A blind user cannot glance at the screen and notice that the AI got it wrong. There is no second channel, no visual sanity check, no half-second where something looks off. There is the output, and there is the decision they make based on it.
That single constraint shaped everything I now believe about building AI products.
SMARTON is an AI-powered accessibility platform we built at Sunbots Innovations. It reads printed and digital documents aloud, describes images, tables, and diagrams, supports academic learning, and handles daily tasks like identifying currency notes. It runs today with over a thousand users and more than fifty partner institutions.
Here's what building it taught me.
Sighted users are your error-correction layer, and you didn't design it
Most AI products quietly depend on the user catching mistakes. The summary looks wrong, so they reread the source. The number seems off, so they check. The generated text is clumsy, so they edit before sending.
You didn't design that safety net. You inherited it. And it's doing enormous work in your product's reliability story.
Remove it and everything changes. If a user is going to act on your output without reviewing it, the honesty of that output stops being a nice-to-have and becomes the entire product.
Confidence is not a technical detail
Early on, we treated model confidence as an internal signal — useful for logging, not for the user. That was wrong.
When someone asks what a currency note is worth and acts on the answer immediately, the difference between "this is a 500" and "this looks like a 500, but the image is blurry" is not a UX nicety. It's the difference between a useful tool and a dangerous one.
A system that hides its uncertainty is lying by omission. We started surfacing confidence directly in the voice output. Users didn't find it annoying. They found it trustworthy — and started relying on the system more, not less, because they finally knew when to double-check.
Speed matters more than you think, and less than accuracy
Accessibility taught me an uncomfortable trade-off. A screen reader that takes eight seconds to describe a scene isn't slow — it's unusable, because the world has moved on. Latency isn't a performance metric here; it's a functional requirement.
But we still refused to trade honesty for speed. A fast wrong answer, delivered confidently, to someone who can't verify it, is worse than no answer.
So we optimized everywhere else: on-device processing where possible, offline capability, aggressive caching of common cases. What we never optimized away was the system's willingness to say it wasn't sure.
The institutional user is a second user
SMARTON is deployed through schools, NGOs, and CSR programs. That means there are two users: the person using the app, and the institution that has to administer, justify, and trust the deployment.
Building for both forced discipline that made the product better. Institutions want audit trails, usage analytics, and content controls. Those requirements sound like enterprise overhead — but they're the same transparency the end user needs, expressed at a different level.
The standard travels
The most useful thing about building under these constraints is that they transfer.
When we built Marketing Autopilot for agencies, the users could all see their screens. But the same questions applied: can they tell what the AI used? Can they tell how sure it is? Can they undo it? An agency sending AI-generated content to a client's audience is, functionally, acting on output they can't fully verify either.
Designing for the highest-stakes user doesn't produce a product that only works for them. It produces a product that's honest by default — and that turns out to be what everyone wanted.
FAQ
What is SMARTON? SMARTON is an AI-powered accessibility platform built by Sunbots Innovations. It combines a voice-first mobile app, visual intelligence, and a web platform to give blind and visually impaired users independent access to documents, images, and everyday tasks.
What does accessibility-first AI design mean? Designing for users who cannot visually verify the AI's output, which removes the informal error-correction most products depend on and forces the system to be honest about its own uncertainty.
Why does designing for accessibility improve products for everyone? Because it eliminates shortcuts. A product that works when the user can't check the AI's work is more honest and more reliable for users who can.
How should AI communicate uncertainty? Directly, in the output itself, in whatever channel the user is already using — not buried in metadata or a settings panel.
I'm Ravi Jadav, Chief Product Officer and Co-Founder at Sunbots Innovations, where we built SMARTON. If your organization is working on accessibility or AI deployment, get in touch.