EXECUTIVE VIEW
It handles the routine calls, in Swahili and English, and passes the rest to your team with context already attached. Unresolved calls feed back into the system automatically. The answers improve over time.
Teknolojia ya kesho, leo. — Tomorrow's technology, today.
THE STATUS QUO
Average monthly salary per call center agent in Kenya (Glassdoor 2026)
Estimated headcount for a large airline or bank call operation at scale
Estimated annual staffing cost before overhead, management, and training
Most of those calls are asking about flight status, account balances, and opening hours.
Per minute of AI-handled call (ElevenLabs Business plan)
Cost of 6,000 AI-handled minutes. Less than one agent's monthly salary.
From kickoff to your first live AI-handled call.
Your agents focus on complex issues. The AI handles everything else. Both get better over time.
THE REAL NUMBERS
~12 agents at KES 45,000 / month
Monthly salaries: KES 540,000
Annual incl. overhead: KES 8–9M
ElevenLabs: ~KES 10,400 / month (10,000 min at $0.08)
Twilio routing: ~KES 5,200 / month
n8n infrastructure: ~KES 7,800 / month
Total monthly: ~KES 23,400
Potential annual saving: KES 6.2M+
Same availability. 24/7. No sick days. No onboarding cost.
~80–100 agents at KES 55,000 / month
Monthly salaries: KES 4.4–5.5M
Annual incl. overhead: KES 65–75M
ElevenLabs: ~KES 52,000 / month (50,000 min at $0.08)
Twilio routing: ~KES 20,800 / month
n8n infrastructure: ~KES 13,000 / month
Total monthly: ~KES 85,800
Potential annual saving: KES 50M+
70% of calls handled without a human agent. Your team handles what actually needs judgment.
~150 agents at KES 55,000 / month
Monthly salaries: KES 8.25M
Annual incl. overhead: KES 120–140M
ElevenLabs: ~KES 78,000 / month (75,000 min at $0.08)
Twilio routing: ~KES 26,000 / month
n8n infrastructure: ~KES 13,000 / month
Total monthly: ~KES 117,000
Potential annual saving: KES 95M+
Every call logged. Audit trail built in. Compliance-ready by design.
Salary data: Glassdoor Kenya 2026. AI platform costs based on ElevenLabs Business plan ($0.08 per minute), Twilio estimated inbound routing, and n8n community edition self-hosted. Exchange rate: 1 USD = 130 KES. Figures assume 70% AI call deflection. The remaining 30% continues to reach your existing human agents. No headcount reduction required in Phase 1.
BUILT FOR THIS MARKET
THE BIGGER PICTURE
First deployment. 30 days. Custom FAQ, Swahili voice, human escalation wired in. The infrastructure is built.
The second deployment is a config swap and a new knowledge base seed. Same infrastructure. Deployed in days. The work done for client one pays forward.
Every new client pays for speed, not your learning curve. The moat compounds with every deployment. The system improves for everyone.
"The first client funds the build. The second proves the model. The third is pure margin."
Ready to be client one?
Let's TalkVOICE LAYER
How the system handles a call, from the moment it comes in to the moment it resolves or reaches your team.
YOUR CHOICE OF STACK
n8n is a visual workflow automation tool. Think of it as the nervous system of the call, it decides what happens at each step: which knowledge base to query, when to escalate, when to log an outcome. The cloud version means no server management required. Every workflow is visual and auditable, you can see exactly what the system did on any given call.
ElevenLabs is the AI voice layer. It handles both understanding what the caller says (speech recognition) and responding in a natural voice. It supports Swahili, English, and 29 other languages natively. The voice sounds natural, not like a phone tree.
Twilio is the bridge between your existing phone number and the AI system. Callers dial the same number they always have. Twilio routes the call to the AI gateway. No new phone infrastructure required on your end.
In Phase 1, the knowledge base is built from your existing documents, FAQ sheets, call scripts, policy documents. The system reads and structures them. When a caller asks a question, the AI queries this knowledge base to find the right answer.
When the AI cannot resolve a call, it transfers to a human agent, but not blind. The agent receives a summary: what the caller asked, what the AI tried, and why it escalated. The agent never starts from scratch.
Self-hosted n8n runs on your own infrastructure, giving you full data control and no execution limits. Make.com is an alternative if your team prefers a managed cloud environment. Both produce a full visual audit trail of every call workflow.
ElevenLabs is the AI voice layer. It handles both understanding what the caller says (speech recognition) and responding in a natural voice. It supports Swahili, English, and 29 other languages natively. The voice sounds natural, not like a phone tree.
Qdrant is a vector database, it stores knowledge in a way that allows the AI to find semantically similar answers, not just exact keyword matches. RAG stands for Retrieval-Augmented Generation: the AI retrieves relevant knowledge before generating its response, making answers more accurate and grounded in your actual data.
Phase 2 adds automatic ticket creation. If a call escalates to a human agent and still cannot be resolved, n8n triggers a webhook that creates a ticket in Jira or Freshdesk automatically, with the full call context attached.
The intelligence layer runs asynchronously, meaning it analyzes call outcomes in the background, never touching the live call path. It identifies patterns in unresolved calls, drafts knowledge base updates, and queues them for approval. The system improves without manual input.
For clients where data residency and audit compliance matter, n8n self-hosted keeps everything on their own infrastructure. Both options produce a full visual log of every call workflow. And if requirements change, either can be swapped without rebuilding the product.
Voice layer: ElevenLabs, with alternatives available at scale.
ElevenLabs is the only confirmed provider with native Swahili support and a full conversational AI platform, speech recognition, LLM routing, and voice response in one. Microsoft Azure and Google Cloud also support Swahili and cost significantly less per minute, at the tradeoff of voice expressiveness. The architecture treats the voice provider as a swappable node. If call volume justifies the switch, Azure or Google TTS can replace ElevenLabs without rebuilding the product.
DOES IT SCALE?
Every 20 additional calls per hour requires a new agent hire. Each hire adds salary, onboarding, management overhead, and desk space. Costs scale in large discrete jumps.
Cost increases per minute of AI-handled call, with no management overhead. 10,000 calls or 100,000 calls, the team size on your end stays the same.
The gateway detects caller language in the first three seconds of speech and switches without a menu, without a prompt, without a delay. A caller in Nairobi speaks Swahili. A caller from Dubai speaks Arabic. A layover passenger from Tokyo speaks Japanese. The system handles all three in the same queue, simultaneously, with no configuration change per call.
See how it handles a real call.
Let's TalkTHE INTELLIGENCE LAYER
Every unresolved call becomes an instruction. Every failure becomes a fix. And every successful resolution becomes a template, applied automatically the next time a similar call comes in, and to every future client deployment. The system improves in the background, without involvement from your team.
A continuous cycle running in the background. Every call makes the next one better.
UNDER THE HOOD
Deterministic. Auditable. Repeatable. Every call follows a defined, testable flow. If a response is wrong, you can trace exactly which node failed and fix it in minutes. This is your compliance and audit layer.
Non-deterministic. Learning. Improving. Runs async — never in the live call path. Analyzes patterns across hundreds of calls, identifies what the spinal cord is missing, and proposes updates. You stay in control. The system gets smarter.
"The spinal cord works even when the brain is thinking."
Gaps identified from escalation patterns. New entries drafted and queued automatically.
New intent categories discovered from caller phrasing. Workflow branches proposed and tested.
False positives reduced over time. Fewer unnecessary handoffs. Human agents handle harder issues only.
The system improves every week. With or without your involvement.
Let's TalkKNOWLEDGE CORE
The knowledge base is what separates an AI that deflects callers from one that actually resolves their issues. It starts with your existing documents. After that, it updates itself.
INPUTS
Word doc, PDF, spreadsheet, or plain text. The system ingests and structures it.
Existing agent scripts and decision trees become structured intent flows.
Terms, refund policies, booking rules. Ingested once, referenced on every relevant call.
Your team updates the knowledge base via a simple interface. Changes deploy within minutes.
Every client deployment lives in its own isolated namespace. Data never crosses between tenants. Caller data is queried from your systems via API during the call and discarded when it ends. Nothing is retained on the gateway side.
A caller asks a question the system cannot answer confidently. The call escalates. The outcome is logged automatically.
The system reviews the escalation. In this case: a question about baggage policy with no matching entry in the knowledge base.
A new knowledge base entry is drafted automatically and queued for your review. You see exactly what it says before it goes live.
You approve. It deploys. The next caller who asks the same question gets a clear answer, without reaching a human agent.
Your knowledge base starts with what you already have.
Let's TalkIf a call reaches your team and still cannot be resolved, the system creates a ticket automatically, with the full call context attached. Your ops team picks it up with everything they need.
Issue created in your existing project. Priority, labels, and description auto-populated from call context.
Ticket created with full call summary. Agent assigned based on your existing routing rules.
This module is optional and activated per tenant. It adds zero complexity to the base voice gateway deployment.
COMING IN PHASE 3
Clients on Phase 1 are automatically considered for early access.
Nothing falls through the cracks.
Let's Talk