Research that shaped a live government service
During my co-op placement at the Ontario Ministry of Transportation, I contributed to UX research and design for the Ontario Occupational Exposure Registry (OER) โ a public-facing self-tracker that allows workers to record and monitor their exposures to hazardous substances in the workplace.
My core contribution was a two-part research study: Part A synthesized usability insights from three participant sessions on the OER prototype, and Part B evaluated the use of AI tools in the research workflow itself โ measuring time savings, quality trade-offs, and practical guardrails for responsible AI integration in UX research.
I also served as Student Lead for the co-op cohort, co-organising team activities and a hybrid cross-office event across three Ontario locations.
Usability gaps in a tool where errors carry real consequences
The Challenge
The OER self-tracker prototype had several usability gaps that risked real harm: users confused the personal tracker with official regulatory reporting, unclear "What happens next" guidance, dead-end pathways for unknown substances, and affordance issues that caused abandonment. The challenge was to surface these issues rigorously through usability testing and translate them into prioritised, actionable recommendations.
The Approach
Three moderated usability sessions with participants from diverse backgrounds. Human note-taking captured nuance and edge cases. AI tools assisted with transcription, theme clustering, and drafting artifact shells โ then each output was validated against the raw session recordings. A time study measured exactly where AI added value and where human judgment was non-negotiable.
Users thought they were filing an official report โ they weren't
Mental Model Mismatch
Across all three sessions, participants sometimes interpreted the OER experience as "official reporting" rather than "personal tracking." When clarified, their perceived value shifted โ they saw it as documentation support, recall aid, and self-protection. If this mismatch isn't addressed, users may delay medical care or workplace reporting under the false belief that submitting a self-tracker entry triggers external action.
This finding reframed the entire design priority: the platform's purpose must be made explicit at every stage, not just on the first screen.
Six patterns repeated across every session
Documentation is the adoption driver โ users value the tool as evidence, recall support, and self-protection, not as a reporting mechanism.
Proof and retention are trust-critical. Users need a download, email confirmation, or confirmation number to trust the record was saved.
"What happens next" must be explicit, with clear guidance distinguishing between urgent medical action and routine workplace reporting.
Substance selection must match real-world knowledge โ users need synonyms, "Other โ specify," and "I'm not sure" options throughout.
Flexible date precision prevents false precision and abandonment. Users shouldn't be forced to specify exact dates they can't recall.
Accessibility and affordances are essential โ primary actions need clear visual hierarchy, scroll cues, and predictable navigation throughout.
Three priority tiers, ranked by user impact
Findings were organised into three priority tiers based on user impact and feasibility:
Must โ Highest Priority
- Clarify the tool's purpose (personal tracker, not official report) early and repeatedly throughout the flow
- Add explicit "What happens next" guidance with separate pathways for urgent help and official reporting
- Fix all "Other" dead ends โ always include "Other โ please specify" to prevent abandonment
- Support flexible dates and "I'm not sure" options consistently across every date field
Should + Could
- Default proof/retention: download + email + confirmation number after every submission
- Define minimum critical dataset; progressively disclose optional fields to reduce overwhelm
- Improve affordances: primary button prominence, scroll indicators, and consistent navigation
- Support multi-exposure entries and a review/edit workflow before finalising proof
- Guided "unknown substance" workflow (brand name, use, location, SDS sheet lookup)
AI cut research time by ~70% โ with human review non-negotiable
Part B of the study was an independent research initiative I led: a structured evaluation of where AI tools genuinely help in a UX research workflow, and where human judgment remains non-negotiable.
Measured Time Savings
Manual research workflow: 12h 1m total (review 6h 18m + synthesis 5h 43m). AI-assisted workflow: 3h 37m total (transcription 2h 20m + synthesis 1h 17m). Overall reduction: ~70% โ with ~63% saved on review/transcription and ~78% on synthesis.
Workflow time comparison
Per-session savings: P1 โ 2h 39m ยท P2 โ 3h 27m ยท P3 โ 2h 18m
~70% overall reduction in research workflow time
Savings breakdown: ~63% on review/transcription, ~78% on synthesis โ but with the caveat that every AI output required human validation before use.
Where AI Added Value
- Transcription first pass (required human correction for accuracy, formatting, and anonymisation)
- Structured note formatting from session recordings
- Cross-session theme clustering โ accelerated pattern identification
- Drafting artifact shells (personas, journey maps) for human editing and validation
Where Human Review Was Essential
- Multi-exposure scenarios: AI under-emphasised that users may need to record multiple substances from a single exposure event without restarting
- Prototype-fidelity blind spot: AI misread prototype limitations (e.g., an unbuilt "specify" path) as genuine usability failures
- Implicit bias: AI generated male personas even when all participants were female โ explicit demographic constraints and bias checklists are required
- Compound edge cases require explicit prompting and verification against raw notes
Five rules for trusting AI exactly as far as it deserves
Based on the pilot, I developed five practical guardrails for responsible AI integration in UX research:
- Include a "prototype limitations" section in every AI prompt so the model distinguishes real usability issues from prototype gaps
- Constrain persona generation: specify participant attributes explicitly (including demographics) and require evidence links for every insight
- Run a bias checklist on all generated artifacts before sharing โ demographics, stereotypes, and unstated defaults
- Require traceability: every theme must link back to a supporting quote and session reference
- Label all outputs as "AI-assisted draft" and mandate human review before use in decision-making
AI is most valuable as a draft and clustering accelerator. The analyst's job shifts from transcription to validation and critical review โ which is where human judgment has the highest leverage.
The registry is live for Ontario workers today
The Ontario Occupational Exposure Registry is publicly accessible on the Ontario government website. Research findings from this study informed the usability, accessibility, and guidance improvements to the registry's self-tracker.
View the live OER โEfficiency and rigor are not in opposition
Government UX research operates under constraints that private-sector work rarely faces: regulatory language requirements, multi-stakeholder sign-off, strict accessibility mandates, and the reality that errors in public services have real consequences for real people. Staying rigorous under those constraints โ especially when AI tools are accelerating parts of the workflow โ required constant critical review.
The AI integration pilot taught me that efficiency and rigor are not in opposition: AI at ~70% time savings is only valuable if the human review that follows is systematic. The most important skill wasn't knowing which tools to use โ it was knowing exactly where to trust them and exactly where not to.