosure at the start costs nothing in completion rate in our experience and avoids the far worse outcome of a caller realising halfway through and feeling deceived. It also makes the escalation offer natural rather than an admission.
Can it work with our existing phone system?+
Usually. Most modern contact centre platforms and SIP-based systems can route selected numbers or queues to an external endpoint, which is all that is needed. We confirm the specific integration path during scoping rather than assuming it, because the exception — an old on-premise PBX with no SIP trunk — does exist and changes the plan.
What does it cost to run?+
Per-minute rather than per-seat, which changes the economics in both directions: cheap at low volume, and worth modelling carefully at high volume before launch. We model expected cost at your actual call profile during the demonstration phase, because a system that is uneconomic at peak is not a system you want to discover after go-live.
In this cluster
Next step
See it on your own call flows
A working demonstration built on two of your real call flows and your own knowledge base, so you can judge it on your calls rather than a scripted demo. Two weeks, fixed scope.

