Building an AI Feature Into Your Product: Where to Start
"We need AI in the app," said Mutua, the managing director, opening the kind of meeting about AI software development in Kenya that every second boardroom seems to be having this year.
"Sawa. What should it do?" asked Chebet, the product lead.
"You know. AI things. Our competitors have it."
"What do their customers use it for?"
Long pause. Somebody refilled the chai. Somebody else checked their phone. This conversation is happening in boardrooms across Nairobi every week, and it is a perfectly good place to start, as long as it does not also become the place where you finish. Here is how to get from "we need AI" to a feature that earns its place.
Find the painful, repetitive task
The best AI features are rarely the flashy ones. They quietly remove a chore your users hate. Look for tasks in your product, or in the work your customers do around it, that are repetitive, text-heavy or require reading a lot of information to make a small decision.
Customers typing long descriptions that staff then have to categorise manually
Support teams answering the same questions from your help documents
Users filling in forms with information that already exists in an uploaded document
Managers reading long reports to find the three things that need attention
Search that fails because people do not use the exact right words
Talk to your users and your support team. They will tell you exactly where the pain is, usually with feeling.
Do not reach for a chatbot by default
When people imagine AI in a product, they usually imagine a chat window in the corner. Sometimes that is right. Often it is not. A chat box asks users to figure out what to type, and most people would rather click a button that does the thing.
Compare "Ask our AI assistant anything" with a button that says "Summarise this case file" or "Draft a reply". The second is more useful because the product already knows what the user is trying to do. Build the AI into the workflow, not beside it.
AI software development in Kenya rarely means training your own model
You almost certainly do not need to train your own AI model. The large AI providers offer models through APIs that your developers can call from your application, paying based on usage. Your product sends some text, perhaps with relevant data from your own database, and gets a response back.
The real work is in what surrounds that call: choosing what information to send, writing clear instructions, handling the response and presenting it well. Where the AI needs to answer from your own content, such as your help articles or policies, a common approach is to look up the relevant documents first and include them with the request, so answers are grounded in your material rather than the model's general knowledge.
Design for the AI being wrong
AI models are impressive and also, regularly, wrong. They can misread a document, misunderstand a request or state something false with complete confidence. Your design needs to assume this will happen.
Show AI output as a suggestion the user can edit, not a final decision
Make it obvious what was generated by AI
Let users give feedback with a simple thumbs up or down
Keep a human in the loop for anything involving money, legal matters or health
Never let the AI take irreversible actions without confirmation
A draft reply that is right most of the time and easy to correct is a great feature. An automatic reply that is right most of the time and occasionally insults a customer is a crisis.
Think about cost, data and speed early
API calls cost money every time they run, and the cost grows with how much text you send and receive. A feature used by a handful of staff is cheap. The same feature used by every customer on every page load may not be. Estimate usage before you build, and design so you are not calling the AI when you do not need to.
Consider data too. If you send customer information to an AI provider, check their data terms, think about Kenya's Data Protection Act 2019, and send only what is necessary. And consider speed: AI responses can take a few seconds, which feels like forever on a slow connection, so show progress and keep things responsive.
Ship small, measure, then grow
Pick one feature, build a simple version and give it to a small group of real users. Watch how they use it. Do they accept the suggestions, edit them heavily or ignore the button entirely? Measure things that matter: time saved, tickets resolved, tasks completed.
If it works, improve it and roll it out. If it does not, you have learned something useful without betting the whole product on it. That is the boring secret of good AI features: they are built like any other feature, with clear goals and honest testing.
If you have a product and a hunch about where AI could help, our software development team is happy to talk it through. Bring the hunch. We will bring the chai and the awkward questions, like "what should it actually do?"