IVR and AI Voice Agents in Arabic and English

This guide is for customer service directors, contact centre managers and CX technology leads who want to automate phone calls in Arabic and English without frustrating callers. It compares classic IVR with conversational AI voice agents, shows how to design call flows and handovers, and explains how to measure whether automation is actually helping.

What is the difference between IVR and an AI voice agent?

A classic IVR (interactive voice response) plays recorded or synthesised prompts and routes callers by key press or a few recognised words. An AI voice agent understands free speech, keeps track of the conversation and can complete tasks through your systems. They solve different problems, and most contact centres end up using both.

Where classic IVR still wins

A key-press menu is predictable, quick to change and works on poor lines. It is the right choice for a language selection, a short routing choice ("for an existing booking, press 1"), confirmations ("press 1 to confirm your appointment") and emergency or after-hours messages. Standards such as W3C VoiceXML and SSML underpin a lot of this tooling, and the design principles have not changed: few options, most common first, and a way to reach a person.

Where AI voice agents add value

AI voice agents earn their place when callers describe their need in their own words, when the menu would otherwise be deep, or when the call ends in a transaction your systems can complete. Examples are "Where is my order?", "I want to move my appointment to Thursday", or "Why was I charged twice?". The agent identifies the intent, collects the details, calls your systems and either resolves the request or hands over with context.

How do you design a call flow that callers do not hate?

Design the call around the caller's most common reasons for calling, not your organisation chart. Start from data: recent call reasons, transcripts or agent notes, and the share of calls for each reason.

Call flow design procedure

1. List the top call reasons and group them into intents, for example order status, change booking, billing question, complaint, new sale.
2. Classify each intent as automate fully, automate partly (collect details, then hand over) or route straight to a human.
3. Write the greeting. Brand, language choice and an open question: "Welcome to Najm Clinics. You're speaking with the virtual assistant. How can I help you today?" Add the recording notice here.
4. Design each automated intent as a short path: identify the caller, confirm the request, act, confirm the result, offer more help.
5. Define the exits: what triggers a handover, where it goes and what data travels with it.
6. Plan for failure: no input, unrecognised input, system errors, out-of-hours. Each must end somewhere useful.
7. Write prompts for the ear: short sentences, one question at a time, key information last ("Your appointment is on Thursday at 4 pm").
8. Prototype and listen. Read the flow aloud with colleagues, then test with real callers before scaling.

Identification and verification

Decide early how callers prove who they are. Caller ID alone is weak because numbers can be spoofed. Common options are a one-time code by SMS or WhatsApp, account details the caller knows, or a call-back to the registered number. Avoid relying on voice biometrics alone for high-risk actions; NIST SP 800-63B, for example, states that biometric comparison based on voice shall not be used for authentication under its guidelines. For OTP channel choices, see our voice OTP guide.

What should you consider for Arabic dialects?

Arabic voice automation must cope with several dialects, accents and frequent switching between Arabic and English. Text written in Modern Standard Arabic (MSA) may look correct but sound formal or stiff when spoken, and callers rarely speak MSA on the phone.

Native-speaker review is not optional. Ask reviewers from each target group to listen to prompts and to judge transcripts, not just to read scripts. Our guide to Arabic customer messaging covers dialect choice and bilingual design for written channels, and most of it applies to voice.

How does intent detection work, and how do you keep it accurate?

Intent detection turns what the caller said into a request your system can act on, such as "change booking" with a date and a branch. In an AI voice agent, speech recognition produces a transcript, a language model interprets it and extracts details, and the agent decides the next step.

Practical rules for accuracy:

• Confirm before acting. "You'd like to move your appointment from Monday to Thursday. Is that right?" catches most recognition errors before they become problems.
• Ask one clarifying question, not three. If the intent is still unclear after one clarification, hand over.
• Use confidence thresholds. Let the agent proceed on high confidence, confirm on medium and clarify or hand over on low.
• Restrict actions with tools. The agent should change bookings only through defined system actions with validation, never by free-form instructions.
• Ground answers in approved content. Answers about fees, policies or medical and legal matters should come from your knowledge base, not from the model's general knowledge.
• Review transcripts weekly. Look for misheard intents, new reasons for calling and phrases the agent does not handle.

How should handover to a human work?

A good handover moves the caller to the right person quickly and passes everything the agent learned, so the caller never repeats themselves. A bad one drops the caller into a general queue with no context.

Handover triggers

• The caller asks for a person, in any wording or language.
• The intent is on your "always human" list.
• Confidence stays low after one clarification, or the caller repeats the same request.
• The caller sounds distressed or angry, or mentions vulnerability, fraud or an emergency.
• A system needed to complete the task is unavailable.

What context to pass

The human agent should see this on screen as the call connects. If the call must wait in a queue, say so honestly and offer a call-back. Our guide to chatbot to human handoff explains the same principles for messaging, and the logic carries over to voice.

How should call recording, transcription and consent be handled?

Recording and transcription are central to improving AI voice agents, but they involve personal data and sometimes sensitive data. Treat them as a privacy decision, not a default setting.

• Give notice at the start of the call that it may be recorded and transcribed, and why. Offer an alternative for callers who object where your legal basis requires it.
• Disclose the AI. Tell callers they are speaking to a virtual assistant. The EU AI Act includes transparency obligations for AI systems that interact with people, and disclosure is good practice in every market.
• Choose a lawful basis for recording, for analytics and for model training separately. Under the GDPR, and in Saudi Arabia under the Personal Data Protection Law overseen by SDAIA, purposes and bases need to be defined; training models on recordings may need more than quality monitoring does. The UK ICO's guidance on AI and data protection is a useful practical reference.
• Minimise and redact. Mask card numbers, codes and national ID numbers in recordings and transcripts, or pause recording while they are collected.
• Set retention periods and delete on schedule.
• Control access to recordings and transcripts by role, and log access.
• Check local rules in each market you operate, including data residency and sector rules for banking or healthcare. This guide is not legal advice.

How do you measure containment and customer satisfaction?

Measure automation by whether callers' problems were solved, not only by how many calls avoided an agent. Use a small set of metrics, read together and broken down by intent and language.

Set baselines from your current IVR or human-only service before launch, and compare the new flow against them on the same intents. Our guide to omnichannel support KPIs covers how these metrics fit with messaging channels.

When should you keep a human on the call?

Keep humans for calls where judgment, empathy or accountability matter more than speed. Automation should make human agents more available for these calls, not harder to reach.

• Complaints and disputes, especially repeat contacts.
• Vulnerable callers, bereavement and health concerns.
• Fraud reports and security incidents.
• Regulated advice in banking, insurance and healthcare.
• High-value sales and retention conversations.
• Anything the caller asks a person for.

What does a sensible rollout plan look like?

Roll out in stages, starting with narrow, high-volume, low-risk intents and expanding only when the metrics support it.

1. Discovery. Analyse call reasons, volumes by hour and language, and current outcomes.
2. Scope. Pick one or two intents to automate first, such as order status or appointment changes.
3. Design. Write the flows, prompts, handover rules, disclosures and recording notices in Arabic and English.
4. Integrate. Connect the systems the agent needs, with tightly defined actions.
5. Test internally with staff from each dialect group, including noisy lines and code-switching.
6. Pilot on a share of calls or one line, with humans monitoring transcripts daily.
7. Measure against your baseline: containment, repeat calls, transfers and CSAT per intent.
8. Expand to more intents, languages or hours, and keep a weekly review cadence.

Worked example: a Gulf clinic group

"Najm Clinics" (synthetic) runs clinics in Riyadh, Dubai and Doha, with heavy call volumes for appointments in Arabic and English.

• Today: a four-level key-press IVR routes callers to a central team; many callers press 0 immediately.
• New greeting: a short bilingual welcome, AI disclosure and recording notice, then "How can I help you?" in the caller's chosen language.
• Automated intents: book, move and cancel appointments, clinic locations and opening hours. The agent verifies the patient with a code sent by WhatsApp or SMS.
• Always human: symptoms, test results, billing disputes and anything urgent. The agent says it is connecting the caller to the team and passes intent, verified ID, language and a summary.
• Arabic: the agent speaks warm Gulf-leaning Arabic, recognises Egyptian and Levantine callers and handles English clinic and insurer names mid-sentence.
• Measurement: appointment intents are compared with the old IVR for completion, repeat calls within a week, time to reach a human and post-call CSAT.

What mistakes should you avoid?

• Automating everything at once instead of starting with a few intents.
• Hiding the human option, which drives anger and repeat calls.
• Testing only in MSA, then launching to callers who speak dialect.
• Letting the agent improvise on policies, prices or medical topics.
• Handing over without context, so callers repeat everything.
• Reporting containment alone and ignoring repeat calls and CSAT.
• Recording by default without notice, retention limits or redaction.
• Not disclosing the AI, which damages trust when callers realise.
• Launching without a baseline, so improvement cannot be proven.
• Limits: AI voice agents can mishear and can be wrong. Regulations on recording, data and AI disclosure differ by market and are changing; confirm current rules with your legal team.

How OmniChannels.ai helps

OmniChannels.ai provides voice alongside messaging: multilingual IVR, text-to-speech including Arabic, and AI voice agents, including in Arabic, that hand over to a human agent. Through BroadNet AI, agents support English and Arabic understanding, persistent memory and tool use. The voice offer also covers inbound contact centre, dialers, call recording and transcription, and bringing your own SIP trunks. It is the same capability set as DialerBee, BroadNet's call-centre software. Voice is provisioned with our account team for your numbers, languages and call flows. Read the voice developer guide to see how provisioning works.