Every voice AI vendor demo works. That is what a demo is for. The interesting question is what happens in month four, when the calls are real, the accents are unfamiliar and the edge cases arrive.
Voice AI is being sold into the Gulf at scale. xAIa, a Dubai company in the PRO TECHnology group, publishes figures including more than a million production calls, 63-plus agents in production and 350,000 calls a month. Whether or not a particular vendor's numbers hold, buyers across the region now face the same evaluation problem.
Here is how to approach it.
1. What does a "call handled" mean?
Volume claims are the headline number in this industry and the least standardised. A "handled" call might mean fully resolved without human involvement, or answered and immediately transferred, or a thirty-second outbound that reached voicemail.
Ask for the containment rate — the share of calls that ended without a human — and ask how transfers and abandons are counted. A vendor who has the number readily is running a measured operation.
2. Has it been tested on your callers' Arabic?
This is the question that matters most in this region and gets asked least. Support for "Arabic" is not one capability. Egyptian, Gulf, Levantine and Modern Standard Arabic differ enough that performance on one predicts little about another, and real callers code-switch into English mid-sentence.
Insist on a test against a sample of your own recorded calls, not the vendor's. Any supplier confident in the system will agree.
3. What happens when it does not know?
The failure mode that damages a brand is not the system saying "I cannot help with that". It is the system confidently saying something wrong.
Ask to see the escalation path, and ask what the agent is permitted to assert without a source. Then ask to hear a recording of a call where it escalated.
4. Who is accountable, and what is logged?
For anything regulated — financial services, healthcare, insurance — the compliance question decides the purchase. What is recorded, where is it stored, how long is it kept, and can you reconstruct exactly what the system told a customer eleven months ago?
That vendors now sell governance and compliance-analysis products alongside the agents themselves suggests the market has learned this the hard way.
5. Where does the data live?
Voice data is personal data. For Egyptian organisations, Egypt's Personal Data Protection Law (Law No. 151 of 2020) governs its processing, and firms serving European customers additionally fall under GDPR. Cross-border processing needs a lawful basis and appropriate safeguards.
Ask where recordings are processed and stored, which sub-processors touch them, and whether the transcripts are used to train models.
6. What does month 13 cost?
Per-minute and per-call pricing looks cheap against salaries at pilot volumes and behaves differently at scale. Model the cost at three times your expected volume, and ask what happens to pricing at renewal once the system is embedded in your operations and switching means re-integrating everything.
7. What happens to the team?
This is a management question, not a technical one, and it determines whether a deployment succeeds. The staff who currently handle these calls know every edge case the system will hit. If they believe the project is aimed at them, that knowledge will not reach it.
The deployments that work tend to be the ones where the same people move to the escalated, judgement-heavy work and are told so plainly at the start.
The underlying point
Voice AI at this maturity is genuinely useful for narrow, repetitive, high-volume work — renewals, qualification, routine answers. That is a real and substantial category, and vendors describing it in those terms are being straight with buyers.
The organisations that get hurt are the ones that buy the demo rather than the deployment. For a sector like Egypt's outsourcing industry, where the International Telecommunication Union tracks the connectivity that made it possible in the first place, the difference between those two purchases is a strategic one.
How to structure the pilot itself
The pilot design matters as much as the vendor choice, and most pilots are structured to succeed rather than to inform.
Pick your worst call type, not your best. A pilot on the cleanest, most scripted queue will pass and tell you nothing. Run it on something genuinely messy, because that is where the deployment will actually live.
Define failure in advance, in writing. Agree the containment rate, the escalation rate and the customer-satisfaction floor before the pilot starts. Metrics chosen afterwards are chosen to flatter whatever happened.
Run it against a human control group. The same queue, over the same period, handled conventionally. Without a control you are comparing the system against your memory of how things used to be, which is not evidence.
Listen to fifty calls yourself. Not a highlight reel, but a random sample including the escalations. Managers who skip this step routinely discover months later that the metrics were technically true and operationally misleading.
The question behind all seven
Every question above is a version of one underlying question: does this vendor understand my operation, or do they only understand their product?
Vendors who have run real deployments talk fluently about failure modes, escalation design and change management, because those are what consumed their time. Vendors who have run demos talk about capability and roadmap. The difference is audible within twenty minutes of a first meeting, and it predicts the outcome better than any specification sheet.
