Building Africa's first LLM-powered customer-service stack for a telco.

Mzansi Mobile's contact centre was a cost centre that everyone agreed needed to change and nobody had successfully changed. Three previous initiatives — a chatbot in 2021, an IVR rebuild in 2022, and an offshore outsourcing experiment in 2023 — had each been declared finished, then quietly rolled back. Customer satisfaction kept falling. The Chief Customer Officer's brief to us was specific: do not bring a product, bring a system that respects the fact that 60% of customer calls happen in isiZulu, Sesotho, or Setswana, and that most agents are paid by the call.

• 14-minute average handle time, 31% first-contact resolution
• Three previous AI/automation initiatives rolled back
• No structured data layer across CRM, billing, and network systems
• Six languages spoken on inbound calls, two of them with negligible LLM training coverage

We bet on retrieval, not generation. The model was always going to be the easy part — what mattered was whether the agent could ground every response in Mzansi's real billing, network, and CRM state in under 800ms. We spent the first three months building a unified retrieval layer over the seven systems that hold a customer's truth, with strict access control, redaction for PII, and a fallback path for every record. Only then did we start tuning the model.

For language coverage, we collected and consented eighteen months of de-identified isiZulu, Sesotho, and Setswana call transcripts, partnered with a Johannesburg university linguistics department to build a benchmark, and fine-tuned an open-weight model on the result. Voice was handled by a streaming ASR pipeline tuned for code-switching, which is how most South African customer calls actually sound.

The platform sits in front of Mzansi's existing voice and WhatsApp channels. An incoming contact is transcribed by a streaming ASR pipeline tuned for code-switching, intent-classified, and routed to a retrieval-augmented agent that can call sixteen internal tools — checking balance, raising a SIM swap, scheduling a network technician — under a policy layer that requires explicit approval for anything that mutates state.

• Retrieval layer unifying seven backend systems with 800ms p95 latency and per-tenant PII redaction
• Fine-tuned open-weight LLM with strong performance on isiZulu, Sesotho, Setswana, English, Afrikaans, and Xhosa
• Streaming ASR pipeline tuned for South African code-switching
• Tool-use policy engine that gates any mutating action behind explicit approval rules
• Human-handoff protocol that hands the agent a full conversation summary and recommended next action

Six months after full rollout, 67% of inbound contacts resolve without a human. The cases that do reach an agent are 71% shorter because the agent arrives with a transcript, summary, and recommended action. Customer satisfaction, measured by post-call survey, rose from 3.1 to 4.4 out of 5. Net contact-centre operating cost fell by R 41M annually, and the team that used to manage the IVR has been redeployed to a quality and policy function.

Retrieval is the product. The model is a feature of the retrieval layer. Every previous attempt at this problem had inverted that hierarchy and bought into a vendor's chatbot before solving the data plumbing. The other lesson — equally important and easier to forget — was about language. Building a credible isiZulu and Sesotho experience required collecting and consenting real conversational data, partnering with linguists, and accepting that the open-weight model would do most of the work once the data was in shape.

• Build the retrieval layer before you choose the model
• Tool-use needs a policy engine — do not let the model mutate anything important directly
• For African-language voice, the ASR is harder than the LLM
• Human handoff is the highest-leverage UX in the entire system — invest in it heavily