Details sanitized and shared with appropriate discretion. Specific metrics and artifacts available upon request.
Building an AI-Assisted Research Intelligence Practice
Designed and led a methodology-first practice that uses AI as a force multiplier across the research workflow, improving operational efficiency by ~30% without compromising rigor or participant privacy.
Detailed case study available on request
This project contains confidential process details and internal metrics. The full case study — including methodology decisions, process documentation, stakeholder artifacts, and specific outcome data — is available to hiring managers and recruiters.
Summary
Research demand at United was growing faster than a five-person team could absorb. Rather than let AI adoption happen ad hoc — with uneven quality and real privacy risk — I designed and led a methodology-first AI research practice: a clear framework for where AI augments the work, a shared prompt library, governance guardrails, and team enablement. The result was roughly a 30% improvement in operational efficiency while keeping human judgment in control of every insight that reached a stakeholder.
Role: Lead UX Researcher · Timeline: 2024–Present [VERIFY] · Team: 5 researchers · Methods: workflow design, structured prompting (CO-STAR), enablement, governance
Problem & context
Our intake queue was outpacing capacity, and turnaround time was a recurring source of stakeholder friction. At the same time, generative AI tools were becoming widely available, and individual researchers were beginning to experiment with them independently. That created two risks at once: inconsistent quality (prompts and outputs varied widely person to person) and privacy exposure (the temptation to paste participant data into consumer tools). The opportunity was real, but capturing it without undermining research rigor or trust required deliberate design — not a tool mandate.
My role
I owned this end to end as Lead UX Researcher: I defined the practice and its guardrails, selected and matched tools to tasks, built the shared frameworks, and ran the enablement so the whole team could adopt it consistently rather than reinventing it individually.
Approach & decisions
The central decision was to be methodology-first, not tool-first. AI augments the steps where it is both safe and high-leverage — scaffolding research plans, drafting discussion guides, first-pass synthesis, literature and competitive scans, template generation — while the researcher retains ownership of the decisions that require judgment: sampling, method selection, interpretation, and anything touching participant data.
A few specific choices:
- Standardized on the CO-STAR prompting framework (Context, Objective, Style, Tone, Audience, Response) so prompts were consistent, reusable, and shareable across the team — turning prompting from a personal trick into a team capability.
- Matched tools to tasks rather than picking one winner — using Claude, Gemini, Copilot, and Figma AI for what each does best, and weighing tradeoffs openly.
- Built guardrails before scaling: no raw PII or confidential participant data into non-approved tools, and a mandatory human-verification gate before any AI-assisted output became a finding.
- Chose enablement over mandate: a prompt library plus hands-on working sessions, so adoption was driven by usefulness, not policy.
Process
I piloted the approach on my own projects first, codified what worked into a reusable prompt library and CO-STAR templates, then ran multiple working sessions and workshops to train the team. From there we integrated the workflows into our standard ResearchOps process and measured the effect on throughput and turnaround. [PLACEHOLDER: add a sentence on the pilot scope and how many sessions you ran]
Deliverables
- A shared, versioned prompt library organized by research task
- CO-STAR prompt templates for recurring deliverables
- A one-page AI governance and guardrails guide (privacy, verification, approved tools)
- Workshop and enablement materials for the team
- Before/after workflow maps showing where AI was inserted [PLACEHOLDER: sanitized screenshot or diagram]
Impact & outcomes
- ~30% improvement in operational efficiency across AI-augmented workflows.
- The team adopted a consistent, quality-controlled approach rather than fragmented individual experimentation.
- Faster drafting and synthesis fed directly into reduced research turnaround and freed researcher time for higher-order analysis and stakeholder work.
- [VERIFY/PLACEHOLDER: any specific downstream outcome — e.g., number of studies delivered, stakeholder satisfaction, throughput change]
Reflection
The guardrails turned out to be as important as the prompts — adoption only scaled because people trusted that quality and privacy were protected. If I were starting again, I’d formalize a lightweight output-quality evaluation rubric earlier so we could measure AI-assisted vs. fully manual deliverables more rigorously, and I’d watch more deliberately for over-reliance, keeping the human-judgment steps explicitly non-negotiable.