SIP Trunks for Voice AI Agents in Dubai, UAE

A voice AI agent that works in a demo still has no way to ring. It exists as a SIP endpoint, an address on a network, and a phone number is something a licensed operator assigns to a business. The thing that joins the two is a SIP trunk. This guide covers what a trunk is, how a call moves over it, what UAE law requires, how you get one, and the four ways to connect it to the agent, including where a session border controller belongs.

What a SIP trunk is

A SIP trunk is a connection between your phone system and a licensed operator's network, carried over IP and sold in channels. One channel carries one simultaneous call, inbound or outbound, across every number pointed at the trunk. The numbers themselves, called DIDs, are assigned separately by the operator and can be moved between destinations without touching the trunk.

Channels and numbers get confused constantly. Two hundred numbers can sit on a trunk with twenty channels; the twenty-first simultaneous call is the one the operator rejects, because a trunk has no queue. Operators also cap call attempts per second, and a burst above that cap is rejected even when channels are free. Ask for both figures in writing.

How a call moves over a SIP trunk

Two flows run at once. The signalling is SIP, defined in RFC 3261: one side sends INVITE with the audio formats it can speak, the other answers 200 OK, and somebody eventually sends BYE. The audio travels separately as RTP packets, often to a different address and sometimes over a different path. Call quality is a property of the audio path; the operator's uptime dashboard is usually a property of the signalling.

Three settings on that audio path decide what the agent hears.

• Codec. UAE operator trunks deliver G.711 by default, sampled at 8 kHz. G.729 compresses the same band with a model that reconstructs speech, which is cheap on bandwidth and lossy exactly where a speech recogniser reads. G.722 and Opus carry more of the spectrum but only help when the whole path supports them; a call that started on the public phone network arrives band-limited already. On a UAE trunk the practical answer is G.711, with wideband worth negotiating only end to end.
• Keypad digits. Send them as named telephone events under RFC 4733, in their own RTP payload type. Digits mixed into the audio as tones end up inside what the recogniser is transcribing.
• Latency. The trunk adds delay at four points: packetisation, the jitter buffer, transcoding, and distance. Platform latency figures include none of them. A media path that hairpins through an edge on another continent pays that round trip on every turn, so keep the media path short and the codec conversions to zero where you can.

The UAE rules for SIP trunks and AI calling

Carrying calls to and from the public phone network in the UAE is a licensed activity. The TDRA licenses the operators, and enterprise voice, including SIP trunks and the numbers riding on them, is contracted between a licensed operator and a business holding a trade licence. That chain shapes every flow below. xAIa connects to telephony and does not supply it: your operator keeps the trunk, the numbers and the regulatory relationship, and the agent registers against that infrastructure like any other endpoint.

A business SIM sits outside the trunk conversation. Operator enterprise terms forbid automated calling except where the operator has approved it in advance, so it is a per-client, per-operator approval, obtained by you for your defined use case and confirmed in writing before anything is built. Personal SIM cards are never used for automated calling.

The calling rulebook attaches to the calling itself, whatever technology places the call. Cabinet Resolution No. 56 of 2024 expressly permits automated systems for marketing under Article 5(6), then holds them to everything else in the resolution: prior approval to practise telemarketing under 4(1), no calls to Do Not Call Registry numbers under 4(5), recording disclosed at the start of the call under 4(7), company and purpose identified under 4(11), and a calling window of 09:00 to 18:00 under 5(3). Cabinet Resolution No. 57 of 2024 prices the failures, with automated calling in violation carrying AED 10,000, rising to 25,000 and 50,000 on repeat, and operating without prior approval starting at AED 75,000. The clause-by-clause version is in the UAE AI call center guide.

Article 5(3) is also a capacity question. A full day of outbound has to fit inside nine hours, which raises peak concurrency, so size the trunk against the peak that window forces.

How to get a SIP trunk in the UAE

A SIP trunk is ordered from a licensed UAE operator on a business account, against your trade licence. The operator assigns the DIDs and provisions the trunk on its own clock, so start this in week one. Before you sign, get these in writing:

1. Channel count and the calls-per-second cap.
2. The codec the trunk delivers, and whether G.711 is guaranteed end to end.
3. The static IP addresses the operator signals from, and the UDP port range its media uses, so your firewall can be opened in both directions.
4. Whether the operator requires a fixed public IP or a session border controller on your side to terminate the trunk. Many enterprise trunks do.
5. Whether TLS for signalling and SRTP for media are offered.
6. Which numbers the trunk is entitled to present as caller ID. Operators police this, and a mismatch is either rejected at the edge or silently overwritten with the trunk's default number.
7. One DID carved out for the pilot, so a single number points at the agent while everything else stays where it is.

Connecting the trunk to the voice AI agent

The agent side is a SIP endpoint. On Vapi, the platform xAIa builds on as a listed service partner, an assistant can be reached directly at a SIP URI, and your own operator trunk can be registered as a SIP trunk credential: the operator's gateway address plus authentication, either username and password or IP-based. Vapi's own guidance is to prefer credential authentication, because IP-based authentication only works reliably when the operator gives you a dedicated termination address. Signalling runs on port 5060 for UDP and TCP and 5061 for TLS, media on UDP ports 40000 to 60000, and SRTP is available for encrypted media; the addresses to allow-list per region are in Vapi's SIP networking guide.

There are four ways to put the trunk in front of that endpoint.

Flow 1: operator trunk straight to the platform

The operator points the trunk, or a block of DIDs on it, at the platform's SIP endpoint. The platform holds the trunk credential, your firewall allows the platform's signalling addresses and media range, and inbound calls to the DID reach the agent while outbound calls leave through the same trunk presenting a number you hold. This is the shortest path and the right one when nothing else is already on the line.

Flow 2: through the PBX or contact centre you already run

If your voice lives inside a cloud contact centre, a hosted PBX or an on-premise IP PBX, the operator trunk is settled and the agent joins the platform you have: as a registered extension, a queue destination or a SIP route in the dial plan, depending on what your system calls it. The PBX keeps the IVR, the numbers, the recording and the reporting, and hands selected calls to the agent over a SIP leg. The agent hands calls back the same way, with a blind transfer over SIP REFER that releases the channel, or a bridged warm transfer that keeps two channels open while the handover lasts. Either way the context has to be pushed into the CRM or the transfer target on purpose, because the protocol moves the audio and nothing else.

Flow 3: through a session border controller

An SBC is a purpose-built edge device, physical or virtual, that sits between your network and everyone else's. You need one when the operator will only terminate the trunk on a fixed public IP, when TLS and SRTP are mandated by your security policy or your regulator, when the trunk carries both human seats and the agent and the routing has to be split, or when you want a single hardened SIP endpoint exposed to the outside instead of the PBX itself. The SBC handles NAT traversal, encryption, transcoding where two sides disagree on codec, topology hiding, and the recording and compliance taps that audit teams ask for. For the agent it presents one secured SIP endpoint, and behind it the trunk, the PBX and the routing rules stay yours. Keep the SBC in region: every turn of the conversation crosses it twice.

Flow 4: on-premise

The strictest flow, and for some buyers the only permitted one. The trunk terminates inside your perimeter, the models run on hardware you control, and the audio is understood in the building it arrived in. An SBC still guards the edge; the platform side now lives behind it. Government entities, banks and insurers whose mandate says the conversation must be processed under national jurisdiction land here, and the full argument is in on-premise voice AI for Gulf government and enterprise. The wider choice of route, including international numbers for markets outside the UAE, is in connecting voice AI agents to UAE phone numbers.

What the agent needs from the trunk

Five things, each a question for whoever supplies the trunk.

• Audio clean enough for a recogniser. A person's brain repairs a bad line; a recogniser transcribes it as delivered, packet-loss repairs included.
• Predictable jitter. A steady delay is absorbed once. A wandering one lands on the silences, and silence is what the agent measures to decide the caller has finished.
• A media path open both ways while the agent speaks. Barge-in works only when the agent can hear the caller over its own playback. Echo suppression or half-duplex behaviour anywhere in the path produces an agent that cannot be interrupted.
• Keypad digits as RFC 4733 events, outside the audio.
• A transfer path that carries the context, into the CRM, the screen pop or the receiving agent, before the line moves.

The trunk sets the ceiling on everything the agent can do. Its channel count caps the concurrency, its codec decides what the recogniser hears, its jitter buffer sits inside every response, and its transfer path decides whether a human inherits the context or only the audio. Across more than 1,000,000 production calls, every trunk problem we traced landed on one of those four, and each has an answer the operator can give in writing before the first call is placed.

Team xAIa

Where xAIa fits

xAIa deploys and runs voice AI employees on Vapi for enterprises across the Gulf and is listed in Vapi's partner directory as a service partner. Our deployments have handled more than 1,000,000 production calls and run at over 600,000 calls a month across 63 agents. Telephony is mapped in week one of every deployment: the trunk contract, the channel and setup-rate caps, the codec, the demarcation point and the caller ID entitlement, before anything is switched on. Typical deployments run 30 days from scoping to production, with telephony provisioning the item most often on the critical path, which is why the operator order goes in first.

Frequently asked questions

What is a SIP trunk?

A SIP trunk is a connection between your phone system and a licensed operator's network, carried over IP and sold in channels, where one channel carries one simultaneous call. The phone numbers, called DIDs, are assigned separately by the operator and pointed at the trunk.

Does xAIa provide SIP trunks or UAE phone numbers?

No. In the UAE, SIP trunks and numbers come from licensed operators, contracted to your business against your trade licence. xAIa connects the voice AI agent to the trunk, PBX or SBC you hold, and works with telephony partners to get you there.

Do I need a session border controller to connect a SIP trunk to a voice AI agent?

Only in specific cases: the operator requires a fixed public IP to terminate the trunk, your policy mandates TLS and SRTP, one trunk carries both human seats and the agent, or you want a single hardened SIP endpoint exposed instead of the PBX. A greenfield agent on a new DID can point the operator trunk straight at the platform without one.

How does a Vapi agent connect to a UAE SIP trunk?

Your operator trunk is registered on the platform as a SIP trunk credential with its gateway address and authentication, your firewall allows the platform's signalling addresses and media port range in both directions, and a DID on the trunk is pointed at the assistant. Vapi recommends username and password authentication over IP-based authentication unless the operator provides a dedicated termination address.

Which codec should a UAE trunk deliver for a voice AI agent?

G.711, and confirm it in writing. G.729 discards the part of the spectrum a speech recogniser needs most. Wideband codecs such as G.722 or Opus help only when the whole call path supports them end to end.

Can I use a SIM card instead of a SIP trunk for AI calls in the UAE?

Yes, with a business SIM card, once the use is approved in advance by your operator for your specific use case and confirmed in writing. Personal SIM cards are never used for automated calling. For any real call volume, a SIP trunk is the right instrument because it is sized in channels and carries your caller ID entitlement.

---

Speak to us and we will map your trunk, your peak hour and your codec before anything is switched on.