A decision framework that helps newer UX researchers know how much AI to use, and when.
Role
UX Researcher & Designer
Timeline
Spring 2026, 14 weeks
Methods
20 interviews, thematic analysis, concept testing, prototyping
Tools
Figma, HTML/CSS prototype
Newer UX researchers are entering a profession where AI is increasing expected, but rarely scaffolded. This project explored the moment before prompting when a researcher is choosing intent, stakes, platform, framing, and language all at once.
Using AI is easy. Using it well takes judgement.
CHALLENGE
Speed without distance
Researchers are expected to move faster without sacrificing the nuance, judgment, and human connection that make research valuable.
AUDIENCE
Newer UX researchers
Researchers who are expected to use AI before they have built the years of pattern recognition senior practitioners rely on.
OUTCOME
Decision framework
A platform-agnostic tool that helps define the task, select a prompt scaffold, add context, and bring the final prompt into any approved AI platform.
AI is now embedded in nearly every stage of UX research, not because researchers necessarily sought it out, but because the industry is rapidly adopting it. Researchers are expected to move faster while maintaining the same level of rigor, forcing them to balance efficiency with confidence in their findings. Whether they are new to the field or experienced practitioners, they are being asked to rely on AI tools that can summarize, synthesize, and generate insights, often without transparency into how those outputs were produced. This creates a growing gap between speed and human judgment.
If AI accelerates research processes, how might we help researchers use it more intentionally?
The project started with an open-ended question: How are UX researchers actually using AI day to day? I conducted twenty 45 to 75-minute semi-structured interviews with researchers across startups, SaaS companies, enterprises, banking, insurance, and other regulated contexts. Questions focused on workflows, trust, hesitation, company pressure, and moments where AI supported or disrupted the work.
Reported feeling external pressure from their company to incorporate AI into their workflow.
Described prompting as one of the biggest barriers to getting useful AI output.
Developed personal workarounds to reconnect with the original research data.
Studying behavior around AI, not just the tools.
I began by mapping which tools researchers used and where those tools appeared in the research process. But the same platforms and concerns kept appearing across companies. The more meaningful differences were in what researchers did when they did not trust the output.
Researchers were re-transcribing quotes by hand, rebuilding charts AI had already generated, returning to source transcripts, and writing "do not hallucinate" into their prompts. These workarounds were creating safeguards to protect the accuracy and integrity of their work.
Every researcher I interviewed had developed a workaround to reconnect with the data
Researchers were most hesitant when AI summarized qualitative data, attributed quotes, or generated insights without showing how it reached them. Their response was to verify transcripts, request timestamps, and limit how much data they uploaded at once.
Some researchers were evaluated on AI literacy. Others could only use one approved platform because of security and compliance restrictions. The environments were different, but researchers in both were creating safeguards to protect their work.
One researcher described confidence as a curve. It began low while they learned the prompt, rose once the output became useful, and fell again as the conversation grew longer and the model began losing earlier context.
The project went through four real pivots: from cataloging tools, to studying behavior, to realizing that company context drove AI use more than individual preference did. A senior researcher at a company with strict compliance rules and a researcher graded on "AI literacy" at a fast-moving company were doing the same thing for opposite reasons: protecting the integrity of their work.
What I learned by narrowing
The project became stronger as the research removed possibilities. What began as an investigation into AI tools shifted toward the behavior surrounding them. Company pressure explained why AI entered the workflow, but the workarounds revealed the deeper concern: researchers were trying to move faster without losing their understanding of the data. The final opportunity became supporting judgment before AI begins shaping the work.
Think First. Refine Later. helps newer researchers decide how much AI integration is appropriate for a task, then guides them through creating a structured, context-grounded prompt that can travel into any AI platform.
A framework that lives in the prompting moment
01 · Quadrant
A simple 2x2 (open vs. closed task, low vs. high stakes) sorts any task into one of four modes: Discover, Pressure Testing, Iterate, or Polish. Naming the mode sets expectations for what AI should (and shouldn't) do.
02 · Library
Newer researchers found a blank prompt box as intimidating as a blank document. The Library step offers a scaffolded starting prompt matched to the chosen quadrant, cutting straight through the prompting struggle 87% of participants named.
03 · Customize
The researcher fills in role, project context, and momentary context (what's true right now, not just in general). A Context Bank surfaces personal and team notes so hard-won context doesn't have to be retyped from memory every time.
04 · Copy
The finished prompt is copied into whatever AI platform the researcher's company allows. Nothing is sent to an AI model from inside the tool itself, and nothing is locked to one platform, a deliberate choice, since every participant's tool stack looked different and none of them were staying still.
It's platform-agnostic. No tool named in the interviews was safe from being replaced within a year. A framework tied to a specific AI would have an expiration date.
It works inside compliance constraints, because it produces a prompt rather than sending data anywhere, it functions even in banking and insurance environments where AI use is tightly restricted.
In concept testing, two things landed immediately: the quadrant vocabulary (several researchers said it named a distinction they were already making but had no word for) and the Context Bank, valued as a way to keep team knowledge present without rewriting it every time.
The biggest shift was learning that research is about narrowing possibilities until the right problem becomes clear
I spent the first half of the project collecting: more themes, more participants, more possible directions. The work got sharper the moment I let the data cut things: three directions became one, "researchers" became "newer researchers," a tool that does AI work for people became a tool that helps them decide what AI work is actually needed.
None of the 20 researchers I talked to called what they were doing a "workaround." They called it "what I just do." Making that implicit behavior explicit, and designing around it instead of around the tools, was the actual design move.
Letting the audience narrow
Prototype and bring literature in earlier
Concept validation is not behavioral validation
Meet researchers inside existing workflows