AI voice agent for call centres: dialects, delay, handoff and recording

Picture two call centres in Amman that launch voice agents in the same week (an illustrative case). A water utility gives its AI voice agent one job, taking fault reports, and passes any caller who sounds stuck to a person with a short summary. A private hospital group lets its agent answer bookings, test results and billing disputes, with no clear way out. A few weeks later the utility is ready to add a second call type, and the hospital group is switching its agent off. An AI voice agent for call centres works when it takes narrow calls, understands how callers really speak, replies fast enough to feel like a conversation and hands over to people cleanly.

Which calls should an AI voice agent for call centres take first?

We recommend starting with calls that are frequent, short and low risk, where the answer comes from a system you already trust. Leave judgement calls to people until the agent has a record. The trade-off is that the first release looks modest, but every call it handles can be checked.

Pull recent call reasons from your contact-centre system; if the codes are unreliable, tag a sample of calls by hand. That sample becomes your first test set.

How do you make the agent understand Arabic dialects on the phone?

Phone audio is narrower and noisier than an app recording. A speech-to-text engine that shines in a demo can struggle with a caller on a mobile line in a car. Callers in Jordan and the Gulf also mix dialects, use English product names and say numbers differently.

Build a dialect test set from your own calls

• Collect recorded calls you may legally use and remove names, ID numbers and card details.
• Have people transcribe and label each call by dialect and topic.
• Include the hard parts: numbers, dates, street names, English brand names and background noise.
• Score two things per dialect: how many words the engine gets wrong, and how often the agent picks the right call reason.
• Set a pass mark per dialect before you test, so a good average cannot hide a weak dialect.

Our guide to Arabic LLM evaluation explains how to build and score this kind of test set.

Choose how the agent speaks

The agent can reply in simple Modern Standard Arabic close to everyday speech, or in the caller's dialect. We recommend simple MSA for the first release because one voice script is easier to test and sounds neutral to callers of different nationalities. A dialect voice can feel warmer, but each one needs its own script review and tests. Add it only where tests show callers prefer it.

Take an illustrative Riyadh retailer whose callers include Saudis, Egyptian and Levantine residents and South Asian residents who prefer English. A sensible first prompt offers Arabic or English in one sentence, then keeps the call in that language.

How fast does a voice agent need to reply?

Every reply passes through a chain: the phone network, speech-to-text, a check that the caller has finished, the language model, text-to-speech and the network again. On a silent line, callers think the call dropped and start talking again, which confuses the agent.

• Measure the real gap. Time it from the end of the caller's speech to the first sound of the reply, on local mobile networks, not on office Wi-Fi.
• Watch the slowest calls. Track the slowest one in twenty, not only the average; callers remember the slow ones.
• Let callers interrupt. When the caller starts speaking, the agent should stop talking and listen.
• Fill necessary waits. For a slow system lookup, the agent says what it is doing ("I am checking your booking now") instead of going silent.
• Keep components close. Host speech and model services near your telephony provider.

We recommend fixing delay before adding call types. A slow agent on two call types tends to send more callers to people than a fast agent on one.

When and how should the call pass to a person?

Tell callers at the start that they can ask for a person at any time, in any wording or dialect.

Transfer triggers to set from day one

• The caller asks for a person, whether they say "agent", "موظف" or "بدي أحكي مع حدا".
• The agent fails to understand the caller twice in a row.
• The caller sounds upset, or mentions health, safety, fraud or legal action.
• The request falls outside the approved call types.
• An identity check fails or the caller cannot complete it.

Use a warm transfer: the person who picks up sees a short written summary of who is calling, why, what the agent already checked and what is still open. The caller should never repeat the story. Outside staffed hours, offer a callback at a stated time instead of a loop.

Judge the agent by calls resolved, transfers by reason and repeat calls within a week, not by calls that never reached a person. Our guide to Arabic AI chatbots for customer service explains why that number misleads.

What are the call recording and disclosure rules in Jordan, Saudi Arabia and the UAE?

A recorded call, its transcript and the summary passed to staff are personal data in all three countries. The points below come from official pages we read on 19 October 2026. This is general information, not legal advice; your obligations depend on your sector, your data and where the processing runs.

• Jordan. The Personal Data Protection Law No. (24) of 2023 says personal data may not be processed without the person's prior consent, except in cases the law permits (Article 4). Consent must be explicit, documented in writing or electronically, and specific about duration and purpose (Article 5). Before processing, the controller must inform the person in writing or electronically of the data, purpose, period and any processor involved, among other points (Article 9). Ask counsel whether a spoken notice at call start is enough or needs a written or text-message follow-up.
• Saudi Arabia. The Personal Data Protection Law is overseen by SDAIA. The SDAIA data protection knowledge centre says the law also covers processing by entities outside the Kingdom of data about people in the Kingdom, and lists a separate regulation on transferring personal data outside the Kingdom. That matters if your speech or model services run on servers abroad. Article 13 of the law also requires telling callers, when their data is collected, the legal basis, the purpose and whether it will be processed outside the Kingdom.
• UAE. The UAE government's data protection page says Federal Decree-Law No. 45 of 2021 prohibits processing without consent except in specified cases, and cites Article 31 of the Constitution, which protects the confidentiality of communications. Decree-Law 45 (Article 2 of the official Arabic text) does not cover government data, health and banking data that have their own laws, or free zones with their own data protection law, such as DIFC (DIFC Law No. 5 of 2020) and ADGM, so check which law applies to you.

Other rules on recording calls

Data protection law may not be the only rule on recording. Before the pilot, ask counsel whether criminal, telecom or sector rules require consent to record calls, and what banks, telecom operators and health providers must follow on retention and hosting.

Telling callers they are speaking to AI

The European Commission's AI Act page says people should be made aware when they interact with a machine such as a chatbot, under transparency rules that apply from August 2026. It can apply to you if your agent is used with callers in the EU; ask counsel whether it does. We recommend the same disclosure for every caller.

A short opening script

"You are speaking with the automated assistant of [organisation]. This call is recorded to handle your request and to test and improve this assistant. You can say 'agent' at any time to speak with a person."

Behind it, record how long recordings are kept, who can listen, where they are stored and which processors handle them.

A step-by-step launch sequence

1. Pick one or two call types from your own call reasons.
2. Write the opening disclosure and recording notice, and have counsel review them for each country you serve.
3. Build the dialect test set and agree a pass mark per dialect.
4. Connect only the system lookups those call types need, read-only where possible, with limits and an audit log for every action.
5. Set the transfer triggers and the summary format, and test them with the staff who will receive the calls.
6. Measure delay on local mobile networks and fix the slowest step.
7. Pilot on a share of calls during staffed hours and review a sample of recordings every week.
8. Add the next call type only when transfer reasons and repeat calls have stayed steady for several weeks.

How to start

Our Agentic AI & Chat/Voice Agents service (TT2) builds chat and voice agents for web, WhatsApp, Teams and phone, with limits, budgets, audit logs and approvals. It sits in the Transformation Lab; the support and voice agents page describes the use case. If you are not sure where to begin, the AI Exposure Scan is a free 45-minute conversation that ends in a written next step.

Related lab: Transformation Lab ↗