There’s a lot of noise about AI. Every SaaS product has added a “powered by AI” badge. Most of it is ChatGPT wrapped in a button. What I want to talk about is the narrower, more useful question: what does meaningful AI integration actually look like for African businesses, and where does it fail?

I’ve built AI-powered systems for clients in Kenya and globally. This is what I’ve learned.

The starting point: what AI is actually good at

Large language models (LLMs) are fundamentally good at three things:

Transforming text — summarising, classifying, extracting structured data from unstructured input
Generating text — drafting, completing, translating
Answering questions — when grounded in context (via RAG or fine-tuning)

They are not good at precise maths, real-time data retrieval without tooling, or reliable multi-step reasoning in production. If your AI use case requires accuracy on arithmetic or current stock prices, you’re building in the wrong direction.

Three patterns that work in production for African businesses

1. Customer support triage and FAQ automation

This is the most reliable use case for AI in SME contexts. A well-built chatbot can handle 60–80% of support queries — FAQs, order status questions, standard pricing enquiries — with high accuracy. The 20–40% that are complex or sensitive get routed to a human.

What makes this work in a Kenyan context:

WhatsApp as the primary channel (not a web widget most users won’t see)
Swahili + English mixed input handled by multilingual models
Clear escalation paths to M-PESA transaction IDs or human agents

What makes it fail: trying to make the bot handle everything. The moment customers can’t get through to a human when they need one, trust collapses fast.

2. Document intelligence — extracting structure from unstructured input

Many Kenyan businesses deal with semi-structured documents: invoices, import declarations, HR forms, land registry documents. LLMs are remarkably good at extracting structured fields from these — names, dates, amounts, IDs — even when the format varies.

An API integration pipeline that takes a PDF, extracts the key fields with an LLM, and pushes them into an ERP or accounting system can eliminate hours of manual data entry per week. The economics are compelling even for small businesses.

Key implementation note: this requires validation loops. LLMs hallucinate. Any extracted field that matters — especially financial figures — needs a confidence score and a human review gate for low-confidence outputs.

3. Internal knowledge base assistant

Staff at growing companies spend enormous time answering the same internal questions: “what’s our leave policy?”, “how do I process a client refund?”, “what’s the password for the staging server?” (please rotate those credentials).

A RAG pipeline connected to your internal documentation — Google Docs, Notion, Confluence — can answer these accurately and instantly. It doesn’t replace your documentation; it makes it actually searchable.

This is most effective when the documentation exists and is maintained. If your internal docs are a graveyard of outdated SOPs, an AI assistant makes you faster at surfacing wrong answers.

What the African context changes

Connectivity and latency

LLM API calls take 1–5 seconds for a full response. On a fast Nairobi broadband connection this is fine. On a 3G connection upcountry it’s borderline. Streaming responses (displaying text as it’s generated) significantly improve perceived performance on slower connections — it’s a non-negotiable for user-facing AI features in Kenya.

Language

Swahili support in major LLMs has improved significantly since 2024. Claude and GPT-4 can handle Swahili reasonably well. Sheng and mixed-code Nairobi speech is trickier — models will interpret it but won’t always get the register right. For customer-facing products, always test with native speakers before deploying.

Cost

LLM API costs at scale add up quickly. A 1,000-token GPT-4o call costs roughly $0.005. That’s cheap for a one-off, but if you have 10,000 customer interactions per day, you’re looking at $50/day — $18,000/year — before your own infrastructure costs.

For cost management: cloud architecture choices matter. Use caching aggressively (identical prompts should not hit the API twice). Use smaller models (Haiku, GPT-4o mini) for classification and triage tasks, and reserve larger models for generation. Fine-tuned smaller models often outperform large general models on specific tasks at a fraction of the cost.

Where AI integration fails

Building on a demo, not a product spec — every LLM product looks impressive in a demo. The question is: what happens at the edges? What happens when the input is ambiguous, malformed, in a language the model wasn’t tested on, or designed to trick the system? Build the failure cases before you build the happy path.

No evaluation framework — how do you know your AI is performing well? Without a test dataset and evaluation metrics, you’re flying blind. For any production AI feature, define what “good” looks like before you build.

Over-automating too early — the best AI deployments we’ve seen start with a human-in-the-loop design. The AI suggests; a human reviews and approves. Over time, as confidence thresholds are validated, automation increases. Full automation from day one is a recipe for customer-facing errors.

Ignoring data privacy — sending customer data to an external LLM API raises data residency questions, especially if you’re subject to Kenyan Data Protection Act obligations or handling health, financial, or identity information. This needs a legal review before you start sending data to OpenAI or Anthropic.

If you’re evaluating AI integration for your business and want a technical partner who has built these systems in production — not just talked about them — get in touch. We can scope what’s realistic for your specific context.