Human-Centred AI: Designing for the Underbanked
Building conversational AI for low-literacy, multilingual users taught me more about good AI design than any enterprise deployment.
The most complex AI design problem I’ve ever worked on wasn’t inside a tier-one bank. It was designing conversational AI interfaces for migrant workers in the GCC , people who might speak Bengali, Malayalam, or Tagalog as their first language, who may have limited literacy, who are using a smartphone in a labour camp on a Friday afternoon.
That context changes everything about how you design.
The illusion of the “average user”
Most enterprise AI products are designed for a hypothetical average user , someone with reasonable literacy, who speaks the product’s language fluently, who has time and patience to navigate errors. Even when designers think they’re being inclusive, the defaults reveal their assumptions.
At myZoi, we had no such luxury. Our users were the edge case that most fintech products ignore. And designing for them turned out to be the most clarifying design constraint I’ve ever worked within.
When you design for a user who can’t read well, natural language becomes everything. Error messages have to be genuinely helpful, not technically accurate. Voice interfaces can’t rely on precise vocabulary. Every assumption you’ve ever made about how people interact with financial technology gets interrogated.
What changes when you design for real people
Language is not a translation problem. Localising an app into Bengali is not the same as designing for a Bengali-speaking migrant worker in Dubai. The idioms are different. The context is different. The level of trust in digital financial systems is different. You have to do the research , real ethnographic work, not a survey.
Conversational AI needs to handle failure gracefully. In our voice interface work at myZoi, we found that how the system handled misunderstandings mattered far more than its accuracy rate. An AI that confidently gives the wrong answer to a question about a money transfer creates real harm for real people. Designing uncertainty into the interface , “I want to make sure I understood you correctly , you want to send 500 dirhams to your account in Bangladesh?” , was one of the most important design decisions we made.
Trust is built in small moments. For a migrant worker managing money far from home, every successful transaction builds trust. Every confusing error message erodes it. The product has to earn the right to handle someone’s money. That means being honest about what it can and can’t do, never overpromising, and consistently doing what it says it will.
What this taught me about Human Centered use of AI generally
Designing for the hardest users produces better products for everyone. When your AI handles ambiguity gracefully for a low-literacy user, it handles it gracefully for all users. When your error messages are genuinely helpful to someone with limited technical vocabulary, they’re helpful to everyone.
The enterprise AI world tends to optimise for the best case. The real test of a well-designed AI system is how it behaves when the user isn’t the ideal user , when they’re confused, under pressure, using it in their second language, or just having a hard day.
That’s the design challenge I care about. And it’s the one I think produces the most durable products.
Christian Buchholz is a Certified HCD Instructor and Chief Innovation Officer with deep experience designing AI products for underserved communities across the GCC and South Asia.