Accessibility agents
Redesigning compliance reporting and AI chat into deterministic, WCAG 2.2 AA conversational workflows.



Executive summary
- Role & Scope
- Product Designer, AI UX at Level Access (April - September 2026). I architected the conversational Reporting Agent overhaul, prompt refinement and artifact versioning flows, and WCAG 2.2 AA streaming response patterns for Ask Level AI.
- Core Decision
- I traded off unconstrained, free-form LLM prose generation for a guided conversational refinement loop backed by schema-locked report versions, preserving natural-language speed while guaranteeing deterministic compliance audit accuracy.
- Verified Outcome
- +23% increase in generated and exported compliance reports; zero WCAG 2.2 AA blockers across VoiceOver, NVDA, and JAWS screen-reader QA.
- Conversational UX
- Chat reporting
- Prompt versioning
- Screen reader architecture
- WCAG 2.2 AA
The constraint and problem
High abandonment and broken screen-reader parity
Level Access's original Reporting Agent trapped users inside rigid multi-field prompt forms, while Ask Level AI lacked chat history, contextual triggers, and accessible streaming patterns. Users suffered steep drop-off when prompts conflicted, and native assistive technology users could not reliably navigate live LLM outputs.
Conversational vs. deterministic tension
Free-form chat risked hallucinated compliance metrics, while rigid prompt forms forced users to restart on edge cases.
Streaming LLM accessibility gaps
Off-the-shelf chat components flooded screen readers token-by-token and broke keyboard focus during generation.
Dead-end chat outputs
Users had no persistent history, artifact versioning, or action triggers to move from insight into remediation.
Key decisions and tradeoffs
Due to confidentiality, I can share scoped views of the Reporting Agent overhaul and Ask Level AI upgrade.
Reporting Agent overhaul: guided flexibility over free-form chat
Tradeoff: Rather than letting the LLM rewrite report layouts arbitrarily on each prompt, I constrained outputs to a versioned, schema-locked artifact canvas paired with conversational prompt refinement. Users iterate rapidly via natural language without risking non-deterministic CSV/PDF export parity.
Prompt refinement for conflicting requests

What I did
- I architected a split conversational-and-artifact workspace with non-destructive version diffing (v1, v2, v3).
- I designed inline disambiguation prompts to resolve conflicting filter criteria without wiping prior report parameters.
Impact
- Increased generated and exported reports by +23%.
- Eliminated data-extraction dead ends when handling multi-constraint enterprise audits.
Ask Level AI upgrade: actionable context and history
I replaced passive Q&A replies with platform-grounded suggestions, explicit feedback triggers, and accessible chat history flows so users close the Find-Fix-Prove loop directly from conversation.
Platform context awareness

What I did
- I introduced contextual follow-up chips, persistent thread history, and explicit thumbs/reason feedback signals.
- I anchored assistant suggestions to active dashboard violations rather than generic documentation links.
Impact
- Lifted daily active engagement across Level AI workflows.
- Enabled keyboard-first navigation between conversation turns and persistent artifacts.
Failure modes and edge cases
Streaming focus management and polite ARIA live regions
Raw token streaming interrupts VoiceOver, NVDA, and JAWS mid-sentence. I engineered semantic announcement batching via aria-live="polite" status summaries, prevented automatic focus hijacking during generation, and enforced explicit Cmd/Ctrl+Enter submission rules so keyboard users never misfire incomplete prompts.

Verified outcomes and what I'd do differently
Verified outcomes
+23% report generation and exports
- 23% increase in compliance reports generated and exported after launching conversational prompt refinement and versioning.
- Sustained growth in daily active users interacting with Ask Level AI capabilities.
Zero WCAG 2.2 AA blockers
Passed strict accessibility QA across keyboard-only navigation, Axe DevTools, VoiceOver, NVDA, and JAWS for streaming LLM output.
Architecture scale
Cross-artifact export parity
Extended the versioned reporting schema to support enterprise CSV, PDF, and Jira ticket workflows.
Native assistive tech benchmarking
Established ongoing usability sessions with native screen-reader users to stress-test multi-turn edge cases.
What I'd do differently
Senior retrospective
I would involve native screen-reader testers during low-fidelity wireframe prototyping rather than waiting for functional build QA, catching verbose aria-live phrasing two sprints earlier. I would also expose a dedicated deterministic filter drawer beside chat from day one for power analysts who already know exact query parameters.



