By Ryan McGarry, UX Research Manager, Akraya | Published March 2026
|
There’s an interesting paradox at the center of modern UX research on AI: we spend our days studying how people interact with AI systems, investigating where they trust the output and where they don’t, where the mental model breaks down, and where the feature delivers. Yet, when we return to our desks, many researchers still rely on manual, 2019-era workflows to synthesize data without using the very tools we test.
This blog post is about closing that gap. It’s about how UX researchers, especially those working in AI-heavy product environments such as enterprise software and developer tools, can meaningfully integrate AI into their research workflows. It’s also about the pitfalls, the policy boundaries, and the kinds of AI-assisted workflows that save time versus the ones that create more problems than they solve.
Everything here comes from our team’s lived experience at Akraya, supporting Google’s suite of developer and cloud products, and from conversations with UX researchers across teams who are navigating this same frontier.
A typical end-to-end UX research project involves tasks that are time-consuming, repetitive, and cognitively taxing: synthesizing hours of session recordings, identifying themes across transcripts, writing executive summaries, formatting reports, building recruitment screeners, and drafting discussion guides. Each of these is a place where AI tools can meaningfully reduce friction if used correctly.
The researchers who will lead the field in the coming years aren’t the ones who avoid these tools out of scepticism. They’re the ones who learn to deploy them with precision, understanding both their capabilities and their failure modes. Which, conveniently, is exactly the expertise that UX researchers build when they study these tools in the first place.
1. Transcription and Initial Data Processing
Session recordings are the bedrock of qualitative research, and they’re also one of the most time-intensive things to work with. AI-powered transcription tools have matured to the point where they can reliably produce usable transcripts from recorded sessions, reducing manual transcription time from hours to minutes.
The important caveat: transcripts produced by AI require review. They frequently mishear technical terminology (especially in developer tool studies where participants use domain-specific language), miss speaker attributions, and occasionally mangle quotes in ways that subtly change meaning. Treat AI transcription as a first draft, not a final artifact.
My recommendation is to build a verification step into your workflow: ensure that you spot-check the transcript against the recording, especially for any quote you plan to use directly in a report or presentation. Otherwise, these simple mistakes can lead to tremendous problems with stakeholders reading a study report and walking away with invalid findings based on hallucinated quotes that participants never actually said.
2. Thematic Analysis and Finding Synthesis
This is where things get genuinely exciting and where we’ve seen the most variation in how researchers use AI tools well versus poorly.
The effective approach looks something like this: after reviewing and organizing your session notes, you bring a structured dataset (themes you’ve already started identifying, notable quotes organized by topic, frequency counts of key feedback) into a tool like Gemini or NotebookLM and use it to help you pressure-test your analysis. Ask it: “Are there patterns here I might have missed?” or “Which of these themes seems most strongly supported by the data?”
One of our researchers recently tried a meta-prompting approach that’s worth sharing: she asked Gemini to help her use NotebookLLM more effectively for literature review. She told Gemini about her research goals and asked for optimal prompts she could use within NotebookLLM to structure her output. The result was a set of tailored prompts that helped her extract structured, actionable literature review output far faster than her previous approach. Using one AI tool to optimize your use of another is a genuinely useful technique, and it’s one most researchers haven’t tried yet.
The ineffective approach: dumping an entire raw transcript into an AI tool and asking it to “give me the findings.” This produces something that looks like research but frequently misses context, misinterprets participant tone, and flattens nuance in ways that can corrupt your analysis. AI is a collaborator in synthesis, not a replacement for it.
3. Report Writing and Executive Summaries
Writing research reports is one of the most universally dreaded parts of the job. The finding is clear in your head. Translating it into stakeholder-ready language, at the right level of detail, with the right framing, is tedious.
AI tools can meaningfully accelerate the drafting phase. The technique that works best: write your key findings yourself, even in rough, unpolished form, and then use AI to help you refine language, improve structure, and produce executive summary versions for different audiences. Prompting for a specific output format helps significantly. “Write this as a three-paragraph executive summary for a non-technical product manager” produces far better results than “summarize this.”
Similarly, giving the AI a role helps frame the output. Prompting with “Think like a senior UX researcher presenting findings to a VP-level stakeholder” produces more structured, appropriately hedged output than an unframed prompt. This is the kind of prompt engineering that’s worth experimenting with systematically.
4. Synthetic Participants and Rapid Concept Testing
This is an emerging use case that’s genuinely useful when used thoughtfully and genuinely problematic when misapplied. AI can generate synthetic participant responses, essentially simulated user reactions to a concept or prototype, that can help teams think through a design direction before investing in full participant research.
The appropriate use case: early-stage concept exploration where the goal is to identify obvious gaps or failure modes before you bring in real participants. Synthetic participants can help a team decide between three rough directions before committing to a study protocol.
The inappropriate use case: replacing real participant research, using synthetic feedback as evidence for product decisions, or presenting AI-generated responses as though they represent actual user behavior. They don’t. And in regulated industries or products with significant user safety implications, the risk of synthetic research leading to bad product decisions is material.
For teams with constrained research budgets, synthetic participants can be a way to maintain some research cadence between funded studies. Just be explicit about what they are and what they’re not.
If there’s one meta-skill that underpins all of the above use cases, it’s prompt engineering, the ability to give AI tools instructions that produce useful output. This is not a technical skill in the software engineering sense. It’s a communication skill, and UX researchers are well-positioned to develop it quickly.
The core principles we’ve found most useful are: be specific about the format you want (“give me a numbered list of five findings, each with a one-sentence implication”); assign a role or persona (“think like a senior UX researcher”); give context before the task (“This is qualitative data from a study on a developer tool used by software engineers at large enterprises”); and iterate. The first output is seldom the best one.
We’ve talked to researchers who tried an AI tool once, got mediocre output, and gave up. That’s roughly equivalent to running one usability session and writing off the entire methodology. Prompting is a skill. It gets better with practice.
One area where AI-assisted research workflows are evolving rapidly is video analysis. For UX researchers who regularly work with recorded sessions, especially when studies produce large volumes of footage, the ability to process video through AI tools holds enormous promise.
In our recent workshop, participants shared experimentation with tools like Google AI Studio for video analysis, as well as creative approaches like slicing video into frames for image-based analysis. These techniques are still maturing, and the results vary significantly based on the tool, the quality of the recording, and how the analysis task is framed.
Our recommendation: treat AI video analysis as supplementary, not primary, for now. Use it to help you identify timestamps worth reviewing, surface potential themes to investigate, or process session clips for specific behaviors. Don’t rely on it to produce a final synthesis without researcher review.
As these tools improve and they will improve quickly, researchers who have been experimenting with them will have a significant advantage. The learning curve is real; start now.
None of the above is an invitation to use any AI tool for any task without understanding the rules that govern your context. For researchers working in enterprise environments, especially in organizations with strict data governance frameworks, the constraints are real and consequential.
The core questions to answer before using any AI tool in your research workflow: Is this tool approved for use with the type of data I’m working with? Could the data I’m providing be used to train or improve the model? What are the data retention policies for the inputs I provide? Am I potentially exposing PII, sensitive business information, or proprietary code?
At Google, internal guidelines explicitly govern which AI tools can be used with which types of data. Researchers are expected to be familiar with these guidelines and to review them regularly, because the rules evolve as the tools do. The consequences of accidentally exposing sensitive information to an external AI system can extend well beyond a reprimand; they can constitute policy violations with real professional and organizational impact.
The principle we operate by: if you’re unsure whether a particular use is allowed, assume it isn’t until you’ve verified. Curiosity is great. Compliance is non-negotiable.
We’ll close with a practical framework for how to start building or refining your own AI-assisted research workflow.
The researchers who will define the next decade of the discipline are the ones who can study AI rigorously and use it wisely in equal measure. These aren’t separate skills that happen to coexist in the same job description. They reinforce each other. Knowing how AI systems fail, how they hallucinate, how they introduce variance, and how they can subtly mislead makes you a better consumer of AI-generated research outputs. And using AI tools regularly in your workflow makes you a better researcher of them.
At Akraya, we’ve been building this muscle across our team over the last two years. The workshops, the shared protocols, the hard-won session learnings, all of it is pointing toward the same conclusion. The competitive advantage in UX research isn’t going to come from refusing to engage with AI or from adopting it uncritically. It’s going to come from developing genuine expertise in both dimensions simultaneously.
That work starts now.
Ryan McGarry is the UX Research Manager at Akraya. He and his team support UX research programs for Google’s Cloud and developer tooling products. Interested in how Akraya can support your AI UX research program? Get in touch.