--- title: "Accessibility Agents: Building Human-Centered AI Workflows for Trusted Accessibility Automation at Scale" lang: en author: "Jeff Bishop - BITS, an affiliate of the American Council of the Blind" description: "The Accessing Higher Ground 2026 deck, with speaker notes." --- # Accessibility Agents Building Human-Centered AI Workflows for Trusted Accessibility Automation at Scale - Event: Accessing Higher Ground 2026 - Presenter: Jeff Bishop - BITS, an affiliate of the American Council of the Blind - Join: letitglow.app/w/your-code Speaker notes are included under each slide. 26 slides. --- ## 1. Accessibility Agents Accessing Higher Ground 2026 Building Human-Centered AI Workflows for Trusted Accessibility Automation at Scale Today you build an accessibility agent. By four o'clock it is part of a team, and part of an open-source project, with your name on it. Jeff Bishop - BITS, an affiliate of the American Council of the Blind ### Speaker notes, slide 1 > The proctor introduces the session. Put the microphone on before you say a word, and keep it on all day: when someone speaks from the floor, pass them a mic or repeat what they said into yours before you answer. > > Say the key point before any logistics, in one sentence: accessibility in higher education does not scale by fixing, it scales by making more people who can, and agents are how one person's know-how reaches a whole course. > > Describe the slide: the title, the subtitle from the program, and one sentence promising what they leave with. --- ## 2. Forty thousand files and three people The problem WCAG 2.2 is the standard. Title II made it a deadline for public colleges and universities. Every syllabus, slide deck, scanned reading and course page is in scope, and most of them are documents, not web pages. Nobody fixes that by working harder. The question is how to scale without losing quality, accountability or trust. ### Speaker notes, slide 2 > Ask, warmly: who here has a backlog? Who has a spreadsheet with the word final in its name more than once? Let the laugh happen; it is recognition, not mockery. > > Title II's rule points to WCAG 2.1 AA; 2.2 is the current standard and includes everything 2.1 does. Say it once and move on. --- ## 3. The finished result, first Where we will be at 4:00 This afternoon a team of agents works through a whole course and writes one report. Every finding has its evidence, its WCAG criterion, and a named person who checks it. One section of that report, from Maria's agent: | Part | What it says | |---|---| | File | The scanned required reading | | Routed to | Alternate format planner, by Maria Alvarez | | Evidence | GLOW audit: pictures of text, no text layer, untagged | | Plan | Get a clean source, then large print and audio, and post it for every student | | Reviewer | Alternate format specialist. Status: proposed | You will build one of the agents that writes a section like this. ### Speaker notes, slide 3 > Read the table out loud, row by row. It is the destination; people relax when they can see it. > > Maria is a colleague in a disability resource center. She is fictional, and she is one step ahead of the room all day. --- ## 4. The journey The map for today 1. The problem: a course full of barriers, and a team too small. 2. Your agent: five plain-English answers. 3. Your agent at work: Copilot uses it on one course document. 4. Your agent, grounded: real evidence in, cited answers out. 5. Your agent, shared: in an open-source collection, with your name on it. 6. Your agent in the team: the whole course, at once. 7. Your campus: a 30-day plan. ### Speaker notes, slide 4 > Read all seven, slowly. This list comes back at the start of every block with one step marked, so nobody ever wonders where we are. > > Promise them out loud: every step has a finished example to fall back on. Nobody gets stuck and stays stuck. --- ## 5. How today runs Housekeeping - 10:30 to 4:30 Mountain Time. Lunch at 12:15. A break at 2:10. - Everything is at letitglow.app/ahg. Not set up yet? A helper is coming to you. - Seven blocks. Every one starts by showing you the finished result. - The last five minutes, at 4:25, are the session evaluation. - Ask for anything, any time: large print, a seat, a pause, a repeat. ### Speaker notes, slide 5 > Name the exits and the restrooms, physically pointing. > > Say the last bullet slowly, then add: that is not an interruption of this workshop - it is this workshop. > > If anyone is not set up, do not wait for them. Helpers sit with them while the room carries on; the morning is designed so they catch up by lunch. --- ## 6. Private stays private, and people decide Two rules, all day 1. Never paste anything private into any AI: student records, health or disability information, anyone's name. We use a sample course all day. 2. Every agent ends with a person checking its work. Agents draft. People decide. If nobody is named, nobody reviewed it. ### Speaker notes, slide 6 > Set the AI context here, in under a minute. What today covers: agents that analyze course content, find barriers and draft fixes, grounded in checkers and cited standards. What it does not: replacing anyone, or trusting any answer without evidence. > > Say what makes it different from the other AI sessions this week: it is about higher education accessibility work at scale, built by the people who do that work, and every agent is open source. > > Name the reasons people hesitate - privacy, accuracy, other people's jobs - once, without dwelling. Rule one answers the first. Grounding answers the second. Rule two answers the third. --- ## 7. What an agent is No code, really - Role: Who it is, and who it works for. - Task: The one job it does. - Trusted guidance: The standards and policies it must cite: WCAG 2.2, your own procedures. - Output format: What comes back, in a shape a busy person can use. - Human review: Who checks it before anything goes out. An agent is those five answers, written in plain English. Writing one is writing, not coding. ### Speaker notes, slide 7 > This is the slide that lowers the shoulders. Say it twice: if you can write instructions for a new colleague, you can write an agent. > > The file format is the same one the open-source Accessibility Agents project uses, which is why their agents and yours can work in the same team this afternoon. --- ## 8. Your accessibility office Agents in action The agent team you will join this afternoon. A coordinator routes each file to these specialists. | Specialist | Takes | |---|---| | Word documents | Syllabi, handouts | | PowerPoint slides | Lecture decks | | Excel workbooks | Gradebooks, templates | | PDF documents | Readings, forms, scans | | Web pages | Course pages, with axe and Accessibility Insights | | Captions and media | Lecture captions | | Plain language | Anything a student must read and act on | | Standards reviewer | Everyone's findings, last | | Your agent | Whatever you build it to do | ### Speaker notes, slide 8 > Read the specialists down the list. Pause on the last row: the team has an empty chair, and it is theirs. > > These specialists are adapted from the open-source Accessibility Agents project. That is where their agents go this afternoon too. --- ## 9. Design your agent Block 2 of 7 - 10:50-11:50 - 60 minutes By the end of this block: your own agent, in a file, with your name on it. - Documents and alternate formats: An alternate format planner, or a document triage agent. - Course content and faculty coaching: A faculty coach, or a course page coach. - Compliance and procurement: A remediation log keeper, or a vendor report reader. Pick one card. Bring one real problem from your job, with nothing private in it. ### Speaker notes, slide 9 > Say where we are on the journey: step 2 of 7, "Your agent: five plain-English answers." Then show what they will have at the end of this block before anyone starts. > > Every card has two ready-made agents in examples/agents. Nobody starts from a blank page; everybody starts from something that already works and makes it theirs. --- ## 10. Maria's five answers Watch me first Maria, from the Disability Resource Center, designing her alternate format planner: | Question | Maria's answer | |---|---| | Role | An alternate format planner in a disability resource center | | Task | Plan the formats a student needs from a checker's report | | Trusted guidance | WCAG 2.2 1.4.5 and 1.1.1, ACB large print, our procedure | | Output format | Numbered steps with who does each, then a checklist | | Human review | The specialist checks page by page; the student confirms | ### Speaker notes, slide 10 > Read Maria's answers aloud. Then type /design-my-agent in Copilot Chat on the projector and answer as Maria, so they hear the whole conversation once before they have theirs. > > Joke if the room is warm: Maria's coffee machine has been out of order since August. Her agent does not drink coffee, which is its main advantage. --- ## 11. Copilot asks, you answer Your turn 1. Open Copilot Chat: Control+Alt+I on Windows, Command+Control+I on a Mac. 2. Type /design-my-agent and press Enter. 3. Answer the questions, one at a time. Borrow from the example whenever you like. 4. Copilot writes your answers into `my-agent/SKILL.md` and reads it back. You are on track if Copilot tells you one thing your agent does well. Stuck? Copy a ready-made agent from examples/agents into my-agent. That counts. You are still in. ### Speaker notes, slide 11 > Helpers move now. Look for the person who has not typed anything yet, and sit with them quietly. > > Step card 2 has every key and every expected screen in words, for screen reader users and anyone who prefers paper. --- ## 12. Your agent at work Block 3 of 7 - 11:50-12:15 - 25 minutes By the end of this block: your agent's first answer on a real course document. 1. In Copilot Chat, type /try-my-agent. 2. Pick a file from the PSY 101 sample course. 3. Read what your agent says. It will guess a little. That is not a mistake. It is the setup for this afternoon. ### Speaker notes, slide 12 > Say where we are on the journey: step 3 of 7, "Your agent at work: Copilot uses it on one course document." Then show what they will have at the end of this block before anyone starts. > > PSY 101 is fictional: a course about memory and procrastination that keeps forgetting things and running late. The barriers in it are planted on purpose; the answer key is in the facilitator folder. --- ## 13. Maria's first answer Watch me > "Run Acrobat's Make Accessible action, add alternative text to all images, and increase the font size. This will make the document compliant with WCAG 2.1." - It never saw the file, so it did not know the reading is a scan. - Alt text on pictures of text would describe the pictures, not turn them into text. - It said "compliant", which her agent is told never to say. ### Speaker notes, slide 13 > Read the quote in a confident voice, then the three problems in a normal one. The contrast is the joke, and the lesson. > > This is a written example, labelled as one in the kit. On the day, show a real first answer from the room if someone offers one. --- ## 14. Back at 1:15 Lunch - 12:15-1:15 You have an agent. That was the hard part. After lunch it stops guessing. ### Speaker notes, slide 14 > Say well done, and mean it. Most people in this room had never written an agent at 10:30. > > Anyone still setting up: lunch is when a helper finishes it with them. --- ## 15. Ground it Block 4 of 7 - 1:15-2:10 - 55 minutes By the end of this block: the same agent, giving cited answers from real evidence. A before and an after. 1. In Copilot Chat, type /ground-my-agent. 2. Use the same file as this morning. 3. Your agent reads that file's checker report from sample-course/evidence: GLOW for documents, axe for the web page. 4. Compare the two answers. ### Speaker notes, slide 15 > Say where we are on the journey: step 4 of 7, "Your agent, grounded: real evidence in, cited answers out." Then show what they will have at the end of this block before anyone starts. > > Grounding is the trusted in the session title. The checkers find; the agent explains and plans; a person decides. --- ## 16. Before and after Watch me Maria's agent, on the scanned reading: | Before evidence | After evidence | |---|---| | Guessed the file was an ordinary PDF | Knew it was two pages of pictures of text | | WCAG 2.1, no links | WCAG 2.2, 1.4.5 Images of Text, linked | | "Compliant" | What was checked, and what a person must check | | Fix the PDF | Get a clean source, then large print and audio for everyone | ### Speaker notes, slide 16 > Read both columns, row by row. > > Then ask one volunteer to read their own before and after. One is enough, and applaud it. --- ## 17. Any checker, and a person The skill, not the tool The same works with Word's and PowerPoint's own Accessibility Checker, Acrobat's checker, or Accessibility Insights. The skill is giving your agent evidence. And some barriers no checker sees: - The accommodations statement, last, in 8-point gray, contradicted two sections earlier by a no-makeup-exams rule in capitals. - Alt text that says "image.png". - A lab that only works if you can see color. - Captions that turn Ebbinghaus into "ebb in house". ### Speaker notes, slide 17 > If Word is handy, show Review, then Check Accessibility on the syllabus, and paste its results into Copilot with the agent. Same skill, different tool. > > Those four barriers are why every agent has a human review step. Read each one; the room will recognise every one of them from their own campus. --- ## 18. Back at 2:20 Break - 10 minutes Next, your agent goes public. Nicely. ### Speaker notes, slide 18 > Short break on purpose. The next two blocks are where the day comes together. --- ## 19. Share it Block 5 of 7 - 2:20-3:00 - 40 minutes By the end of this block: your agent in the open-source AHG 2026 collection, committed in your name. 1. Open letitglow.app/ahg/share, and choose your `my-agent/SKILL.md`. 2. Press Open the share form on GitHub. The form opens, already filled in. 3. Tick the privacy box, and press Submit new issue. 4. Within a minute, the reply: Your agent is in. You did this once already, with your practice agent. Same again. ### Speaker notes, slide 19 > Say where we are on the journey: step 5 of 7, "Your agent, shared: in an open-source collection, with your name on it." Then show what they will have at the end of this block before anyone starts. > > Everyone shared a practice agent before the day, so this is the second time, not the first. Say so; it changes the room. > > Helpers watch the issues labelled needs-a-hand and go to those people first. A Nearly there reply always says exactly what to fix. --- ## 20. In the collection, with your name on it Open source, live As your agents arrive, the collection grows, here, on the screen. Tonight every one of them goes into the Accessibility Agents project, with every author credited, for anyone in higher education to use and improve. ### Speaker notes, slide 20 > Project agents/README.md in github.com/Community-Access/ahg-2026 and refresh it as agents arrive. Read each name and agent title out loud. Applause is allowed and encouraged. > > Each agent is committed in its author's name, so it shows on their own GitHub profile. Started from a ready-made agent? It still counts. Say so: what matters is that it is yours now. --- ## 21. Build the office Block 6 of 7 - 3:00-3:45 - 45 minutes By the end of this block: your agent working inside a team, across the whole course, and one team report. 1. In Copilot Chat in VS Code, type /run-the-office. 2. The coordinator sends each course file to its specialists, and yours where it fits. 3. The standards reviewer checks everyone's work. 4. You get the team report, with your agent's section in it. ### Speaker notes, slide 21 > Say where we are on the journey: step 6 of 7, "Your agent in the team: the whole course, at once." Then show what they will have at the end of this block before anyone starts. > > This is building an agent team: they added a specialist, and their coordinator is running it. Copilot announces each file as it finishes, so screen reader users hear progress, not silence. --- ## 22. One course, one report What scale looks like - Seven files, eight specialists, plus yours. - Every finding: evidence, a WCAG 2.2 criterion with its link, and a named reviewer. - A list of what no checker could see, for a person. - Three lines at the top that a director will actually read. That is quality, accountability and transparency, at scale. It is also a Tuesday. ### Speaker notes, slide 22 > Ask: what would it take to run this across ten courses? Let them answer. The honest answer is a person's review time, which is exactly where it should go. --- ## 23. Everyone's agents, together Agents in action One more run, on the screen: the office team with every agent this room shared. Listen for yours. ### Speaker notes, slide 23 > Copy every agents/// folder from the ahg-2026 repository into a room/ folder in your kit, then run /run-the-office on the projector. Read out each agent's name as the coordinator calls it. > > If the run is slow, narrate it. If it stumbles, say what happened and why; a real failure explained well teaches more than a perfect demo. --- ## 24. Take it home Block 7 of 7 - 3:45-4:15 - 30 minutes By the end of this block: a 30-day plan for your own campus. 1. In Copilot Chat, type /my-30-day-plan. 2. Then add your one-sentence commitment at letitglow.app/w/your-code/11 for the wall. What you can do on Monday, with tools you already have: | You learned | Monday, with | |---|---| | Writing an agent | Any AI assistant, including your campus Copilot | | Grounding it | Word's, PowerPoint's and Acrobat's checkers, Accessibility Insights, axe | | Human review, in writing | Your own office's process | | Agent teams | One specialist at a time, or the Accessibility Agents project | ### Speaker notes, slide 24 > Say where we are on the journey: step 7 of 7, "Your campus: a 30-day plan." Then show what they will have at the end of this block before anyone starts. > > Small is the point. "Run the planner on the next five requests" beats "transform our office". --- ## 25. The commitment wall 4:15 Every commitment in this room, on one screen, with no names on it. ### Speaker notes, slide 25 > Project the wall from GLOW. Read three aloud. Do not comment on them, do not rank them. > > Then go straight to the evaluation. The last five minutes belong to it. --- ## 26. Go make one more champion 4:25 - Thank you You came in with a problem. You are leaving with an agent, a team and a plan. Before you go: the session evaluation. It shapes next year's conference. Help any time, before or after today: support@community-access.org. GLOW is a community project of BITS, an affiliate of the American Council of the Blind. ### Speaker notes, slide 26 > Say it, thank them, and hand over to the proctor for the evaluation. Then stop talking while people fill it in. > > No question block here. Answer at the tables while people pack up. ---