The dominant design philosophy in AI product development is optimise for speed. Reduce friction. Get the user to the answer faster. Make the interface disappear. Make the output instantaneous. For many use cases, this is right. For accessibility use cases, it's often wrong.
What speed optimisation misses.
Speed optimisation assumes that the bottleneck is information retrieval. That what stands between a user and a good outcome is the time it takes to get from question to answer. For many accessibility contexts, this assumption fails. The bottleneck isn't retrieval. It's processing, decision-making, confidence, and agency. The user doesn't just need the answer faster — they need to understand the answer, trust it, and feel confident acting on it.
Consider someone with cognitive disability using an AI health tool. A fast output that's dense, probabilistic, or qualified with six caveats may be technically correct — but it doesn't help. It may actively harm, by creating anxiety without resolution. The friction that matters isn't the two-second response time. It's the cognitive friction of processing an answer that doesn't give you a clear path forward.
Or consider someone with low digital literacy navigating an AI-powered government service. The interaction design might be smooth and fast for an experienced user. For someone encountering this kind of interface for the first time, "fast" isn't a feature — it's a source of error. The best thing the AI can do is slow down. Confirm understanding. Ask again. Give space.
The form that takes longer on purpose.
One of the most instructive examples I've encountered is a digital form built for a community health organisation. The original form was built to be as frictionless as possible — minimal steps, quick completion, optimised for throughput.
The community it served were people navigating complex health needs in crisis. The form was fast, but completion rates were terrible. People abandoned midway. The data collected was incomplete. Staff had to follow up constantly.
When the organisation ran a co-design process with the community, what came back wasn't "make it faster." It was "give us more time. Let us know what's coming. Don't make us feel rushed." The redesigned form took longer to complete. It introduced more steps — not fewer. It added confirmation screens, plain-language explanations of why each question was being asked, and natural pauses between sections.
Completion rates went up significantly. Data quality improved. Staff follow-up calls dropped. The "slower" form was more effective because it met people where they were — rather than where a frictionless-interaction philosophy assumed they were.
What this means for AI design.
Output format matters more than output speed. A response that arrives in two seconds but requires significant cognitive work to interpret is less accessible than a response that takes five seconds and arrives in a format the user can immediately act on.
Confirmation and reflection have value. For high-stakes decisions — health, safety, legal, financial — slowing down is a feature. An AI that confirms your understanding before presenting a recommendation respects the weight of the decision being made.
The user's confidence matters as much as the output's accuracy. A technically accurate answer that the user doesn't trust, or that creates anxiety rather than resolution, hasn't done its job.
Co-design changes the requirements. Every example in this article has the same root: the people who built the tool closest to the people who would use it discovered requirements that a frictionless-interaction philosophy would never have surfaced. Accessible AI is co-designed AI. There's no shortcut.
A different optimisation target.
The shift isn't from fast to slow. It's from optimising for speed to optimising for confidence, agency, and clarity. When they're in tension, clarity wins.
For organisations building AI tools that will serve people with complex or high-stakes needs: know what you're optimising for. Speed is not a proxy for quality. Frictionlessness is not a proxy for accessibility. Sometimes the most useful thing an AI can do is ask again.