Large Language Models in Africa: Real Uses, Costs and Risks

A large language model (LLM) is software trained on huge amounts of text so it can read, write, summarise and answer questions in plain language. ChatGPT, Claude and Gemini are the best-known examples. African businesses are already using them to answer customer questions on WhatsApp and SMS, give farmers and patients advice in Swahili, draft documents and sort through messages, and the cheapest models now cost a fraction of a US cent per conversation.

The catch is not the price of the model. It is accuracy, local languages and the data protection rules you sign up to the moment customer data leaves your office. This guide covers all three, with verified examples and the arithmetic shown.

Key Takeaways

• Start with one job that is mostly text and mostly repetitive, such as answering the same 20 customer questions, not with “AI strategy”.
• Budget for people and setup, not the model: for a small business the model fee is often under a few dollars a month, while review time and integration cost far more.
• Keep a human in the loop for anything that affects money, health or a customer’s rights. Kenya’s best-run LLM services all do this.
• Test the model in the languages your customers actually write in, including Sheng and code-switching, before you go live.
• Before you send customer data to an overseas AI provider, sign a data processing agreement and check the cross-border rules in the Data Protection Act.

What is a large language model, in plain terms?

An LLM predicts the next piece of text, one small chunk (a “token”) at a time, based on patterns it learned from training data. That lets it draft an email, summarise a tender document or answer “do you deliver to Thika?” in natural language.

What it does not do is know your business: your prices, stock or returns policy. You supply that in the instructions or by connecting your documents. When it has no good information, it can produce a confident, wrong answer. People call this “hallucination”, and it is the single biggest reason LLM projects go wrong.

There are three ways businesses use LLMs:

• Chat apps such as ChatGPT, Claude or Gemini, used by staff directly in a browser or phone.
• APIs, where a developer connects a model to your WhatsApp line, website or internal system and you pay per token used.
• Open or local models such as Meta’s Llama family, which a technical team can adapt and host itself.

Our view: most Kenyan SMEs should start with the first option, prove the value, and only then pay for an API build. If you are choosing between the main chat apps, our Gemini vs Claude vs ChatGPT comparison for Kenyan businesses goes through the trade-offs.

How are African businesses actually using LLMs in 2026?

These examples all have a public, named source.

Maternal health advice by SMS (Jacaranda Health, Kenya). Jacaranda’s PROMPTS service answers pregnant women and new mothers by SMS in English, Swahili and Sheng. According to NPR’s September 2026 report, it runs at public clinics in 24 Kenyan counties, now handles about 15,000 questions a day, and flags roughly 7% of messages as possibly urgent for a nurse’s priority queue. A team of 18 nurses monitors and audits the answers around the clock. Canada’s IDRC describes Jacaranda’s Kiswahili health model, UlizaLlama, as open source, and reports Jacaranda’s own estimate that the program still makes errors in 30% of cases, which is why a human takes over when a mother is not satisfied.

That last number is the most useful fact in this article. One of the most respected LLM deployments in Kenya still relies on humans to catch mistakes. Plan for the same.

Customer care at scale (Safaricom). In June 2025, TechAfrica News reported that Safaricom had extended its work with AWS to use generative AI in customer experience, using large language models through Amazon Bedrock. Its Zuri chatbot handles a range of customer queries and reduces traffic to the Care Centre.

Farm advice in local languages (Digital Green, Kenya). Farmer.Chat is an AI advisory chatbot for smallholder farmers. The GitLab Foundation, which funded it, reports 14,000 users and 260,000 questions answered in six languages by the end of 2023. A randomised trial across 600 villages in Nakuru County, led by IFPRI, runs from September 2025 to March 2027 to test whether it actually improves yields and income. The honest status: usage is proven, impact is still being measured.

African-language models (South Africa and the region). Johannesburg’s Lelapa AI, founded in 2022, builds InkubaLM, a model designed for African languages including code-switching, and offers the Vulavula API for speech and language tasks, according to Techpoint Africa. Separately, six operators (Airtel, Axian Telecom, Ethio Telecom, MTN, Orange and Vodacom) have joined the GSMA to build models trained on African languages, and developers demonstrated the first open Swahili reasoning model at MWC Barcelona 2026, Techpoint reported. The same report notes that Africa has more than 2,000 languages while most LLMs are trained mainly on English.

Our own use. Our blog and news desk run on a pipeline where a language model drafts each article and automated checks reject invented clients and unsourced statistics. What we learned: an AI writer will invent statistics and client stories unless checks stop it. The details are in our AI publishing system case study.

Notice the pattern. Every working example is narrow (one job), runs in the channel people already use (SMS, WhatsApp, an app) and has humans checking the output. For more everyday examples, see how Kenyan businesses use ChatGPT today.

Which business tasks suit an LLM, and which do not?

If customers write to you in Swahili, read our guide to serving customers in Kiswahili with AI before you choose a model.

How much do large language models cost a business?

There are two price models.

Subscriptions for staff. Claude’s Pro plan is $20 a month billed monthly, and a Team Standard seat is $25 a month billed monthly, on Claude’s pricing page at the time of writing.

API pricing per million tokens (at the time of writing):

• Google’s Gemini 2.5 Flash-Lite: $0.10 input and $0.40 output per million tokens, with a free tier.
• OpenAI’s gpt-5-nano: $0.05 input and $0.40 output; gpt-5-mini: $0.25 input and $2.00 output.
• Anthropic’s Claude Haiku 5.5: $0.10 input and $0.50 output for prompts up to 100K tokens; Claude Sonnet 5.5: $2 input and $10 output.

Worked example (hypothetical): an agro-dealer in Eldoret with five staff wants a WhatsApp assistant to answer product and price questions. Assume 3,000 conversations a month, each using 1,500 input tokens (the question plus the shop’s product list and instructions) and 300 output tokens (the reply).

• Input: 3,000 x 1,500 = 4,500,000 tokens, or 4.5 million.
• Output: 3,000 x 300 = 900,000 tokens, or 0.9 million.

Apply a buffer: prompts grow as you add products, so budget for three times your estimate. Even Sonnet becomes 3 x $18 = $54 a month, still small next to one staff member’s time.

So where does the money go? Into building the connection to WhatsApp (Meta bills WhatsApp Business Platform messages separately: template messages are charged per message, while replies inside an open customer service window are free), keeping your product and price information current, testing, and the staff hours spent reviewing answers. If you want to estimate payback properly, use our guide on how to know if an AI investment will pay off.

What are the risks, and what does Kenyan law require?

Wrong answers. Covered above. Jacaranda’s own error estimate is the reminder: build review into the process from day one.

Data protection. The Office of the Data Protection Commissioner published a Guidance Note on Artificial Intelligence in July 2026, opened for public consultation from 1 to 17 August 2026 according to Digital Policy Alert. It applies the Data Protection Act to AI and says, among other things:

• AI that involves large-scale profiling, automated decisions with significant effects, sensitive data or children’s data will typically need a Data Protection Impact Assessment, under Section 31 of the Act.
• If a cloud AI provider processes personal data for you, you need a written data processing agreement that limits what it can do with that data.
• Personal data cannot go to an offshore AI processor without a lawful transfer basis under Part VI of the Act. If your provider runs its models on servers outside Kenya, this applies to you.

In practice: do not paste customer ID numbers, M-Pesa statements or medical details into a free chat app, and use plans whose terms say your data is not used for training. Our guide to Kenya’s AI compliance requirements covers this in more depth.

Policy direction. Kenya’s National AI Strategy 2025-2030, unveiled in 2025, names healthcare, education, agriculture and public service as priority sectors. Expect more rules, not fewer.

What is “large language model optimisation”?

“LLM optimisation” (LLMO) has two meanings. Engineers use it for making models faster and cheaper to run. Marketers use it for getting your business mentioned when someone asks ChatGPT or Google’s AI a question.

For the marketing meaning, Google’s own guidance is clear. Its page on AI features and your website says there are no additional requirements or special optimisations needed to appear in AI Overviews or AI Mode, and no special AI text files or schema to add. Our opinion: anyone selling you a secret “LLMO package” is selling ordinary SEO under a new name. Spend the money on pages that answer real customer questions.

How do you start using LLMs in your business this week?

1. List your repeat questions. Go through last month’s WhatsApp, email and SMS. Write down the 20 questions you answer most often and how long each takes.
2. Pick one job. Choose the task that is most repetitive and lowest risk. Customer FAQs or drafting quotes beat anything touching payments or health.
3. Write a one-page fact sheet. Prices, delivery areas, hours, returns policy, what to escalate. This is what the model will answer from, so it must be accurate and dated.
4. Test with a chat app first. Paste the fact sheet into ChatGPT, Claude or Gemini on a business plan, then ask your 20 real questions, including the Swahili and Sheng versions. Note every wrong or vague answer.
5. Fix the instructions, then retest. Most errors come from missing facts in your sheet. Add them and rerun the same questions.
6. Decide the human handover. Write down which questions always go to a person (refunds, complaints, anything medical or legal) and who answers them.
7. Check the data rules. Read the provider’s data terms, list what customer data would be shared, and decide if you need a data processing agreement or impact assessment.
8. Only then consider an API build. If the test works, connect the model to your WhatsApp line or website, and review a sample of real conversations every week.

Mistakes to avoid

• Going live without a test set. If you have not tried your 20 real questions, you do not know how often it is wrong.
• Letting the chatbot quote prices from memory. Prices must come from your current fact sheet or system, or customers will hold you to wrong ones.
• Publishing AI-written statistics. Models invent numbers that look real. Check every figure against its source.
• No route to a person. A customer stuck in a loop with a bot is a lost customer.
• Pasting personal data into free tools. It can breach the Data Protection Act and your customers’ trust.
• Paying for the most expensive model by default. Test the cheaper ones first. The worked example above shows a gap of about 20 times between the cheapest and the dearest option.

Frequently Asked Questions

What is the difference between ChatGPT and a large language model?

A large language model is the underlying technology. ChatGPT is a product built on OpenAI’s models, just as Claude and Gemini are products built on Anthropic’s and Google’s models. A business can use the products directly or connect to the models through an API.

Can LLMs understand Swahili and Sheng?

Quality in Swahili and Sheng varies between models, and most models are trained mainly on English. Jacaranda Health’s service in Kenya works in English, Swahili and Sheng, but with nurses auditing the answers. Test any model on your customers’ real messages before relying on it.

Is it legal to use ChatGPT with customer data in Kenya?

It can be, if you meet the Data Protection Act’s requirements. That means a lawful basis for processing, a data processing agreement with the provider where it handles personal data for you, and a lawful basis for sending data outside Kenya. The ODPC’s July 2026 AI guidance note sets out what it expects.

How much does it cost a small business to use an LLM?

A staff subscription on Claude’s Pro or Team Standard plan costs $20 to $25 a month per person when billed monthly, at the time of writing. API usage for a small customer service bot can cost under a dollar a month in model fees, as the worked example shows. The bigger costs are setup, integration and the time spent reviewing answers.

Will an LLM replace my customer service staff?

Unlikely in the near term. The examples here use LLMs for routine questions and for flagging urgent ones, while people handle complaints, exceptions and anything sensitive.

The bottom line

Large language models are now cheap enough for any African business to try, and the examples that work in Kenya share a shape: one narrow job, the channel customers already use, and a person checking the output. Start with a fact sheet and a test, not a big build. If you would like a second opinion on where an LLM fits in your business, our free AI Business Assessment is a good first step, and we quote any build in writing after a conversation about scope.

Questions about your own setup? WhatsApp us on 0711 344 702.