Why start with the channel, not the model?
In many African markets, businesses reach customers through a phone before they reach them through a website. WhatsApp is widely used for service and sales, mobile money handles payments for people without cards, and USSD and SMS reach users on basic handsets.
An assistant that lives only on a desktop site, or in an app that needs a large download, misses much of that audience. Start by mapping how customers contact you today, then place AI where it removes a queue on those channels.
Which journeys suit AI first?
| Journey | What the assistant does | Hand over to a person when |
|---|---|---|
| Order status | Looks up the order and replies on WhatsApp | Delivery fails or the customer disputes it |
| Onboarding | Collects and checks ID and address documents | A document is unclear or details do not match |
| Collections and reminders | Sends payment reminders and confirms receipt | The customer reports hardship or a dispute |
| Support | Answers from approved content and raises tickets | The request is outside policy or the customer is upset |
| Distributor orders | Takes orders by message and updates stock | Quantities or prices look unusual |
Pick the journey with the highest volume and the clearest rules. It is the easiest to measure, and its results give you the evidence to agree harder cases later. Resist launching five journeys at once: one that works builds more trust with customers and staff than several that almost do.
How does mobile money change the design?
Mobile money services such as M-Pesa, MTN MoMo and Airtel Money let customers pay by approving a prompt on their phone. Flows must expect that people may not respond, may respond late, may approve twice or may lose signal halfway.
- Idempotent payments: the same request never charges twice.
- Clear status messages: pending, confirmed and failed are different states, each with its own reply.
- Reconciliation: match provider confirmations to orders automatically, and queue mismatches for a person.
- Official interfaces: integrate through the provider's APIs or an authorised aggregator, which usually requires onboarding in each country.
- Fraud awareness: never ask for a PIN in chat, and say so in the conversation.
Which languages should the assistant support?
Plan language by market and test each one. English and French cover large parts of the continent, Arabic is central in Egypt and North Africa, Swahili is widely spoken in East Africa, and Afrikaans is used in South Africa alongside many other languages, including Zulu, Xhosa, Hausa, Yoruba and Igbo.
Current AI models vary in quality across them, so a sensible approach is honest and incremental.
- Collect real messages in each language, including switching between languages and informal spelling.
- Measure accuracy by language and report it separately.
- Where quality is weaker, narrow the scope, use buttons and structured menus, and offer a human speaker.
- Use native-speaker reviewers to check tone as well as facts.
How do you build for low bandwidth and basic devices?
- Keep pages and app downloads small, and avoid heavy scripts and large images.
- Make apps offline-first, so work is stored on the device and synced later.
- Prefer messaging channels, with short messages and buttons, over long web forms.
- Offer SMS or USSD fallbacks where customers need them, through telecom aggregators.
- Test on mid-range Android phones and throttled networks rather than office wifi.
Which data protection laws should you ask about?
| Country | Main law |
|---|---|
| South Africa | Protection of Personal Information Act, 2013 (POPIA) |
| Nigeria | Nigeria Data Protection Act 2023 |
| Kenya | Data Protection Act, 2019 |
| Egypt | Personal Data Protection Law, Law No. 151 of 2020 |
Each law has its own rules on consent, security, breach notification and transfers across borders. Some treat certain categories, such as children's or health data, with extra care, and sectors such as banking and telecoms have additional rules. This is general information. Use qualified local counsel in each country, and check how the rules have developed since this was written.
How do you hand over to a person on WhatsApp?
A good handover is the difference between a helpful assistant and a frustrating one. Capture the customer's name, request and what the assistant has already tried. Route it to the right queue, show staff hours, and send a message saying when a person will reply. Make sure the person answers in the same thread, so the customer does not start again.
In markets where one phone is shared by a family or a shop, confirm identity before discussing account details.
Which sectors move first?
- Fintech and payments: onboarding checks, support and dispute handling.
- Telecoms: customer care, bundle and billing queries.
- Logistics: delivery status and proof of delivery documents.
- Agriculture: advisory assistants and order flows in local languages.
- Retail and distribution: order taking and stock reconciliation.
How do you measure success?
Choose the measures before launch: the share of conversations resolved without a person, time to first reply, payment completion rate, satisfaction by language and the share handed over. Review them by country and by language rather than overall, because averages hide weak spots.
How should you roll out across several countries?
- Pilot one journey in one country, measured against a baseline.
- Settle the data protection position with local counsel before launch.
- Add the second country with its own test set, payment integration and legal review.
- Share the common components, and keep languages and rules configurable per market.
Our page on AI development for African businesses explains how we work across the time zones, our AI chatbot development service covers WhatsApp assistants, and you can try the live support agent demo or contact us.