Accessibility Is the First Question

Before you start

How to use this guide

This is not a transcript. It is the conversation of November 3, organized so that you can come back to it, use it in your own meetings and share it with your team.

Each section follows the order of the talk. The dark boxes reproduce the slides as real text, so a screen reader reads them like any other part of the page. After each box you will find what was said, cleaned up for reading, and, in many sections, a short exercise.

• If you have five minutes, go straight to the four questions. They are the most useful thing in this guide.
• If you have thirty minutes, read Parts 1 to 3. They explain why AI tends to learn from the average and what to do about it.
• If you want to run this with your team, use the session plan near the end. It needs no slides and no technical knowledge.

Contents

1. Opening: “my money doesn’t”
2. Normal people don’t exist
3. Part 1. Design from the edges
4. Part 2. The data that was never collected
5. Part 3. When AI learns from the average
6. Part 4. What AI gives back
7. The four questions
8. Questions from the session
9. Run it with your team
10. From the book: It Was Never About Technology
11. Resources and sources
12. The talk, slide by slide
13. About the author and this guide

Opening

“I have cerebral palsy, but my money doesn’t”

I lived in Canada for about ten years and worked for Microsoft Canada, where one of my roles was Accessibility Lead. So this conversation started where my own learning started: with a friend.

Years ago, in Toronto, I went to a meeting with my friend Dave Dame, who later joined Microsoft. We were meeting one of the largest banks in Canada. Our goal was simple to say and hard to do: we wanted them to change their workplace policies for professionals with disabilities.

Dave has cerebral palsy. He looked at them and opened the meeting with one sentence: “I have cerebral palsy, but my money doesn’t.”

The room changed. He was not asking for charity, or for a favour. He was talking business: customers, talent, and money the bank was leaving on the table. In one sentence, he moved the conversation from pity to value.

Then he asked a second question, a personal one: “Have any of you ever had to hire someone when you were eighteen years old?” He had. To live on his own at university, he had to hire someone to help him with the most personal parts of his day. At eighteen, he was already a manager, choosing someone and trusting that person with his own life.

Later he told me something I never forgot. People in his situation train a muscle that most leaders never train: vulnerability. You learn to ask for help, to trust and to delegate. He believes it should be a required skill for every leader.

Keep that word with you as you read: vulnerability. We come back to it at the end.

Opening

Normal people don’t exist

We invented “normal”. It is a useful idea for statistics and a dangerous idea for design. Think about how many forms you have filled in with a box that did not fit you: a name that was too long, a family that did not match the options, a body that did not match the chart. That small moment of not fitting is what millions of people feel every day, with almost every system.

Here is the concern. AI learns from data, and data is full of averages. Many AI systems are, in practice, being taught that a normal person exists. If we teach machines that, they will design for that person, at scale, everywhere, very fast.

The average pilot

In the late 1940s, United States Air Force planes kept crashing, often with nothing wrong with the machine. The cockpit had been designed in 1926 around the average body of the pilots of that time, with seats, pedals and controls all fixed.

In 1950, a young lieutenant, Gilbert Daniels, measured 4,063 pilots on ten dimensions, using generous ranges: you only had to be close to average, not exactly average. How many pilots fit the average on all ten? Zero. The average pilot did not exist, so the cockpit had been designed for nobody.

The solution was not a better average. It was adjustable design: seats that move, pedals that adjust, straps that fit the person. Accidents went down, and this is where your car seat comes from. The story is told by Todd Rose in The End of Average, by Kat Holmes in Mismatch, and in Chapter 1 of my book.

The average started to move

For two hundred years, the average was a ruler. It measured you after you arrived: at school, at work, at the doctor. Today the average is alive. It learns from every click, every route on your phone, every pause on a video, every like and every comment. It no longer waits for you. It predicts you: the next song, the next route, the next product, sometimes the next job candidate.

Most of the time this is convenient. But a prediction is always a guess based on the people who came before you. That changes everything for people at the edges, especially people with disabilities. Before, they were excluded by the ruler. Now they can be excluded by the prediction, before they even arrive.

Part 1 of 4

Design from the edges

Inclusive design starts by recognizing who is left out.

The best toolkit I know for this comes from Microsoft’s Inclusive Design team. I used it every day at Microsoft, and it is public and free (see Resources). It starts from a simple fact: we all have biases, people with disabilities included, and we tend to assume that others think and live like us. They don’t. So the first move is to recognize exclusion, find the people who are not part of what you are building, and invite them in as co-designers rather than spectators. As the disability rights movement says: nothing about us without us.

Disability as a mismatch

Stairs make a building inaccessible: the problem is the stairs, not the wheelchair. A video without captions excludes a person who is hard of hearing: the problem is the video. A website that does not work with a screen reader excludes a person with sight loss: the problem is the website. An AI that only understands the average voice is a mismatch too.

The good news is that a mismatch is not a fact of nature. It is a design choice, and design choices can be changed. So when something does not work for a person, the first question is not “what is wrong with this person?” It is “what is wrong with this design?” That one we can fix.

Three principles of inclusive design

1. Recognize exclusion. Before you ask “who is our user?”, ask “who are we leaving out?” It sounds small. It changes the whole project.
2. Learn from diversity. The people who live the mismatch every day are the real experts. They know the challenges, the shortcuts and the workarounds. Bring them in as designers from the start, while you can still change direction, not as testers at the end when everything is built.
3. Solve for one, extend to many. When you solve for the person at the edge, you usually solve for many more. The curb cut on the sidewalk was designed for wheelchair users. Today it helps parents with strollers, travellers with suitcases, delivery workers with carts, people on bikes and skates.

Permanent, temporary, situational

One design that works with one hand helps all three people in the first row. If you design only for the permanent case, you think you are serving a small group. You are also serving everyone in the temporary and situational cases, and that is a lot of people. This is why accessibility is never a niche, and why I approach it as an economist and a business person, not only as a matter of charity or legislation.

Not a niche

In Canada, 8 million people aged 15 and older live with a disability that limits their daily activities: 27%, more than one in four. In any group like the one in this session, many people live with a disability or care for someone who does. Some disabilities you can see. Most you cannot.

Around the world, it is about 1.3 billion people. When the Belgian astronomer Adolphe Quetelet invented the idea of the “average man” in the 1830s, the whole world population was smaller than the number of people with disabilities today.

The captions you are reading started at the edge

The session had live captions on, and many people were reading them. Those captions started at the edge. Alexander Graham Bell’s mother was hard of hearing and his wife was Deaf, and his work teaching Deaf people shaped the research that led to the telephone. Vint Cerf, one of the fathers of the internet, is hard of hearing, and his question was never only “how do we make it faster?” but “how do we make sure the message arrives?” He has often said that email mattered to him personally, because text reaches people that sound does not.

Live captions were built for Deaf and hard of hearing people. Today they are how many of us follow a meeting in a noisy airport, or late at night with a sleeping baby in the next room. What started as an accessibility feature became a productivity standard.

A mother at 3 a.m.

At the 2019 Microsoft global hackathon, four of us built an app for parents and caregivers of children with epilepsy, for the phone and the Apple Watch. Our team won globally. From day one we built in accessibility, using the accessibility features that the platform already offers to every developer: adjustable contrast, screen reader support, easy navigation. Not because we thought our users were blind, but because we wanted the app to work in any context, for anyone.

We expected the first feedback to be numbers. It was a mother. At night she turned up the contrast so she did not need her glasses, and let the screen reader speak so her eyes could stay on her child. A feature made for blind users helped a tired mother at three in the morning. Nobody on our team had imagined her. We designed for the edge, and the edge took us to her. We also learned that to care for the child, you have to care for the person who cares.

Part 2 of 4

The data that was never collected

What a building in Toronto taught me about AI.

January 23, 2017. Toronto under snow, everybody walking through the PATH tunnels. I had invited Maayan Ziv, the founder of AccessNow, to visit our office downtown. At nine in the morning my phone rang: she could not find the elevator. I had never noticed that there was no accessible route from the PATH, because I always took the stairs.

I went down and found her in her power wheelchair in a corridor of a technology building, full of people who build the future, a building that had not imagined her. We went through two doors to a service elevator in a forgotten space that opened to the street, at minus twenty degrees. There were two garbage bags inside, and I took them out. On our floor, the glass door was too narrow for her wheelchair, so I found a stool and unlocked the second panel. In the meeting room, the chairs were in the way, and I moved them out one by one while she waited.

I was deeply embarrassed and said sorry more than once. She smiled and said, more or less: “This is my life, Ricardo. It seems nothing was designed for me. Today was different.” Different, because someone came down to look for her. Then she told me something I repeat everywhere I go:

A leader’s real job is to remove barriers. Sometimes the barrier is a staircase, sometimes a door, sometimes a meeting room full of chairs.Maayan Ziv, as I remember her words

And sometimes, as we will see, the barrier is in the data. Notice the two leadership lessons so far: from Dave, vulnerability, the skill of asking for help, trusting and delegating; from Maayan, that leaders remove barriers.

If no one collects the data, the barrier is invisible

Maayan did not stay angry. She built something. For decades, the maps on our phones told us where everything was, but never whether you could get in. Is there a step at the entrance? Is the washroom accessible? Is there a ramp? Nobody collected that, so for the systems those barriers did not exist. AccessNow, which she founded in 2015, collects exactly that: people rate and map the accessibility of businesses and public spaces, together, as a community, around the world.

This is the bridge to AI. If no one collects the data, the barrier is invisible. And what is invisible to the data is invisible to the AI. A model cannot respect a person it has never seen.

Part 3 of 4

When AI learns from the average

Data bias is a design decision no one made on purpose.

Jutta Treviranus, director of the Inclusive Design Research Centre at OCAD University in Toronto, explains this better than anyone I know. Imagine a chart of everybody’s needs. Most people sit in the middle, close together. People with disabilities are spread out at the edges, each one different from the others. For the model, the middle is the signal and the edges look like noise. So the more different you are, the more likely the model is to ignore you or get you wrong.

In 2015, she tested the learning models that guide self-driving cars through intersections, using an unusual but real case: a friend who moves her wheelchair backwards. Every model made the same decision: keep going. In the simulation, every one of them ran her over. The developers said the models were young and needed more data. She tested again with more data, and the models ran her over with more confidence. More data about the average only teaches the model, even more firmly, what the average does.

Now think beyond cars: any system that learns from the past to decide something about a person. Who gets a call back. Who gets flagged. Who gets help first. The same pattern can happen there, quietly and with confidence.

That is actually good news. If no one made the decision on purpose, someone can make it on purpose, the other way. That means choosing who is in the data, who is in the room, and who has the power to say “stop, this does not work for me”.

Five ways bias gets into AI

None of these can be fixed by an engineer alone. Every one of them needs people from the edges in the room, early, with a real voice. If I had to pick one to watch most closely, it would be automation bias, because it is the one that makes good people stop asking questions. It just flows, and nobody notices.

Scenario 1: the voice assistant that does not understand

Scenario 2: the résumé gap

Part 4 of 4

What AI gives back

The fear in the room, and the work that remains human.

It is a fair question, and I will not pretend it is not there. My answer, after years helping organizations adopt AI, is this: AI takes the repetitive. What remains is the human work.

Copying numbers from one screen to another. Searching for that one paragraph. Formatting the same report every week. AI is very good at that, and honestly, nobody will miss it. What remains is judgment, empathy, negotiation and care. Knowing when a rule does not fit the person in front of you. Listening to a person on the phone who is scared and confused, and making them feel heard. No model does that the way a human being does.

And here is the honest part: nobody trained us for this. Schools did not. Companies did not. There is no MBA in empathy. So the real question is not only “what will AI do?” It is “what will we do with the time and attention AI gives back to us?” For people with disabilities, this matters twice, because AI can finally close many of the mismatches in this guide, if we design it with them.

“Being vulnerable is horrible”

My brother lost part of his leg in a motorcycle accident. The first time he went out after the amputation, I was driving him. We stopped at a pharmacy and I asked: “Do you want me to go in for you?” He said: “No. I will go.” He went on his crutches. A car almost hit him, and people coming the other way did not wait for him to pass. They were in a hurry.

When he came back to the car, he cried in a way I had never seen before. When he calmed down, he said: “Being vulnerable is horrible.” This is a man who served in the army, who always did everything on his own.

The barrier that day was not his leg. It was the street, the hurry, and the people who did not see him. Today he runs with one prosthesis and swims with another. And I keep thinking about Dave, at the beginning of this guide.

Vulnerability is a skill. It is where empathy begins. And empathy is where good design begins.

Closing · The tool to keep

The four questions

If you take only one thing from this guide, take these four questions. They came from Katarina Subotic and the Accessibility Advisory Council. I could not have written them better.

1. Who could this create a barrier for?Finds the people. Name them: not “users”, but the caller with a speech disability, the applicant with a career gap.
2. What assumption is the technology making about who uses it?Finds the hidden assumption: about voice, vision, attention, language, time, a career that goes in a straight line.
3. What experience is missing?Finds the gap in the data, and in the room. Whose life is not in the records, and who is not at the table?
4. How would this be different if accessibility were there from the start?Gives you the cost of doing it late, and usually a better design for everyone.

Use them in your next project, your next policy, your next purchase of an AI tool. They are short on purpose. You can ask them in any meeting, in any role. You do not need to be a technical person. And please, ask them in the first meeting, not the last one, when it is too late and too expensive to change.

We started with a simple idea: normal people don’t exist. Here is the other side of that same idea. Accessibility is not the last step. It is not the compliance check the week before launch. It is the first question. Every time we ask it first, someone like Maayan finds the elevator, someone like my brother crosses the street, someone like that mother sleeps a little better.

Q&A

Questions from the session

The Q&A was moderated by Zahra Ali. These are the questions answered live, with my answers written out more fully. Questions that did not fit in the time will be answered in the next version of this guide, at the same link.

Which of the five biases is the most common in tools we already use, and how do we spot it?

Automation bias. Things get automated for convenience, and after a while decisions are simply happening without anyone noticing how, or on what criteria. That is why I think it is the most dangerous. To spot it, ask two questions about any automated step: what criteria is it using? and what is its impact on people? If nobody in the room can answer, you have found it.

Can automation bias be fixed after a model is already live?

Yes, but it takes more work. Once a system is running with customers and users, fixing it can mean pulling it back, sending it “to the maintenance centre”, and dealing with the dissatisfaction of people who depend on it. So it is possible, but the trade-off is big, especially when the solution is live. That is exactly why the four questions belong in the first meeting.

How do we create guardrails for privacy and confidentiality in AI tools, especially for marginalized groups who are afraid of sharing sensitive information that could identify them?

The best practice I have seen is a responsible AI committee: an engaged group connected to the leadership of the organization, looking at every use of AI through five lenses: inclusiveness (not only disability, but everything that makes us human and unique), fairness, accountability, responsibility and security. The members must have real power to say “approved”, “not approved”, “go” or “no-go”, and the people most affected must have a voice inside it. This will not be solved by technical knowledge or by software. It will be solved by people in decision-making roles with the power to make hard decisions.

In my book, the chapter “Privacy Is Power” adds a simple test: collect only the data that is strictly necessary, treat sensitive inferences with explicit caution, and make sure people can really say no.

Put it to work

Run it with your team

A 45-minute session plan. No slides, no technical knowledge, any team.

From the book

It Was Never About Technology

An English extract of the key ideas of “Nunca Foi Sobre Tecnologia: Decisão humana, poder e responsabilidade na era da inteligência artificial” (Alta Books, 2026). Until now, the book has only existed in Portuguese, so this extract is a first.

The book is not a technical manual and not an apocalyptic warning. It is an investigation of how we got here, from factories that standardized bodies, through statistics that turned people into averages, to algorithms that try to turn uniqueness into prediction. Its central question: when we delegate decisions to machines, who is really deciding?

Eight ideas, one per chapter

1. The origin of the average. From the Air Force cockpit to Quetelet’s “average man”, Taylor’s factory and the IQ test, the average was a tool to organize the world. It also carried a powerful fiction: that a standard human exists, and that whole systems can be designed around that person.
2. From the average man to Homo Tech. Each technological revolution arrives faster than the last. Until now we automated things: force, movement, production. Now we automate something much more delicate: decisions.
3. The algorithmic tribe. The internet helped us find our people, and algorithms learned to keep us there. Personalization can connect, and it can also enclose.
4. Innovation at the extreme. Accessibility is not an extra. It is the stress test of a system, because it is at the extreme, not in the comfortable centre, that a system reveals its true assumptions. The “niche” is in fact the largest invisible market in the world.
5. The human, ambidextrous leader. Launching Copilot for Microsoft 365 taught me that we updated the machine and forgot to update the way we treat the people who operate it. Technology moves in silence; culture keeps making the same noise. Leaders now need both business and technology fluency.
6. Privacy is power. Privacy is turning from a quiet right into a visible privilege. Not every convenience is neutral. Some quietly redefine what freedom means.
7. The age of hyper-personalization. Hyper-personalizing is not knowing people better; it is deciding before them. Responsibility does not disappear with automation. It only changes place. Consent is not a click, and without reversibility there is no consent, only compliance.
8. It was never about technology. AI does not create values. It executes them with growing precision. If there is exclusion, it scales it. If there is bias, it amplifies it. If there is care, it sustains it.

It was never about technology. It was always about how we choose to treat people.Nunca Foi Sobre Tecnologia, Chapter 8

Ten questions before you automate

The book closes with twenty questions to ask before automating, scaling or hyper-personalizing anything. Here are ten of them, the ones closest to this conversation.

1. Who becomes invisible when this system works well?
2. What definition of “normal” is built into this design?
3. Am I automating a task, or a judgment?
4. If this system gets it wrong, who pays the price, and how?
5. What is being inferred without the person knowing it?
6. Who can question, review or stop this decision?
7. Was this system tested from the average, or from the extreme?
8. If I were the person most harmed, would I accept this design?
9. What does this system consider a “deviation”, and why is that treated as a problem?
10. What kind of error can this system not perceive? Systems do not see what they cannot measure.

The book is available in Portuguese at Amazon Brazil (amazon.com.br/dp/6582931016). An English edition is in progress.

Go further

Resources and sources

Toolkits and organizations

• Microsoft Inclusive Design: the inclusive design toolkit, the persona spectrum and the guide “In Pursuit of Inclusive AI” (the five biases). inclusive.microsoft.design
• Inclusive Design Research Centre, OCAD University (Jutta Treviranus), Toronto. idrc.ocadu.ca
• AccessNow (Maayan Ziv): rate and map the accessibility of places. accessnow.com

Books

• Todd Rose, The End of Average (2016): the story of the average pilot and Lieutenant Gilbert Daniels.
• Kat Holmes, Mismatch: How Inclusion Shapes Design (2018).
• Ricardo Wagner, Nunca Foi Sobre Tecnologia (Alta Books, 2026), in Portuguese.

Data cited

• Statistics Canada, Canadian Survey on Disability, 2022: 8.0 million Canadians aged 15 and older (27%) have one or more disabilities that limit their daily activities.
• World Health Organization, Global report on health equity for persons with disabilities (2022): an estimated 1.3 billion people, about 16% of the world’s population, experience significant disability.

Appendix

The talk, slide by slide

The sequence of the 28 slides, with the approximate time of each block in a 41-minute talk. Use it to find a moment in the recording, or to retell the talk in your own words.

Times are the planned times from the start of the talk. The live session ran a little longer, and the stories were told in more detail.

About

About the author and this guide

Ricardo Wagner is an economist, author and advisor on AI adoption, based in São Paulo, Brazil. He spent eighteen years at Microsoft, about ten of them in Canada, where he was Accessibility Lead for Microsoft Canada, and later became Director of Copilot for Microsoft 365 in the Americas. He has lived and worked in 47 countries and given more than 500 talks. His book “Nunca Foi Sobre Tecnologia” (It Was Never About Technology) was published by Alta Books in 2026.

Website: oricardowagner.com.br/en
LinkedIn: linkedin.com/in/ricardowagner
Instagram: @rimatswagner

Thank you

To Katarina Subotic, who made this conversation happen and wrote the four questions with the Council; to Stefan Martinovic, Zahra Ali and Lhaiza, who prepared and reviewed every detail; to Janine Dyck, who opened the session; and to everyone who joined live, read the captions and asked questions.

About this guide

Version 1.0, November 2026. English and French editions. Written by Ricardo Wagner from the slides and the recording of the session; the stories are told as he remembers them. The statements in this guide are the author’s own and do not represent the views of the WSIB. You are welcome to share this guide, in full or in part, with credit to the author. The next version will add answers to the questions that did not fit in the Q&A.