Arabic AI Voice Agents for Banking: Gulf Dialects, Data in the UAE

Arabic AI voice agents for banking answer Gulf customers in Arabic and English, name the bank in the greeting and verify the caller before any account detail. Speech-to-text is set to Arabic or a Gulf code such as Arabic (UAE). UAE banks keep customer data in the UAE, so self-hosting there fits, and outbound calls stay service calls.

Key Takeaways

• Set speech-to-text to Arabic or a Gulf code such as Arabic (UAE).
• Greet in Arabic and English, and name the bank in the first sentence.
• UAE banks hold customer data in the UAE, so run the agent there.

Gulf customers ring in Arabic or English, and the greeting plays before the agent can tell which. We set the speech language, greet in both, pick an Arabic voice, keep UAE data in the UAE and hold outbound calls to service, all in Dograh. It is part of our guide to banking voice agents.

Set the language before anything else

An Arabic bank agent starts with one setting: the language its speech-to-text listens for. In the agent's Model tab, the Language dropdown offers Multilingual (Auto-detect), which follows callers across ten languages: English, Spanish, French, German, Hindi, Russian, Portuguese, Japanese, Italian and Dutch. For Arabic, pick it directly instead.

Dograh's own speech-to-text uses Deepgram's language list, which offers Arabic plus regional codes. In the dropdown they appear as Arabic (UAE), Arabic (Saudi Arabia), Arabic (Qatar), Arabic (Kuwait) and others, 16 regional options in all. Pick the Gulf dialect code that matches the callers on that line, and test it with real Gulf speech.

Many Gulf customers switch between Arabic and English in one sentence, especially for product names and numbers, a habit called code-switching. You have two ways to handle that. Give each language its own phone number, since several numbers can point at different agents through each number's Inbound workflow setting. Or, for callers who mix both, use Speechmatics' Arabic and English bilingual model, which Speechmatics built for Arabic and English in the same stream, on your own key.

Add product names, such as your card and account names in both scripts, to the Dictionary in the agent's Settings. Those words reach the speech-to-text provider, and whether they help depends on that provider. Multilingual Voice Agents covers detection and language switching across every language Dograh serves.

Greet in both languages and name the bank

The greeting plays before the agent can hear which language the caller speaks, so write it in both. In the UAE, the Central Bank's Consumer Protection Standards require staff to identify themselves and the institution they represent, and all disclosure information must be available in Arabic and English. A voice agent follows the same rules, so the bank's name comes first.

In Dograh, that line is the Greeting Text on the Start Call node, spoken word for word as the call connects. For example: "أهلاً بكم في بنك المثال. Welcome to Example Bank. How can I help you?" If you want to tell callers they're speaking with an automated assistant, or your location requires it by law, add it to the same greeting. Callers can cut into a greeting after two words by default, so switch off Allow Greeting Interruption on the node when the greeting carries a notice that must play in full.

Then the agent replies in the caller's language and checks who is calling before it shares anything. Caller ID only points to a record. It never proves the caller owns it. Before any balance or card detail, the agent asks for an approved check, such as date of birth plus the last digits of the account number, and your core system answers pass or fail. AI voice agents for KYC in banking covers the check in detail.

Open Source Alternative to Vapi / Retell

Self-hosted voice agent platform — no per-minute fees

dograh-hq/dograh

Star on GitHub

The Arabic voice on the line

The voice matters as much as the hearing, because Gulf customers notice a voice that stumbles over their names. Dograh's own voices include Arabic options, and its speech-to-text handles Arabic. On your own key, ElevenLabs lists Arabic for Saudi Arabia and the UAE in its multilingual voices.

Test each voice with the words your callers hear most: customer names, branch names, amounts and dates. Have your core system return amounts and dates written the way they should be spoken, and listen on a test call before you pick.

Some callers need a person, such as anyone reporting fraud or a lost card. The Transfer Call tool runs on five telephony providers: Twilio, Telnyx, Plivo, Cloudonix and Asterisk ARI. The branch or fraud team's line rings first, and the customer is connected once a colleague answers. Twilio lines can add a short spoken introduction that both of them hear, in Arabic if your voice speaks it. Before the handover, the agent logs the reason in your case system, so the colleague starts with it on screen. QA summaries come out in English by default, so an Arabic-speaking reviewer reads the transcripts of flagged calls.

Keep UAE customer data in the UAE

For a UAE bank, a data localisation rule decides where the call data sits. The CBUAE Consumer Protection Standards require licensed institutions to hold and store all consumer and transaction data within the UAE, with a secure backup in a separate location, and to keep records for at least five years. Call recordings and transcripts are consumer data like any other.

On top of that sits the UAE's data protection law, Federal Decree-Law No. 45 of 2021, in force from 2 January 2022, with its own conditions for sending personal data abroad. The financial free zones, such as the DIFC, have their own data protection laws.

Two layers handle every call. One is Dograh, running the call flow, tools and telephony, and the other is the speech and language models your bank chooses for Arabic. Self-hosted in a UAE data centre, the platform, recordings and transcripts stay there. Arabic speech-to-text can run there too, with Whisper large-v3 through the Local Models (Speaches) provider, and so can an open-weight language model. For the voice, use Dograh's Arabic voices or another hosted Arabic voice. A hosted voice receives the words the agent speaks, balances included, so check where that provider processes them.

Self-hosted, the recording switch applies server-wide, and your own storage schedule and deletion process keep records for the five years. Anonymous usage telemetry starts switched on, and the UAE team turns it off with ENABLE_TELEMETRY=false and by removing POSTHOG_API_KEY. A bank hosting Dograh itself pays no per-minute platform fee, only its servers, model keys and phone minutes, and Dograh Cloud's platform fee is 1¢-a-minute.

Saudi banks work under the Saudi Central Bank (SAMA) and the Personal Data Protection Law, issued by Royal Decree in September 2021 and since amended, which SDAIA oversees. Check those rules for your Saudi lines separately.

Outbound calls stay service calls

In the UAE, an outbound bank call should tell the customer something about their own account, never sell. The CBUAE Standards bar unsolicited calls to consumers for marketing financial products, require express consent for promotional contact, and prohibit marketing loans and financing to individual consumers through direct telephone contact.

So outbound calls cover a payment due date, a card ready for collection or a missing document. A loan or card offer has no place on them. Collection calls have their own rule: the Standards require them to be recorded and the records kept for five years after the debt is settled or written off, so collections teams keep recording on.

Campaign calling hours are set by weekday in one time zone per campaign, so set the calling days to match each country's working week. Treat this post's points on the CBUAE Standards and data protection as general information, and confirm your bank's position with counsel.

Where to start

Start inbound with questions that need no account, such as branch hours and fees, on one Arabic number. Add balance questions behind the check next, then outbound service calls. The same Arabic settings carry over to other Gulf sectors, as our post on AI voice agents for real estate in the UAE shows.

Before launch, call the agent from an ordinary phone. Phone audio usually arrives at 8 kHz against 16 kHz in a browser test, and Arabic names are often the first words to suffer.