Almost every AI voice agent sounds good in a scripted demo. The differences appear on a real phone line, with a real caller who is in a hurry, switches language halfway through a sentence, or asks something your website doesn't answer. A good buying process is designed to surface those differences before you sign.
This guide groups the questions that matter into eight areas. Use them in conversations with vendors, and insist on seeing the answers on a live call rather than a slide.
1. Does it sound good on an actual phone call?
Browser demos use high-quality audio. Phone calls usually travel as narrowband audio sampled at 8 kHz, which strips out much of the detail speech recognition relies on. Ask to ring the agent from a mobile phone, ideally on speakerphone, and from a noisy place.
- How long is the pause between you finishing and the agent replying? Natural conversation leaves gaps of only a fraction of a second; long silences make callers talk over the agent or hang up.
- Can you interrupt it mid-sentence, and does it stop and listen?
- Does it cope with background noise, or does it react to every sound?
Our explainer on voice AI latency covers where the delay comes from.
2. Does it really speak your callers' languages?
"Supports Arabic" can mean very different things. Many systems handle Modern Standard Arabic well and struggle with Gulf dialects, mixed Arabic and English, or Arabic names and place names. If your callers are in the UAE, test in Emirati Arabic and with code-switching. Ask the vendor which speech models they use, whether they have tested on Gulf dialect speech, and whether a native speaker has reviewed the voice. See Arabic and Emirati dialect in voice AI.
3. Where do its answers come from?
The biggest risk with a fluent AI is a fluent wrong answer. Ask:
- Does the agent answer only from information you approve, or does it fill gaps from general knowledge?
- What happens when it doesn't know? Ask it a question your playbook doesn't cover and listen.
- Can you see and change what it has been told, without a support ticket?
- How are changes tested before they reach callers?
4. What can it do, and who approves it?
An agent that only talks is limited. An agent that can act without limits is a risk. Find out exactly which actions it can take, such as booking, logging a job, sending a message or issuing a refund, and which of those wait for a person. Ask whether the voice model itself can trigger actions or whether actions go through a separate, controlled system. A caller should not be able to talk an agent into something your policies forbid.
5. How does it hand over to people?
Every deployment needs a clean path to a human. Ask whether it can transfer a live call, take a message, book a call-back or raise an urgent alert, and how you are told. Ask how the vendor confirms that a transfer actually connected. "The request was sent" is not the same as "a person answered". Make sure the agent never tells a caller someone will ring back unless that has really been arranged.
6. How does it connect to your phone line?
Ask whether you can keep your existing number, whether the vendor provides numbers in your country, and whether it can take calls from your operator's SIP trunk or your phone system. In the UAE, voice services must run through the licensed operators, e& and du, so ask how the vendor's route fits within that. Our SIP trunking guide explains the options.
7. Data protection and security
- Roles and contract. You will usually be the controller and the vendor a processor. Expect a written data processing agreement covering instructions, security, subprocessors, breach notification and deletion.
- Location and transfers. Where are audio and transcripts processed and stored, and by which subprocessors? What safeguards cover transfers out of the UK or the UAE?
- Training. Is your call data used to train anyone's models?
- Retention and access. How long are recordings and transcripts kept, who on the vendor's side can see them, and can you delete a caller's data on request?
- Separation. On a shared platform, how is your data kept apart from other customers'?
8. Price, contract and support
Pricing models vary: per minute, per call, per seat, or a flat monthly plan with an allowance. Model your real call volumes, including peaks, and ask what happens when you exceed the allowance. Check the minimum term, the notice period, whether set-up is charged, and whether you can export your transcripts and playbook if you leave. Ask who you call when something goes wrong at 9pm on a Thursday, and how quickly they respond.
A simple scorecard
| Area | What good looks like |
|---|---|
| Phone quality | Quick replies, handles interruptions and noise on a real mobile call |
| Languages | Natural in your callers' dialects, copes with switching mid-call |
| Accuracy | Answers from approved information; says when it doesn't know |
| Actions | Clear list of actions; consequential ones wait for approval |
| Hand-offs | Honest, confirmed hand-offs and messages |
| Phone line | Works with your number and your operator, within local rules |
| Data | Written processing terms, known locations, no training on your calls |
| Commercials | Predictable pricing, fair exit, reachable support |
Run a pilot, not a pitch
The best evidence is a short pilot on a slice of your real calls, after hours or on overflow, with success measures agreed in advance. Our step-by-step implementation guide shows how to set one up.
Sources and further reading
Hear Clodia for yourself.
Try a live conversation in your browser, or tell us about your front desk.