AI in Swahili and Sheng: Where the Machines Still Struggle
Achieng runs a small online cosmetics shop from her flat in Kilimani. Last month she tried connecting an AI tool to her WhatsApp business line so it could answer customer questions while she slept. The first message that came in at 11pm said: "Niaje, hiyo lotion iko kwa stock ama imeisha? Nataka mbili, nitatuma doh kesho." The bot replied with a polite paragraph about sunscreen in formal Swahili that sounded like a news anchor reading a government notice. The customer sent a single laughing emoji and bought elsewhere. Her first Swahili AI chatbot experiment had lasted exactly one message.
Welcome to one of the most interesting gaps in AI right now: language, especially the way Kenyans actually speak.
Standard Swahili: better than you might expect
Let us give credit where it is due. Today's AI tools handle standard Swahili reasonably well. They can translate a simple notice, draft a formal letter or summarise a short article. For straightforward business writing, such as a price list, opening hours or a polite reminder, the results are often usable with a quick review by a fluent speaker.
That said, the quality still tends to be noticeably better in English than in Swahili. The models have simply read far more English text. You may find the Swahili correct but stiff, with word choices that feel a little like a textbook from Form Two.
Sheng: a moving target
Sheng is where the machines really start to sweat. It changes quickly, varies between neighbourhoods and borrows freely from English, Swahili and several Kenyan languages. A word that is fresh in Eastlands this year might be outdated by the time it spreads to Rongai. AI tools learn mostly from written text that already exists online, so they are always a few steps behind the street.
Common trouble spots include:
Slang for money, like doh, mbao or ngiri, which AI may misread or ignore
Playful spellings and shortened words used in texts and WhatsApp
Words that mean one thing in standard Swahili and something quite different in Sheng
Jokes, sarcasm and teasing, which can easily be taken literally
New phrases that appeared after the tool finished learning
Code-switching: the everyday Kenyan sentence
Most Kenyan customers do not write in pure English or pure Swahili. They mix both in a single sentence, sometimes with a word from their mother tongue thrown in for seasoning. "Please confirm kama delivery itafika leo, niko job hadi saa kumi." For humans, perfectly clear. For many AI tools, this kind of mixing can lead to mixed-up replies, awkward translations or a sudden switch into one language that feels cold.
Voice is another challenge. Speech recognition has improved, but Kenyan accents, background noise from a busy stage and people switching languages mid-sentence can still trip it up. If you are thinking about voice bots, test them with real recordings from real customers, not a quiet office.
Proverbs, politeness and culture
Language is not just words. Kenyan business communication is full of courtesy, indirectness and context. A customer who says "tutaona" might mean "maybe" or might mean "definitely not". An elder may expect a greeting before any business is discussed. Proverbs and idioms carry meaning that a literal translation flattens completely. AI does not reliably pick up on these cues, which matters when you are dealing with customers, partners or communities where respect is everything.
How to set up a Swahili AI chatbot wisely
None of this means you should avoid AI. It means you should use it with your eyes open:
Use it for first drafts of Swahili content, then have a fluent speaker review before publishing
Avoid fully automated replies to customers who write in Sheng or mixed language
Give the tool examples of how your customers actually speak and how you want to reply
Keep a human escape route in any chatbot: a clear way to reach a real person
Test regularly with your own staff, especially younger team members who speak the current slang
Do not use AI translation for contracts, legal notices or anything with serious consequences without professional review
The good news is that African languages are getting more attention from researchers and developers, and things are improving. Local data and local testing will make a big difference over the next few years.
Until then, Achieng has found a middle ground. Her bot now handles simple questions in English and standard Swahili during the day, and anything that sounds like Sheng gets a quick reply: "Sawa, nimekuona! Achieng atakujibu asubuhi." Sales are back up, and the news anchor voice has been retired.
Your next step: collect twenty real customer messages from your own WhatsApp line and run them through any bot you are considering. Count how many replies you would actually send. If you want a chatbot built around how your customers really talk, talk to Philsoft.