What a Voice AI Agent Costs to Run in Kenya

A voice AI agent on a Kenyan phone number costs between about $0.03 and $0.08 per minute all-in, and the biggest variable is not the AI. It is the phone call itself. Choose the wrong carrier for a +254 line and the telephony leg alone can cost twenty times more than it should. This post prices every layer of the stack with current published rates and works through a realistic monthly budget for a Nairobi support line.

The three layers of every voice agent bill

Every voice agent invoice has three parts, and vendors only ever quote you one of them at a time:

• Telephony. The phone number, the per-minute charge for the call leg, and any media-streaming fee to get the audio to your agent.
• The AI pipeline. Speech-to-text to transcribe the caller, an LLM to decide what to say, and text-to-speech to speak back. All three meter by the minute or by the character.
• Orchestration. The real-time plumbing that connects the call to the models: LiveKit Cloud, a platform like Twilio ConversationRelay, or something you host yourself.

The mistake most budgets make is pricing layer two carefully (it is the exciting part) and treating layer one as an afterthought. In Kenya that gets the economics exactly backwards, because layer one is where the prices differ by an order of magnitude.

The telephony layer: what a Kenyan number really costs

There are two realistic ways to put a Kenyan phone number on a voice agent: a local aggregator with direct carrier interconnects, or a global CPaaS provider that terminates calls to Kenya at international rates.

Africa's Talking: local interconnect, local prices

Africa's Talking publishes its Kenya voice rates in shillings, exclusive of 16% VAT:

• Regular number (Safaricom or Airtel): KES 5,000 setup, KES 2,000 per month
• Incoming calls: KES 0.50 per minute
• Outgoing calls: KES 2.50 per minute
• Toll-free 800 series (Safaricom only): KES 50,000 deposit, KES 20,000 setup, KES 15,000 per month, incoming at KES 2.50 per minute
• Concurrency: one call per second included; each additional call per second costs KES 3,000 per month

At roughly KES 129 to the US dollar (the mid-market rate in September 2026), that is about $0.004 per minute inbound and $0.019 per minute outbound. Africa's Talking also offers bring-your-own-carrier and SIP trunking, which matters if you already have a Safaricom line you want to keep.

Twilio: a Kenyan number at international rates

Twilio sells Kenyan local numbers for $16 per month, and inbound calls cost $0.0239 per minute. That is workable, though still about six times the Africa's Talking inbound rate.

Outbound is where it hurts. Twilio prices calls from its platform to Kenya as international termination:

• To a Kenyan landline: $0.3779 per minute (about KES 49)
• To a Kenyan mobile: $0.3933 per minute (about KES 51)

Compare like with like: KES 2.50 versus roughly KES 51 per outbound minute. That is a twenty-fold gap, and it is the single most important number in this post. If your use case involves calling customers (payment reminders, collections, appointment confirmations), a US-centric platform will cost you more in telephony than the entire AI pipeline combined.

Twilio's SIP interface at $0.004 per minute is genuinely cheap, but it only helps if you already have a carrier relationship to bring. If you do, Africa's Talking's BYOC option competes for the same setup.

Telnyx and Plivo are excellent for US numbers (Telnyx lists US outbound voice at $0.007 per minute, Plivo at $0.0115), but neither changes the Kenyan picture: their Kenya termination is international-rate too.

Latency: the cost that hides off the invoice

Telephony choice also affects how the agent feels. A caller starts talking over any agent that pauses too long, and every extra network hop adds delay. The major model APIs are hosted mostly in the US and Europe, and the nearest large cloud regions to Nairobi are in South Africa. Keeping the media path short (a carrier with local interconnect, an agent hosted in the closest region you can get, streaming rather than batch transcription) matters as much to perceived quality as the choice of voice model. Nobody publishes a reliable per-route latency figure, so treat any vendor latency claim you cannot test yourself with suspicion, and test with real Safaricom and Airtel calls before committing.

The AI layer: transcription, the model, and the voice

This layer is now cheap enough that it rarely breaks a budget. Current published rates:

• Speech-to-text: Deepgram charges $0.0043 per minute for streaming Nova-3 Monolingual and $0.0052 for Nova-3 Multilingual on pay-as-you-go. If your callers switch between English and Swahili, budget for the multilingual tier and test accuracy on real call audio first.
• The LLM: this is the smallest line item. LiveKit's per-minute inference list puts GPT-4o mini at $0.0006 per minute and GPT-5 nano at $0.0002; even a strong open model like Kimi K2.6 runs $0.0035 per minute. If you prefer to pay per token, Claude Sonnet 5 is $2 per million input tokens and $10 per million output tokens direct from Anthropic, which works out to well under a cent per minute for typical voice turns.
• Text-to-speech: the widest spread. Deepgram Aura-2 costs $0.018 per minute through LiveKit; Cartesia Sonic 3 costs $0.030; ElevenLabs advertises low-latency TTS as low as 5 cents per minute on its Business plan. Remember the phone network compresses everything to narrowband audio: a premium voice buys less over a phone line than it does on a website demo.
• Bundled option: Deepgram's Voice Agent API wraps STT, LLM and TTS orchestration at $0.075 per minute, and Twilio's ConversationRelay is $0.070 per minute before telephony. Convenient, but you pay for the convenience.

Worked example: 20,000 minutes a month on a Nairobi support line

Take a SACCO or microfinance support line handling about 6,700 inbound calls a month, averaging three minutes: 20,000 minutes. All figures use the published rates above, dollars, VAT excluded unless noted.

Telephony, Africa's Talking:

• Inbound minutes: 20,000 x KES 0.50 = KES 10,000, about $77
• Number rental: KES 2,000 plus VAT, about $18
• One-off setup: KES 5,000 plus VAT, about $45
• Monthly total: roughly $95

Telephony, Twilio (for comparison):

• Inbound minutes: 20,000 x $0.0239 = $478
• Number: $16, plus Media Streams at $0.0044 per minute = $88
• Monthly total: roughly $582

AI pipeline, budget stack (Deepgram Nova-3 Monolingual + GPT-4o mini + Aura-2):

• $0.0043 + $0.0006 + $0.0180 = $0.0229 per minute
• Monthly total: roughly $458

AI pipeline, premium stack (Nova-3 Multilingual + Kimi K2.6 + Cartesia Sonic 3):

• $0.0058 + $0.0035 + $0.030 = $0.0393 per minute
• Monthly total: roughly $786

Orchestration: LiveKit Cloud bills agent sessions at $0.01 per minute, so about $200 for 20,000 minutes (the $50 Ship plan covers the first 5,000). Self-hosting the open-source LiveKit stack removes this line and replaces it with engineering time and a server.

Putting it together:

• Lean build (AT + budget AI + LiveKit): about $750 per month, or $0.037 per minute
• Premium build (AT + premium AI + LiveKit): about $1,080 per month, or $0.054 per minute
• Same premium build on Twilio telephony: about $1,570 per month, and far worse if any of those minutes are outbound

Now flip the scenario to outbound. A collections or loan-reminder campaign running 20,000 outbound minutes pays KES 50,000 (about $386) on Africa's Talking. The same minutes on Twilio cost $7,866. That difference alone would fund the entire AI stack several times over.

For the build-side costs (development, prompt design, integration, testing), our voice AI implementation cost breakdown covers the project budget separately from these running costs, and the AI agent development cost breakdown prices the engineering work if you are scoping a custom build.

Recording, consent, and the Data Protection Act

If you record calls, and for quality assurance you should, those recordings are personal data. Kenya's Data Protection Act 2019 applies to how you collect, store and retain them, and the Office of the Data Protection Commissioner is the active regulator. In practice: announce the recording at the start of the call, tell callers what the agent is, store recordings somewhere you control with a defined retention period, and be ready to delete a specific customer's data on request. A voice agent that cannot answer "show me everything you hold on this caller" is a compliance problem waiting for a complaint. This is general information, not legal advice; get your counsel to review the call flow before launch.

A decision checklist before you buy anything

• Inbound or outbound dominant? Inbound tolerates more carrier choices. Outbound at any volume effectively forces a local aggregator or a direct carrier relationship.
• Do you need toll-free? The 800 series through Africa's Talking is Safaricom-only and carries a KES 50,000 deposit plus KES 15,000 monthly. Only pay it if callers genuinely will not spend their own airtime.
• How many simultaneous calls? One call per second is included at Africa's Talking; each extra costs KES 3,000 a month. A campaign dialler needs this budgeted upfront.
• Which languages? If callers mix English and Swahili, you need multilingual STT and you need to test it on real phone audio, not clean microphone samples.
• Cloud or self-hosted orchestration? LiveKit Cloud at $0.01 per minute is fast to ship; self-hosting cuts the fee but you own the uptime.
• What is your recording and retention story? Decide before launch, not after the first complaint.

When a voice agent is the wrong choice

Honesty helps more than enthusiasm here. Skip the voice agent when:

• Your volumes are tiny. Under roughly 2,000 minutes a month, the setup fees and engineering time swamp any saving over a part-time receptionist.
• The calls are genuinely complex. Disputed loan restructurings, distressed customers, multi-party negotiations: these need a human. The agent should triage and hand off, not pretend.
• Your customers mostly prefer chat. If your support already runs on WhatsApp, a chat agent reaches people where they are for a fraction of the cost, and our guide to AI voice assistants for small business covers when each channel wins.
• You cannot staff the escalation path. An agent that handles 80% of calls brilliantly and strands the other 20% in a queue nobody answers is worse than no agent.

Common mistakes

• Pricing the AI and forgetting the phone call. As the worked example shows, telephony ranges from $95 to $7,866 for identical minutes depending on carrier and direction.
• Buying premium TTS for a phone line. Narrowband call audio flattens the difference between a $0.018 and a $0.050 voice. Test blind before upgrading.
• Ignoring VAT and one-off fees. Africa's Talking prices exclude 16% VAT, and setup fees are real money at small scale.
• Assuming concurrency is free. One call per second is roughly 25 simultaneous three-minute calls. A marketing blast needs paid CPS.
• Skipping the Swahili test. Multilingual STT pricing only matters if accuracy holds on compressed, noisy phone audio. Run fifty real calls through it first.
• No cost ceiling. Put a hard monthly cap and alerting on every API key from day one. A stuck retry loop on an outbound campaign can burn a month of budget in a weekend.

Frequently asked questions

Can I get a Kenyan +254 number from Twilio?

Yes. Twilio lists Kenyan local numbers at $16 per month with inbound calls at $0.0239 per minute. The catch is outbound: calls to Kenyan mobiles terminate at $0.3933 per minute, about twenty times the Africa's Talking local rate, so Twilio numbers suit inbound-heavy use cases only.

How much does a voice agent cost per minute in Kenya?

All-in, between about $0.03 and $0.08 per minute depending on the stack. A lean build on local telephony runs near $0.037; a premium build near $0.054; bundled platforms like Deepgram's Voice Agent API start at $0.075 before the phone leg.

Does speech recognition work for Swahili callers?

Multilingual STT models such as Deepgram's Nova-3 Multilingual exist and are priced only slightly above monolingual tiers, but published language lists do not tell you how a model performs on noisy, narrowband phone audio with code-switching. Budget a testing phase with real call recordings before committing.

Do I need consent to record calls in Kenya?

Call recordings are personal data under the Data Protection Act 2019, and the ODPC is an active regulator. Announce the recording, state the purpose, and keep a retention and deletion process. Have your lawyer review the exact wording and flow.

Can I keep my existing Safaricom number?

Usually yes, via bring-your-own-carrier. Africa's Talking offers BYOC so you can attach an existing line to its APIs without porting, and Twilio's SIP trunking does the same if you already have a carrier relationship.

What is the cheapest way to prototype?

LiveKit's free tier includes 1,000 agent session minutes a month and Deepgram starts with a $200 credit, so a working prototype on a test number costs almost nothing. Spend the prototype budget on call recordings and accuracy testing rather than infrastructure.

Where to go from here

Start with the worked example, swap in your own call volumes and direction split, and the carrier decision usually makes itself. If you would rather have the whole thing (telephony, agent logic, escalation to your team, compliance) designed and run for you, that is exactly what our AI employees service delivers: an always-on agent on your phone lines, your data, your rules. For banks and insurers that need the same capability behind their own firewall with governance documentation, see our enterprise AI practice.