When the App Said 'No' — And How I Learned to Listen to the Numbers
A wrong assumption about user behavior led to a costly mistake in the Arise & Shine Transporters project — but it also revealed the value of letting data guide decisions, not just intuition.
When I first built the AI-powered receipt verification feature for Arise & Shine Transporters, I assumed that drivers would always upload clear images of fuel receipts. After all, that’s what I would do. But the app didn’t behave as I expected. Instead of smooth verification, it began flagging receipts as invalid — and not because of poor image quality, but because the AI was confused by the local currency symbols and formatting used by drivers in Kenya and Tanzania. I had overlooked a basic but crucial detail: the way fuel receipts are presented in the region varied widely, and the AI wasn’t trained on that diversity.
This mistake cost time and trust. The feature didn’t work as intended, and the team had to revisit the AI training data. It was a humbling moment. I had trusted my assumptions, not the data, and that led to a delay in the feature rollout. But it also taught me a valuable lesson: when working with AI, it’s not enough to build for what we think users will do — we need to build for what they actually do.
I revisited the problem with a fresh perspective. Instead of trying to force the AI to conform to a generic format, I adjusted the training data to include real fuel receipts from the region. I also added a feature that allowed drivers to input key details manually if the image wasn’t clear enough. The result was a more robust system that worked with the realities of the users’ daily lives. It wasn’t a perfect solution, but it was one that respected the data and the people using it.
This experience reinforced the importance of testing with real users and real data. AI is a powerful tool, but it doesn’t replace human judgment — it amplifies it. In the end, the mistake didn’t define the project. What it did was remind me that building AI-powered tools for everyday life in Kenya means being open to learning from the unexpected. It also showed me the value of collaboration with AI assistants like Claude and Codex, which helped me refine the feature in ways I wouldn’t have thought of on my own.
If you’ve ever built something and found that it didn’t work the way you expected, you’re not alone. Mistakes like these are part of the process — and they’re often the best teachers. What matters is what we do after we see the app say 'no.'
