Look at any AI announcement from the past year, and notice that what’s being measured is a benchmark score, a context window, a hard exam result. Every one of those numbers describes what the most capable model can do for the most capable operator, on the best hardware, under good conditions. That is a ceiling, and the industry is very good at raising it.

A floor is something different. It’s what the least served person can actually do, and raising the ceiling doesn’t raise the floor (sometimes it does the opposite). Every capability that needs a newer GPU, a bigger budget, or a pair of working eyes lifts the ceiling while leaving the floor exactly where it was.

We started Fair Play to write about teams doing the floor work, so it seemed fair to say what we think the work requires, and to hold ourselves to.

Access isn’t a gradient. Barbershop singers chase what they call the expanded sound: four voices lock a chord precisely enough, and a fifth voice appears that nobody is singing. But the failure mode is the important part. A sloppy quartet doesn’t get a quieter fifth voice, it gets no voice at all. Access works the same way: a form that loses one required field to a screen reader hasn’t lost a tenth of its function, it’s lost all of it. You can’t average access across users and call things mostly fine because it works or it doesn’t. Which means accessibility can’t be a phase two.

We adopted a standard called POURS: Perceivable, Operable, Understandable, Robust, and Safe and Secure. The first four come straight from the W3C’s POUR principles, the backbone of web accessibility guidelines. We then applied it everywhere, not just the web interfaces people look at, but the scheduled jobs and API’s, too. Perceivable applied to an API is asking whether the thing that ran at 3am left behind something a person can actually read.

Then we measured ourselves against it, and we’re publishing the number. We mapped our accessibility against all 55 WCAG 2.2 AA criteria: 5 aligned, 3 known gaps, 47 still unverified. We ran the same exercise on our blind and Deaf support work and scored it 2.6 out of 5, with the lowest score (1.5), on making disability-involved evaluation a routine practice. That’s the most useful thing we learned, because it’s a process gap we can fix by changing who’s in the room. Some of this is genuinely built. Our KeyCenter surface ships a non-visual route with sonification, along with eyes-free, sound-free, and hard-of-hearing presets. Our speech recognition runs entirely on local hardware, and our braille output is still on the roadmap.

None of this is charity, and treating it that way would be condescending. Kyle Keane, a blind researcher who builds the assistive technology he uses, put it best: noise-canceling headphones and a phone camera that describes a room aren’t disability tech, they’re human-enhancing tech. We all move in and out of situational disability, hands full at a door, a noisy room, tired eyes late in the day. That line between assistive and simply useful runs through more of this industry than the product categories suggest. OneCourt is a good example: its haptic tablet translates live sports data into vibrations and audio commentary, so blind and low-vision fans can feel a game as it happens. Built for one group, useful in a way that reframes what watching a game even means.

There’s a cost argument too, and for the small organizations Fair Play covers, it may be the decisive one. If AI gets more expensive the more you use it, it gets expensive exactly as it becomes load-bearing, which is a ceiling business model dressed up as a subscription. We build the other direction: work the system has done many times gets captured into something deterministic that no longer needs a model at all, so a task the platform has mastered costs less every month.

None of this is finished, and we’re not waiting until it is. Our next step is asking people who’d actually know. We continue to meet with the National Federation of the Blind, and what matters in those meetings is handing them our 47 unverified criteria to learn which ones we got wrong. That’s the standard we’re asking to be held to, and it’s the one we’ll use when we write about everyone else.

What does the floor look like in the thing you’re building, and who’s standing on it?