Accessible by Design, Not by Retrofit

At university (so a long time ago) a student in my class read on a monitor the size of a small television, with the text magnified until one or two letters filled the screen at a time. Every sentence was a scroll. That same year I recorded lectures and transcribed the notes by hand for a friend who was deaf. Neither of them asked to be an example of anything, and I’ve kept them anonymous here for that reason, but between them they taught me more than I expected. They made me realize just what being visually impaired, or deaf, could mean for someone trying to do the same work as everyone else in the room.

I carried that for a long time without doing much about it. This post is about why, and what changed.

The miss

Some years later, in presales at VT Team, an early Amazon Connect partner, we were working through a prime contractor with a national institute for the blind. They wanted off their existing call center and were looking at Amazon Connect. Connect’s own agent desktop wasn’t going to work for their agents as it stood, so we proposed a custom agent UI that worked with JAWS and the standard desktop tools a blind agent already uses. (AWS documents screen-reader support for it today, and it’s getting better day by day; at the time, in our assessment, it needed help.)

I wanted that project badly, and the fee had nothing to do with it. I argued for doing it at cost, because it was the right thing to do and because I could see the second act; the institute and VT Team co-licensing the UI so that blind agents anywhere in the world could take calls on Connect. A tool and a new way of doing things, and we’d touch a lot of lives with it.

They decided against it and went another way. I don’t blame them; they made a reasonable call on the options in front of them, and I’ve made the same kind of call from the other side of the table more times than I’d like to admit. It has stayed with me anyway. Of all the deals I’ve lost, that’s the miss I want to fix.

What “accessible by design” actually means

Start with the floor, because Gairm is aiming to ship the floor. Screen-reader-native, so every element on the agent desktop has a name, a role and a place in the reading order. Refreshable Braille supported. Keyboard and alternative input for everything, no mouse required, JAWS integrated, WCAG compliant. That’s what the institute would have got in the custom UI, and it’s what most of the industry still treats as a stretch goal.

“‘A few million’ is the number we tell ourselves so that ‘later’ feels reasonable.”

I don’t think the floor is the ambition, though, and here’s where I’ll admit something. JAWS, to my ear, sounds terrible. I’m not a screen-reader user, and the people who are rate the tool highly (WebAIM’s 2024 survey of 1,539 screen-reader users has 95.6% of JAWS users very or somewhat satisfied), so my ear is the wrong one to trust here. Satisfied, though, is one measurement, taken against what exists. I was happy with my tiny first car right up until I got to drive a fancy one, and then I (very quickly) realized how bad mine was. Maybe there’s a fancier car here and maybe there isn’t; that’s for the drivers to say. Build it with the people who use these tools every day, then let them measure again.

What I hear, listening in, is a read-back of the screen, label by label, in a synthesized voice. People who use a screen reader every day don’t listen that way (in the same survey, 71.6% said the first thing they do on a long page is jump between its headings), and they move at speeds I can’t follow. Still, the software hands them the whole screen and leaves them to find what matters.

Here’s what I’d rather build. Instead of reading the screen to the agent, describe it. “Caller’s been on hold four minutes, she’s called twice this week about the same invoice, and the last agent promised a callback that didn’t happen. Start with greeting and immediate empathy, then solve.” That’s what a good colleague leaning over your shoulder would say; it’s a summary, in a human voice, of what matters right now, and it’s exactly what the models are good at. Be My Eyes shipped a version of this idea for photographs in 2023, and Apple’s VoiceOver now describes screen elements even in apps that never bothered to label them, so the pieces exist. Nobody has built the agent desktop around them.

Then there’s Braille. Braille is good, and I’m not going to argue with it - hey, it made it in to our logo! A 40-cell refreshable display is also $3,759 at list, weighs 1.5 lb and is a foot long, and that price lands on someone before the first call is taken. For some people it’s the only way the screen reaches them at all; a spoken description does nothing for an agent who is deafblind. Will fewer people need one in the future? I honestly don’t know, and that call belongs to Braille readers, not me. We need to support Braille displays for many years to come, but that shouldn’t stop us from building a better tool alongside them, as one more choice.

Which is why the overlay is wrong. An overlay is a line of JavaScript that promises to make an inaccessible site accessible after the fact, and the public record on that promise is settled. The FTC’s final order in April 2025 has accessiBe paying $1 million and barred from claiming, without evidence, that an automated product can make any website WCAG-compliant. More than a thousand accessibility practitioners have signed a fact sheet stating that no overlay on the market can make a site fully compliant with any existing standard, I would add my signature if I thought mine would move the needle. The National Federation of the Blind declined accessiBe’s convention sponsorship in 2021 and said its behavior was harmful to the advancement of blind people. And UsableNet (a remediation vendor, so salt to taste) reported that by mid-2024 more than a fifth of digital accessibility lawsuits were against sites already running a widget.

Accessibility needs to be baked in. One day I’d like to see a standard that assumes the describe-it approach rather than the read-it-back one; I’m calling that a hope, because I don’t have a roadmap. If you use these tools every day, or build them, and want to write that standard together, call me. I’ll buy the first coffee.

The other side of the headset

Every accessibility conversation I’ve had in twenty-five years of contact centers has been about the customer. Accessible IVR, accessible chat, an accessible front door. Almost none of it was about the person answering (who, in every one of those conversations, could see the screen).

I live in a world with intentional naivety, so let me say the naive thing. AI will replace most human contact center agents, and never all of them. The world I want to live in is one where the agents who are left come from the communities the labor market walks past: people with significant disabilities and veterans first (that’s Gairm’s Phase 1), then people facing barriers of race, ethnicity or geography. In time, survivors of trafficking and people who’ve been incarcerated. And not just the agents. The supervisors, the engineers and the architects, all from the same group. The point is underserved communities that get beaten to every job by someone the market finds easier to hire. That is Gairm’s mission to the core.

You can’t staff that world with an accessible front door. It needs the agent desktop, the supervisor view and the engineering toolchain all accessible, from line one, because the person you’re building for is going to be in the room. That’s the side of the headset the institute’s custom UI was for, and it’s the side nobody builds for first.

What it used to cost

Cost was the hardest barrier to break, in a world where you can find minimum-wage agents anywhere. An operator meets it three times.

The benefits system. In my hiring experience, agents with disabilities usually command a higher hourly rate, and there’s nothing mysterious about why. Disability benefits in most jurisdictions come with an earnings line (in the US, Social Security’s “substantial gainful activity” threshold is $1,690 a month in 2026, or $2,830 if you’re blind), and every hour you work takes you closer to losing the benefit. The health coverage that often travels with the benefit makes the cliff steeper. So a minimum-wage job is frequently a net loss, and the rational move is to hold out until it isn’t. That’s arithmetic, and the price is set by the benefits system.

The equipment. The Braille display, the large monitor, the licenses, the kit to get someone to take their first call. Then the speed and productivity hit of working through all of it, and that hit belongs to a desktop that was never designed to be read aloud. (In fairness to the budget line, most accommodations cost less than their reputation - in JAN’s survey of employers, 61% said theirs cost nothing, and where there was a one-time cost the median was $300. Contact-center kit sits at the expensive end of that.)

The employer. The readiness of the workplace and the facilities to have that person work for you, when they are likely the only visually impaired employee out of hundreds or thousands. Set it up for one person and you pay full price for everything, every time. That’s the budget argument retrofit lost every single time, and I lost it along with everyone else.

In the time and tokens era, building quality accessible products should be easy. I want to be careful with “should”, because the research says the default is the opposite - AI-generated UI is inaccessible unless you ask for it, and even then it needs checking. The CodeA11y authors put the share of web pages with accessibility violations at around 96%, and the models learned from that web. Cheap is not the same as automatic.

It’s a choice at line one, made once, and after that it costs about what the rest of the software costs, which is a great deal less than it did (that’s my Time and Tokens post in a sentence). It’s also the gateway. Build the desktop so that it describes itself and you’ve widened how people access technology, and not only how they access your queue.

”We’ll add it later”

Sure, but that sounds a lot like the check is in the mail. It will come, and there will always be something that’s a higher priority, because the effort-versus-reward looks so small; why build for a few million people globally when you could build for the eight billion others?

Two things about that arithmetic. The first is that I don’t judge the CX leader who does it. I honestly had it for 25 years. It was a problem I wanted to fix and couldn’t, because I had those competing priorities too.

The second is that the numerator is wrong. The WHO puts significant disability at 1.3 billion people, one in six of everyone alive, and counts at least 2.2 billion with a near or distance vision impairment. “A few million” is the number we tell ourselves so that “later” feels reasonable.

I’ve just chosen a different path; my stakeholders are different now (I wrote about why). A nonprofit’s stakeholders include the people your customers serve, and when the people you serve are also the people you employ, “later” stops being available as an answer.

So Gairm is set out to build it now, with employees who are exactly who the products will be for. We’ll ship the floor, then we’ll build the thing I wish we’d built for the institute. And I’ll call them one day, show them what we built and ask them what we got wrong.

PS. Hold me accountable - ask me on LinkedIn, or in the grocery line on how we are doing against this goal. I’ll be happy to tell you with honesty and transparency.

---

Further reading: WebAIM, “Screen Reader User Survey #10”, for who actually uses what; the “Overlay Fact Sheet”, the practitioner consensus on overlays in one page; the FTC’s accessiBe order; Mowar et al., “CodeA11y: Making AI Coding Assistants Useful for Accessible Web Development”, and Guriţă and Vatavu, “When LLM-Generated Code Perpetuates User Interface Accessibility Barriers, How Can We Break the Cycle?” (W4A 2025), on why the default is inaccessible and what prompting changes; and the Speech Accessibility Project at the University of Illinois, the same argument made for voice input.