Accessibility is the Blueprint for Trustworthy AI
AI changes how software behaves, but it doesn't change what people need to trust it. Accessible design is the solution.
As AI becomes more conversational, autonomous, and embedded into our daily work, people increasingly expect greater visibility, verification, and control. They want status indicators, transcripts, and ways to pause, review, and confidently move forward.
These are the kinds of needs that accessibility practitioners are accustomed to addressing for people with disabilities. Now that AI has made those needs universal, it’s smart to draw from accessibility principles as a blueprint for trustworthy AI.
Let’s explore:
What users need to trust AI
Usability testing uncovers the gaps
Research reveals where trust breaks down
Apply accessibility principles to AI experiences
Invest in accessibility early
What users need to trust AI
For most of the history of software, trust came from predictability.
The same search returned the same answers. A workflow followed a predictable path. Good UX helped people understand what actions were available and what would happen next.
AI changes that relationship.
AI systems are probabilistic and don’t always generate the same outcome. An agent can make recommendations, complete tasks, or draw conclusions based on incomplete information. Instead of interacting with fixed workflows, people are interacting with systems that interpret, reason, generate, and adapt. That changes what users need in order to trust the experience.
Many of the challenges we describe as “AI trust problems” are actually concerns about visibility, control, and feedback. Beyond wondering whether AI is correct, people are asking:
- Did the system understand what I said?
- Why did it make this recommendation?
- How do I stay in control if it gets something wrong?
Our research on AI voice experiences reflected these questions. Participants consistently asked for ways to verify conversations, understand when the system was listening, and review information before taking action. These requests weren’t really about voice. They were about trust.
Usability testing uncovers the gaps
Long before AI introduced the unpredictability of non-deterministic experiences into mainstream software, people with disabilities often had to navigate digital surfaces without reliable feedback, clear orientation, or a sense of control. Accessibility practitioners have spent decades designing for those moments.
In our inclusive usability testing, multiple participants with different access needs using various forms of assistive technologies evaluated an Agentforce Voice prototype. The connection between visibility and trust became immediately clear. When participants couldn’t tell whether the agent was listening or couldn’t access a live transcript, confidence quickly gave way to uncertainty.
As one participant who uses the NVDA screen reader explained, “It would be helpful to have a sound or something to indicate that the voice feature is active so I can start speaking.”
These weren’t edge-case requests. They were fundamental requests for visibility, feedback, and control. People trust technology when they understand what it’s doing, what it expects from them, and how to stay in control. Accessibility has always been about making those things clear.
Research reveals where trust breaks down
One of the enduring principles of inclusive design is “solve for one, extend to many.” When you solve a constraint for one group of people, you often end up creating a better experience for many others.
Dark mode, for example, was originally developed for users with cataracts and low vision. It’s now used by billions today to reduce eye strain or preserve battery life. Voice assistants were originally created for people with mobility disabilities and are now used every day by people who are driving cars, holding children, or multitasking and need a hands-free alternative.
People with disabilities are often the first to encounter friction in a digital experience – not because their needs are unusual, but because they quickly expose where an experience lacks clarity, feedback, flexibility, or resilience. Their interactions reveal where a system asks too much of the user, leaves too much open to interpretation, or fails to provide enough guidance to move forward.
Our Agentforce Voice research highlighted this clearly. Participants using magnification and alternative navigation methods experienced challenges discovering how to get started and interpreting ambiguous visual cues, such as a sound wave icon that lacked the familiarity of a microphone. Later in testing, when voice mode hid the live transcript by default, low-vision participants lost an important point of reference. As one participant using OS Magnification explained: “I wanted to have a reference, which is this chat box, where I could also go through it while it was speaking to me.”
These weren’t isolated accessibility challenges. They revealed foundational usability problems.
A synchronized transcript is an accessibility feature. But it can help anyone verify information, revisit details, and stay oriented throughout an AI conversation whether they’re using a screen reader, working in a noisy environment, or simply trying to confirm what the system heard.
Participants repeatedly described how hands-free voice interaction simplified complex tasks and made their work more efficient. As one JAWS screen reader user explained: “I place a lot of value on the [voice feature] because that actually saves me time from having to type, and allows me to do what I’m doing with Salesforce a lot faster.”
Apply accessibility principles to AI experiences
For someone using a screen reader, screen magnifier, voice interface, or another form of assistive technology, seemingly small gaps in communication can have outsized consequences. If information appears without context, a state change goes unannounced, or a system behaves unexpectedly, users can quickly lose their sense of orientation. When people lose their sense of orientation, they also lose trust in the experience.
To solve this, accessibility practitioners developed a set of enduring principles for creating digital experiences people can perceive, operate, understand, and reliably use. These principles became the foundation of the Web Content Accessibility Guidelines (WCAG), first released in the 1990s as the internet rapidly expanded and digital experiences rarely accounted for people with disabilities. Together, these principles make up the acronym POUR:
- Perceivable: Can I tell what’s happening?
- Operable: Can I interact with it?
- Understandable: Can I predict how it works?
- Robust: Will it work with the technology I rely on?
The POUR principles have remained central to accessible design, and the questions map well to today’s AI design challenges:
- Can users tell when an AI agent is listening?
- Can they interrupt it?
- Can they understand why it produced a particular answer?
- Can they use it with the tools and technologies they already depend on?
Different technology. The same design questions. And the same core principles.
The internet has transformed dramatically since WCAG was first introduced, evolving from desktop websites to smartphones, touch interfaces, and now AI-powered agentic experiences. As technology has evolved, the success criteria and implementation guidance have evolved with it, helping teams apply the same core accessibility principles to new interaction models. AI is the latest chapter in that story.
Invest in accessibility early
Usability testing and accessibility research are valuable. The barriers and opportunities identified by people with disabilities today often become the usability expectations of everyone tomorrow. As AI becomes more autonomous and adaptive, users increasingly need to understand system actions, recover from unexpected outcomes, or change how they interact. Accessibility research allows product teams to catch these edge cases early to solve friction points before they become widespread frustrations.
The future of trustworthy AI won’t be built by treating accessibility as a checklist. It will be built by treating accessibility as the blueprint.
