How to Connect Voice AI to Kenyan Phone Numbers: Setup and Costs

Putting a voice AI agent on a Kenyan phone line costs anywhere from about KES 12,000 to over KES 160,000 a month, and the difference comes down to one decision most guides skip: which carrier terminates your calls. The AI platforms everyone compares, Vapi, Retell, LiveKit, assume US-style telephony. In Kenya the phone layer is a separate buying decision with 10x cost swings, and this post prices it properly: local termination through Africa's Talking, Twilio's Kenya rates, how SIP trunking connects either one to your AI platform, and the consent rules under the Data Protection Act 2019.

Every number below comes from a vendor pricing page or the Act itself, checked in October 2026. Prices change, so verify before you budget. At the time of writing one US dollar buys about KES 130.

You are paying for three separate layers, and it helps to keep them apart:

• The phone line: a Kenyan number plus per-minute termination into Safaricom and Airtel. This is Africa's Talking or Twilio.
• The orchestration layer: the platform that answers the call, streams audio, manages interruptions and barge-in. Vapi, Retell, or LiveKit.
• The models: speech-to-text, the LLM, and text-to-speech. Sometimes bundled into the platform price, sometimes passed through at cost.

Most "voice AI pricing" content covers only the second and third layers. In Kenya, the first layer is where budgets go wrong.

Africa's Talking: local numbers and local rates

Africa's Talking (AT) is the Nairobi-based carrier API that most Kenyan developers already know from SMS and USSD. Its voice product handles making and receiving calls, IVR menus, call transfer, recording, conferencing, and SIP. For an AI agent, the point is simple: AT gives you a real Kenyan number and terminates calls into Safaricom and Airtel at local rates.

From the AT pricing page, Kenya voice costs:

• Regular virtual number (Safaricom or Airtel): KES 5,000 setup plus 16% VAT, then KES 2,000 per month plus VAT.
• Incoming calls on that number: KES 0.50 per minute.
• Outgoing calls to Kenyan lines: KES 2.50 per minute.
• Toll-free 800 series (Safaricom only): KES 50,000 deposit, KES 20,000 setup, KES 15,000 per month, and KES 2.50 per minute incoming. Expensive to stand up, but callers pay nothing.
• Calls to SIP endpoints registered with AT: KES 0.25 per minute. Conference rooms: KES 0.50 per minute.

One limit matters for outbound campaigns: calls-per-second (CPS). You get 1 CPS free. Two CPS costs KES 3,000 per month plus VAT, and 5 CPS costs KES 12,000. One CPS means one new call started each second, which is plenty for an agent answering inbound calls and fine for a reminder campaign dialing a few thousand numbers a day. It is not enough for a dialer blasting tens of thousands of contacts in an hour.

AT also sells voice in Uganda, Tanzania, Nigeria, and other African markets at local rates, which matters if your agent serves East Africa rather than Kenya alone. SIP outgoing runs UGX 35 per minute in Uganda, TZS 8.61 in Tanzania, and NGN 3.00 in Nigeria.

The catch: AT is telephony, not an AI platform. Its voice API speaks XML actions and webhooks. You connect it to your AI layer over SIP or through a small bridge service. That is normal plumbing, but it is plumbing you or your implementer has to build and maintain.

Twilio in Kenya: fine inbound, painful outbound

Twilio does operate in Kenya, and its Kenya pricing page tells a lopsided story:

• Receiving calls on a Kenyan local number: $0.0239 per minute plus $16 per month for the number. That is reasonable.
• Calling Kenyan numbers: $0.3779 per minute to landlines and $0.3933 per minute to mobiles. That is international termination pricing, roughly KES 49 to 51 per minute, about 20 times the AT outbound rate.
• SIP interface and bring-your-own-carrier trunking: $0.0040 per minute per leg, plus extras like recording at $0.0025 per minute and answering machine detection at $0.0075 per call.

The pattern is common with global CPaaS providers outside their home markets: competitive for inbound, expensive for outbound into local mobile networks, because they pay the same international termination rates everyone else does and add a margin.

Twilio still makes sense in two situations. First, inbound-heavy workloads where the per-minute gap against AT is small ($0.0239 versus KES 0.50, or about $0.004) and you value Twilio's ecosystem. Second, BYOC: you keep Twilio's tooling and trunk your own Kenyan carrier in behind it at $0.004 per minute. The second option mostly makes sense if you already run Twilio elsewhere and want one control plane.

The AI layer: Vapi, Retell, or LiveKit over SIP

Once you have a number and termination, the AI platform connects over SIP. All three major options support bringing your own carrier, which is the key feature to look for.

Vapi. Vapi's pricing is a $0.05 per minute hosting fee on pay-as-you-go, with model costs passed through at provider rates. SIP trunking to your own telephony is free on the Vapi side, so pairing it with AT costs you only the hosting fee plus models. Vapi's own worked example puts 1,000 minutes at $82 to $129 all-in with Deepgram, OpenAI, and ElevenLabs. The free tier is $5 of credits and 4 concurrent calls.

Retell. Retell's pricing is component-based: voice infrastructure at $0.055 per minute, text-to-speech from $0.015, LLMs from $0.0128 per minute for GPT-4.1 mini up to $0.08 for Claude Sonnet-class models, and telephony at $0.015 per minute, which drops to free if you bring your own SIP trunk. A lean configuration lands near $0.083 per minute; a premium one can reach $0.31. Retell bills per second, including silence and hold time, and gives 20 concurrent calls free.

LiveKit. LiveKit is the open-source option, which makes it the natural fit if data residency matters. The cloud Ship plan at $50 per month includes 5,000 agent session minutes; LiveKit's own calculator estimates a phone call at about $0.067 per minute all-in with typical model choices. You can also self-host the whole thing, which is the route banks and insurers take when recordings cannot leave their infrastructure.

A caveat worth repeating from people who build these for a living: no vendor publishes a true all-in per-minute number, because models, voices, and call length move the real figure by 3x or more. A cost breakdown from Hestur puts the realistic all-in range at $0.07 to $0.33 per minute, notes that GPT-4o-class models cost three to five times their mini versions, and finds managed platforms cheaper below roughly 20,000 minutes a month once you count the engineering time a self-hosted stack consumes. Treat any headline per-minute figure as a starting point for a spreadsheet, not a budget line, as Code and Trust's telephony comparison puts it.

Worked example: 5,000 minutes a month

Take a realistic mid-size deployment: a customer service agent handling 5,000 minutes a month, split 3,000 minutes outbound (payment reminders, delivery updates, mostly to Safaricom mobiles) and 2,000 minutes inbound.

Telephony through Africa's Talking:

• Outbound: 3,000 x KES 2.50 = KES 7,500
• Inbound: 2,000 x KES 0.50 = KES 1,000
• Number rental: KES 2,000
• Subtotal KES 10,500, plus 16% VAT = KES 12,180, about $94

The same traffic through Twilio:

• Outbound to mobiles: 3,000 x $0.3933 = $1,180
• Inbound: 2,000 x $0.0239 = $47.80
• Number: $16
• Total about $1,244, roughly KES 161,000

That is a 13x difference on this call mix, and it widens as outbound share grows. If your agent is inbound-only, the gap nearly disappears and the decision turns on features instead.

Now add the AI layer on top of the AT telephony:

• Vapi: $0.05 x 5,000 = $250 hosting, plus models at roughly $150 to $400 depending on choices. Call it $400 to $650, plus the $94 phone layer.
• Retell: lean configuration near $0.083 per minute is about $415; premium models push past $1,500. Own-SIP telephony is free, so add the $94.
• LiveKit Cloud Ship: $50 covers the agent minutes, models around $190 at LiveKit's calculator rates, plus $94 telephony. Roughly $330 total, the cheapest managed route.

So a sane all-in budget for 5,000 minutes a month on Kenyan numbers is $450 to $800 per month on a managed platform, plus one-time setup: the KES 5,800 number activation and whatever integration work the SIP bridge needs. For the wider economics, including what the AI side costs at higher volumes, our post on what a voice AI agent costs to run in Kenya goes deeper, and the voice AI implementation cost breakdown covers build budgets.

Consent, recording, and the Data Protection Act

The phone layer has a legal side in Kenya, and it is not optional reading for outbound use cases. The Data Protection Act, 2019 treats phone numbers and call recordings as personal data, and voiceprints fall under its definition of biometric data.

Three provisions matter most for AI calling:

• Section 37: you cannot use personal data for commercial purposes without the data subject's express consent. An AI agent making marketing or sales calls to numbers you collected needs documented consent. Consent under the Act must be express, unequivocal, free, specific, and informed.
• Recordings: call recordings are personal data. Tell callers the call is recorded, store recordings with appropriate safeguards, and do not keep them longer than necessary.
• Section 43: a breach affecting recordings or call logs triggers notification duties to the Data Commissioner, and to affected people where there is real risk of harm.

The Office of the Data Protection Commissioner has been active: the registration regulations took effect in July 2022, many controllers and processors above the thresholds must register with the ODPC, and the office has published sector guidance including a note for the communications sector, as DLA Piper's Kenya data protection summary tracks. Service calls about an existing customer relationship sit on different ground from cold marketing calls, but the line is fact-specific. Get a Kenyan lawyer to review your call flows before an outbound campaign, not after a complaint.

Latency: keep the agent close to Nairobi

Termination choice also shapes call quality. Every network leg between the caller, the carrier, your SIP trunk, the AI platform, and the model APIs adds delay, and phone callers notice lag faster than web users do. If your agent infrastructure sits in a US region while your callers are on Safaricom in Kisumu, you are adding hundreds of milliseconds before the models even start thinking. Host the agent stack in the nearest sensible region, keep the SIP path short, and budget the milliseconds deliberately. Our voice AI latency breakdown maps where the time goes and what a realistic response-time budget looks like.

When Kenyan telephony is the wrong choice

Skip the local phone layer entirely when:

• Your customers are not in Kenya. If the audience is American, a US number on Twilio or the platform's bundled telephony is cheaper and simpler.
• The product is web-first or WhatsApp-first. A web voice widget or a WhatsApp bot avoids PSTN costs altogether. Plenty of Kenyan customers prefer WhatsApp anyway.
• Volume is trivial. Under a few hundred minutes a month, platform-bundled telephony at $0.015 per minute beats the overhead of a carrier contract and a SIP bridge.
• You need toll-free immediately. AT's toll-free setup wants a KES 50,000 deposit and Safaricom-only reach. If toll-free is core to the use case, price it early.

Common mistakes

• Routing outbound calls through Twilio by default. It is the single most expensive error on this stack: $0.39 per minute where KES 2.50 exists.
• Ignoring CPS limits. One free CPS is fine until your reminder campaign needs to finish 10,000 calls before lunch. Price the extra CPS into the plan.
• Forgetting VAT. AT prices exclude 16% VAT. Twilio bills in dollars, which also means FX exposure on a KES budget.
• Recording without disclosure. "This call may be recorded" is not a nice-to-have; recordings are personal data under the DPA.
• Paying for premium voices on transactional calls. A payment reminder does not need a cloned studio voice. Model and voice choices move cost 3x, so match them to the call.
• Quoting a per-minute price you have not modeled. Vendor headline rates exclude half the stack. Build the spreadsheet first.

Frequently asked questions

Can I use a regular Safaricom SIM as the number for my AI agent? Not sensibly. SIM-box setups are fragile, hard to scale, and sit badly with carrier terms. A virtual number from Africa's Talking costs KES 2,000 a month and comes with an API, SIP, and proper call routing.

Can Twilio give me a Kenyan number? Yes for receiving calls: Twilio lists Kenyan local numbers at $16 a month plus $0.0239 per minute inbound. The problem is the other direction, where calling Kenyan mobiles costs $0.3933 per minute, so Twilio fits inbound workloads, not outbound ones.

Does Africa's Talking work outside Kenya? Yes. AT sells voice in several African markets at local rates, including Uganda, Tanzania, and Nigeria. Rates and available number types differ by country, so check the pricing page for each market you serve.

Do I need consent before an AI agent calls customers in Kenya? For anything commercial, yes. Section 37 of the Data Protection Act requires express consent before personal data is used for commercial purposes. Calls about an existing service relationship are a different analysis, but have a lawyer confirm where your use case falls.

What is the cheapest way to test a voice agent on a Kenyan number? Combine free tiers. LiveKit's Build plan includes 1,000 agent minutes and a free phone number with no card, and Retell and Vapi both offer starter credits. Add an AT number at KES 2,000 a month when you are ready for real Kenyan callers.

Does the phone line affect which languages the agent can speak? No. The carrier carries audio; language ability comes from your speech-to-text, LLM, and text-to-speech choices. Swahili and Sheng performance is a model selection question, and it is worth testing with real Kenyan audio before launch.

Where to go from here

The phone layer is a solved problem once you pick the right carrier for your call mix: Africa's Talking for anything outbound-heavy in Kenya, Twilio for inbound or global consistency, SIP into Vapi, Retell, or LiveKit for the AI itself. If you would rather have the whole thing, telephony, agent, models, and compliance wiring, delivered as one working system, that is what our AI employees service does, with business automation handling what happens after the call ends: the CRM update, the M-Pesa follow-up, the escalation to a human when the caller needs one.