Mental Models for the AI EconomyAccessibility
Your AI Chat Box Is Probably Unusable With a Screen Reader
Streaming replies, focus jumps and silent errors lock out the users chat could help most. What breaks, the eight checks that fix it, and a five-minute test.
Almost every product now has a chat box. Most of them were built in a sprint, by copying the look of ChatGPT, and tested by people who can see the screen. That is how a feature meant to make software easier ends up being the one part of the product a blind customer cannot use.
The good news is that chat is not a new kind of interface. It is a text field, a button and a list of messages that change. The web has had rules for each of those for years, written down in WCAG. What is new is streaming: the answer arrives a few characters at a time, and that one design choice breaks most implementations.
The streaming problem
Sighted users watch text appear and find it reassuring. A screen reader user hears what the page tells their software to announce. If the reply area is a live region that updates on every token, the reader starts speaking, gets interrupted by the next update, starts again, and so on. Depending on the screen reader, people hear a stutter of fragments, only the last word, or nothing at all. WCAG's status messages criterion (4.1.3) asks that changes like “a reply has arrived” be announced without stealing focus. It does not ask that every character be read as it lands.
One reply, three ways to announce it
“I found three flights under $400. The cheapest leaves at 6:10 a.m. Do you want me to hold a seat?”
The fix is small: keep the visual stream, but announce in sentences or once at the end, with a short status such as “Assistant is replying” while it works. Put the finished reply in a polite live region, never an assertive one, so it waits for the user to finish what they were doing.
The four failures I see most
Focus jumps. After sending, focus moves to the new message or the top of the page. Keyboard users lose their place. Keep focus in the message box.
No way to stop. A long answer that cannot be paused by keyboard fails the user who needs to read slowly and the user whose screen reader is mid-sentence. A real Stop button fixes both.
Structure flattened. Models return headings, lists, tables and code. Rendering them as one block of text throws that structure away. Render the Markdown into real HTML elements, the same point I make about formats in HTML or Markdown for AI Agents?.
Silent errors. “Something went wrong” in red text, announced to nobody. Errors need to be announced and need to say what to do next.
Eight checks for a chat interface
Why this is a product decision
If your support is moving into a chat assistant, an inaccessible assistant is an inaccessible support channel. In the EU that can fall under the European Accessibility Act, which covers customer service for in-scope services. In the US, chat widgets already appear in accessibility lawsuits. But the stronger reason is the one from Accessibility Is a Product Decision: the people who rely on a screen reader, voice control or zoom are also the people for whom a conversational interface could be the biggest improvement. Getting it wrong wastes the best use case you had.
Test it the cheap way first: unplug the mouse, turn on VoiceOver or NVDA, and ask your assistant three questions. You will know within five minutes.
Sources
- W3C: Understanding WCAG 2.2 Success Criterion 4.1.3 Status Messages
- MDN: ARIA live regions
- W3C: WCAG 2.2
Across the series
- The European Accessibility Act Is Live. Here’s What It Asks of Your Product.
- Your Visitors Already Said No. Is Your Site Listening?
- Seats, Usage or Outcomes: Pricing an AI Product Without Losing Your Margin
- Can You Train AI on Your Customers’ Data? Ask Them First
- Roadmapping When the Model Gets Better Every Seven Months
- Accessibility Testing: A Green Scan Is Not a Finished Journey
- AI Data Retention: ‘Not Used for Training’ Does Not Mean Deleted
- Agent Evaluation: The Demo Worked. Can You Ship It?
Get the next essay the day it publishes.
A few a month, written from real builds. Unsubscribe in one click.
Performance marketing and search leader who builds — 15+ years in digital growth and product (AT&T/DirecTV, Microsoft, Employbridge, Site Impact), now running his own properties with AI. These essays are field notes from that loop: what AI actually changes about work, product, and the economics underneath. Based in South Florida.
- AccessibilityThe European Accessibility Act Is Live. Here’s What It Asks of Your Product.
Since June 28, 2025, listed digital services sold to EU consumers must be accessible. Who is in scope, what “accessible” means in practice, and a 30-day plan that does not start with an overlay.
9 min read · interactive - AccessibilityAccessibility Testing: A Green Scan Is Not a Finished Journey
What automated scans miss, how WCAG evaluation differs from disabled-user research, and a practical evidence sprint for sign-in, checkout and support—using W3C's July 2026 evaluation guidance.
9 min read · interactive - AccessibilityAccessibility Is a Product Decision, Not a Final Audit
A beautiful interface can still lock people out. How to design for independent completion, test the real journey, and treat access as product quality.
8 min read · interactive - Web & AI SearchReddit, Forums and the AI Recommendation Engine
Why community discussion feeds AI answers, what the Google and OpenAI Reddit deals changed, and a practitioner's playbook for earning mentions without astroturfing.
8 min read - AI EconomicsSynthetic Customers: Test the Funnel Before AI Traffic Finds It
Browser agents can pressure-test a journey before launch. Here is where synthetic users find real friction — and where plausible simulations become expensive theater.
7 min read · interactive - Privacy & TrustYour Visitors Already Said No. Is Your Site Listening?
About a dozen US states treat Global Privacy Control as a legal opt-out, and California browsers must send one from 2027. How the signal works, what it should switch off, and a live check of your own browser.
9 min read · interactive - AI EconomicsAgentic Checkout: Is Your Store Buyable by AI?
Being recommended is no longer the finish line. AI agents need reliable product state, policy, authority, and checkout rails to complete the transaction.
7 min read · interactive - Web & AI SearchAnswer-First Writing: The Inverted Pyramid for Machines
Put the answer in sentence one under a question-shaped heading. The four rules that make a paragraph quotable by snippets, voice and AI assistants — with an interactive before/after.
7 min read · interactive - AI DevelopmentThe Practice: Spec-Driven, Maker-Checker, Governed Scope
Two credible trials on AI coding point in opposite directions — 55.8% faster in one, 19% slower in the other. The difference isn't the model. It's the practice. Three patterns, with interactive math and cited data.
9 min read · interactive