WhatsApp Self-Service at an African Mobile Network Operator — 44% Contact Reduction, English + Swahili
ChatNext-powered WhatsApp self-service for SIM, billing and bundle management — bilingual, integrated with the BSS stack, deflecting 44% of all inbound.
The challenge
The operator — a leading mobile network operator in East Africa, with more than twenty-eight million active subscribers across three countries — was running a contact centre operation that was structurally underwater. Inbound contact volumes had grown 31% year-over-year as subscriber growth, mobile-money adoption and 4G upgrades all drove more reasons to call. The contact centre, despite ongoing hiring, was averaging a 23-minute hold time at peak. Customer-reported NPS was at a five-year low.
The operator already had a basic chatbot — a rules-based tree on its website and Facebook Messenger that handled the simplest FAQ-style questions. Usage was negligible. Customers vastly preferred to call the contact centre, in part because the bot could not actually do anything: it could tell a customer the cost of a bundle but it could not let them buy one, it could explain the SIM-swap process but it could not initiate it, and it could check whether a service was active but it could not activate it.
The operator's executive team had three constraints that had defeated previous attempts to fix this. First, the operator's customer base was predominantly Swahili-speaking with frequent English code-switching, and the previous bot had been built on a global NLP platform that simply did not understand the dialect blend. Second, the operator's BSS (business support systems) stack was a patchwork of an Amdocs CRM, an Oracle billing platform, an in-house SIM-management system and a recently-deployed mobile-money platform — and the previous chatbot had not been integrated with any of them. Third, the operator's regulator required that customer interactions be auditable and replayable for at least seven years.
The approach
We deployed ChatNext (Cn) on WhatsApp Business as the primary channel, with parallel deployment on Facebook Messenger and Telegram. The operator's WhatsApp Business Account had over twenty-two million customers already opted in to receive marketing messages — converting this base to an interactive support channel was the immediate opportunity.
Phase one — fifteen weeks — covered the top thirty customer intents, which together accounted for 78% of contact-centre volume. The intents broke into three categories: information (balance, data usage, bundle prices, retail-shop locator), transaction (bundle purchase, airtime top-up, mobile-money transfer, bill payment), and account management (SIM-swap, PUK retrieval, plan change, fraud reporting). Each intent was built end-to-end — meaning the bot didn't just answer questions but actually executed the transaction against the underlying BSS systems.
The Swahili-English NLP challenge was the work that took the longest. We collected three years of de-identified contact-centre transcripts (approximately 14 million conversations) and used them to fine-tune a custom intent classifier on top of a multilingual base model. The model handles the operator's specific dialect blends — including the Sheng street slang common in urban areas — and routes seamlessly between languages within a single conversation. Where customers use mixed-language inputs ("Naomba kucheck balance yangu" — "Please can I check my balance"), the bot responds in the same blend.
Phase two added six more intents — primarily mobile-money related, including transaction dispute, M-Pesa-equivalent statement requests and merchant-payment troubleshooting. Phase three added agent-handoff to live contact-centre agents for complex cases, with full conversation context handed off so the agent did not need to start from scratch.
The pre-built building blocks
Rather than commission a ground-up build, the engagement leaned on MindMap's pre-built accelerator library — production-tested components that compress what would otherwise be a six-to-nine-month build into weeks.
ChatNext
Multilingual NLP, WhatsApp Business connector, conversation orchestration
NLP Router
Intent routing across bot, human agent and BSS systems
Multi-Channel Agent
Unified WhatsApp / Telegram / Messenger handoff
OnboardX
Step-up authentication for SIM-swap and mobile-money intents
The architecture
ChatNext is deployed in the operator's regional Azure tenant in Johannesburg, with primary processing in-region and an active-passive failover environment in the operator's own data centre. The WhatsApp Business API connection uses Meta's cloud API with end-to-end TLS and message-level encryption for any sensitive content.
The NLP stack runs a fine-tuned Mistral 7B model for primary intent classification (chosen for latency — the model serves at p95 of 320ms) with a Llama 3 70B model as a fallback for low-confidence cases and for the more complex generative responses (such as mobile-money transaction dispute drafting). Both models are served via vLLM on a GPU cluster in the operator's tenant.
Integration with the BSS stack is via an integration layer we built that wraps each backend system in a constrained, audited API. The CRM integration handles customer lookup, profile retrieval and account status. The billing platform integration handles bundle pricing, bundle purchase, airtime top-up and bill payment. The SIM-management system integration handles SIM-swap initiation, PUK retrieval and SIM activation. The mobile-money integration handles balance lookup, transfer initiation, statement retrieval and dispute filing.
Critical transactions — anything that moves money or changes a customer's SIM — require step-up authentication. The bot triggers an OTP via the operator's existing SMS gateway, validates the OTP, and only then executes the transaction. Suspicious patterns (such as a SIM-swap request from a device that has only been seen on the WhatsApp number for less than 24 hours) trigger an additional verification step routed to a human agent.
The entire conversation log is persisted to a regulator-compliant audit store with seven-year retention, full message replay capability and exportable PDFs for any conversation. The operator's compliance team has direct query access to this store.
The numbers behind the story
Inbound contact-centre volume has dropped 44% since full rollout. Total monthly conversation volume on WhatsApp has stabilised at approximately 2.3 million sessions per month. Median time-to-resolution is 12 seconds for informational intents and 47 seconds for transactional intents — both inclusive of authentication.
The contact centre's average hold time has dropped from 23 minutes at peak to 4 minutes, even as overall customer activity has grown. The operator has paused new contact-centre hiring for two quarters. Customer-reported NPS is up 27 points since pre-launch.
Specific outcomes by intent: 81% of bundle purchases now happen on WhatsApp versus 22% pre-launch; 67% of SIM-swap requests are now self-served on WhatsApp (a process that previously required a retail-shop visit); 92% of balance and data-usage queries are now handled by the bot, with median response time of 4 seconds; and 71% of mobile-money transaction-status enquiries are resolved without escalation.
Mobile-money revenue per active user has risen 14% — partly attributable to the fact that customers can now self-serve transfers and merchant payments via WhatsApp 24/7, including outside the operating hours of the contact centre and most retail shops. The bot is the operator's single largest customer touchpoint by interaction volume.
“We had two failed chatbot attempts before MindMap. The difference was that ChatNext actually understood our customers — including the Swahili-English code-switching that had defeated every other engine we tried. Our contact centre has gone from being a bottleneck to having spare capacity. And mobile money revenue per user is up because customers can do things at midnight that used to require a retail-shop visit.”— Chief Customer Experience Officer· East African Mobile Network Operator
Why MindMap was chosen
The operator had previously commissioned two chatbot builds — one with a global cloud platform's conversational AI service, and one with a regional NLP vendor — both of which had failed to handle the language dynamics of the customer base.
MindMap was chosen for three reasons. First, ChatNext's track record in African markets meant we could demonstrate, on similar accents and dialect blends, that our NLP stack actually worked. Our pre-sales demonstration included a side-by-side test of intent classification on a sample of the operator's actual conversation logs — we hit 94% accuracy versus the incumbent's 71% on the same test set.
Second, the integration depth we proposed went beyond what other vendors were offering. We had pre-built integration patterns for Amdocs CRM, Oracle billing and the major SIM-management platforms — meaning the BSS integration work, which had been the bottleneck on previous projects, was four to six weeks of work rather than four to six months.
Third, our pricing model — fixed-fee per intent delivered, with success-fee components tied to deflection rates — aligned commercial outcomes with the operator's business case in a way the incumbent's per-message pricing did not.
Related deployments
AI Voice Agent for Collections
An Arabic-English voice agent replaced the outbound collections dialler, reducing cost per contact by 58% while increasing same-call promise-to-pay rate.
Voice-First African Self-Service
Voice Bot + ChatNext reached 22M subscribers in 6 local African languages — 48% of voice-based customer interaction now AI-handled.
Network Anomaly Detection
Anomaly Detector + AI Ops platform gave the NOC 4-hour-earlier detection of customer-impacting incidents — slashing MTTR by 56%.
Want an outcome like this?
Start with a 2-week AI Readiness Sprint. We deliver a prioritised use-case backlog and business case grounded in what's actually buildable with our accelerator library.