A Practical Guide to AI for UX Research (2026)

While reality hasn’t matched many of the lofty promises made (yet), things are starting to shift. In the past six months both the capabilities of AI models as well as their ability to act on our behalf and interoperate with other tools has exploded. And the UX Research community has taken notice. Last year, UXRConf 2024 had one talk about AI. This year, Research Week dedicated half a day to AI, and many sessions at UXRConf and the Research Leadership Summit centered around AI.

AI won’t take our jobs as Researchers, unless we fail to examine when and how AI can help us do our jobs better, the same way we have with other advancements in technology. This five-part series on AI in Research is designed to help you do that, whether:

Let’s dive in.

(Oh, and make sure to grab our 2026 AI x UXR Tool Comparison Guide so you can follow along with who’s building what.)

We’ve all been there: the dreaded “blank page problem.” Sometimes the hardest part is getting started. When there are a nearly infinite number of things you could be researching or ways you could be approaching the problem, how do you know you’re on the right track?

Instead of staring at a blank doc or pasting a research plan template for the hundredth time (swipe mine below if you like), feed whatever inputs you have (Jira tickets, Product Requirements Docs, design mockups, your research objectives, whatever) into your favorite LLM. Ask it to help you:

Once you have a plan, another place researchers can get stuck is figuring out what questions to ask or how to ask them. Here is another area where AI excels because it enables an interactive experience where you can ask it to rephrase questions, take different perspectives, or elaborate on certain areas.

Try asking it to write five questions for each of your research goals and see what it comes up with. Then ask it to do it again from the perspective of your CEO or a new teammate with little to no background context. Ask it to make questions longer or shorter. If you’re working on a survey, ask it to use different question types (e.g. multiselect, matrix, linear scales). If you’re working on an interview, make sure you balance how many follow-up questions it asks (more on this later).

Again, the goal isn’t to get something to hand off right away. A key tenet of using AI is that you’re still responsible for the quality of your work. Your goal is to get past the tedious part faster and get to something that’s at least good enough to react to.

Asking AI to critique your work might be my favorite use case so far. Similar to asking it to take different perspectives when generating questions, you can ask AI to look for biases, blind spots, or imbalances in your approach. Ask it to think about how different types of users might react to the questions (a la “synthetic users” 🤫).

You can do this at any point in your research process, which is incredibly handy when preparing findings and putting a presentation together. Does your narrative flow smoothly? How might your Sales team react to this compared to Marketing or Product? What might your PM think? Your designer?

Using an AI as a sparring partner is like having an army of tireless (junior) teammates whose sole purpose is to poke holes in your thinking.

Your typical recruiting process looks something like this:

Once you’re comfortable with the quality of the output from the above workflow, you can start to automate the process with a little extra work. Because research tools currently don’t support end-to-end automation, you might have to get your hands a little dirtier than usual in terms of technology. Using platforms like Zapier, n8n, Gumloop, or Make can help you create agents (usually via MCP) to handle all of this for you. The initial outreach, the scheduling email (which, if an AI is handling it, could be completely conversational instead of forcing participants to pick from existing time slots), the reminders, and the reward. All personalized. All automated.

You can even go one step further and have the agent troubleshoot participant interactions, answering FAQs from participants, providing additional instructions in the reminder email (e.g. to be in a quiet place with stable internet), and flagging anything unusual for human review. In an ideal world, your participant management tool can provide a completely prompt-based experience for all of this. Imagine something like:

The UX Researchers Guild recently ran an Intro to AI Moderation session that walks you through the basics of AI moderation. In a nutshell, you give the agent a script (or it generates its own) just like you would give to a researcher, including what to follow up on and how deeply. From there, the rest of the process plays out like it would with a human: there’s outreach to participants, they’re screened in, and they conduct their session. The difference is that it’s an AI moderator instead of a human.

When the company Synthetic Users first started making waves, there was an uproar in the UX community. Despite what their marketing may tell you, you can’t fully replace human users with AI in Research. However, there are some interesting use cases in planning, data gathering, and even knowledge sharing.

It’s crucial to give your LLM enough context about your users (ideally trained on your real data), but you can use synthetic users to strengthen your overall research design and approach. Ask AI to assume the role of different types of users to gauge how they might react to different questions, prompts, or mockups. You can’t use this as a replacement for research, but it’s a great way to prepare yourself for the real thing, especially with complex domains, hard-to-reach audiences, or sensitive topics.

This is possibly the most interesting and novel application of AI I’ve seen come out of the Research community so far. Researchers Jane Leibrock from Anthropic, Jennifer Maples from AirBnB, and Molly Needelman from Google are experimenting with new ways to generate stimuli for research sessions. The old way of working is to have your design team create several static (though ideally clickable) mockups that each participant interacts with, explaining what they would do. This “Wizard of Oz” approach has been the state of the art for approximately 50 years 😳. In 2025 and beyond, we can do better.

Something interesting that’s caught my attention is the divergence in fidelity: some go high, others go low. Instead of the same static mockups for each participant, teams are experimenting with personalized prototypes that make the experience feel more real. You can go a step further and vibe code prototypes, so you can test interactions without the need for engineering or any backend support. Now each session can get into rich detail of how a product or feature might work for their world. Molly’s “Build to Learn” talk at UXRConf went as far as to include co-creation with your users, allowing them to build something best suited for them in-session.

Calling transcription an AI feature at this point feels a bit silly. However, as I learned while writing this very post, there’s a big difference between transcription (which turns what you speak into writing, stammers and all) and summarization (which gets the gist of what you’re trying to say and compiles it into something clear and concise). That difference is crucial for Research, which can easily wind up with 10s of hours of audio and video to parse through.

I’ve been in the industry long enough to remember an era before we recorded sessions at all, relying on human note-takers and transcriptionists. Having AI-powered transcription and summarization baked into every video recording tool is a huge time-saver, and if your insights repository doesn’t have AI features, you can at least take those transcripts and drop them into your LLM of choice to augment your analysis.

Once you have some representation of your raw data (survey responses or interview/usability transcripts) loaded into an LLM, you can start to ask questions of the data using natural language, e.g.

When you’re prepping your analysis, think of AI as an assistant and a sparring partner, similar to how we used AI in the planning phase. Prompt it to challenge your early insights, suggest alternative framings, or identify gaps in evidence. Compare the data you collected to your initial research objectives, using it to frame your narrative, and have AI take the perspective of some of your stakeholders to anticipate how they might react.

Make sure you’re aware of the biases/tendencies of the model you’re using for this, though. If you’re using something like GPT-4o that tends to be overly agreeable, be sure to tell it to be critical.

One huge win is to integrate your insights repository with your company’s AI assistant. If you’re using Glean, Guru, Copilot, or other similar AI knowledge management tools, you can train custom GPTs/agents to respond to queries from anyone at the company. Instead of you or one of your colleagues fielding these questions, your stakeholders can “chat” with the data for questions like:

If you aren’t already creating highlight reels as part of your research practice, you’re missing out. Highlight reels are an excellent way to illustrate the pain points users face in their own words while also avoiding any sampling bias. If a stakeholder watches one interview, they’re more likely to get tunnel vision. If they see a 2-minute clip with 8 different participants all saying similar things, it’s hard to do that.

In an ideal world, your analysis tool can auto-generate these highlight reels for you if you tell it what you’re looking for (e.g. “give me 4 examples of pricing confusion”). But at the very least, using AI during your analysis phase can give you sources with timestamps to vastly decrease the amount of time it takes to produce. AI turns scrubbing through hours of video into one short prompt and a few video edits (or, if you’re lucky, a click of a button 🪄).

With Great Question AI, video highlights (along with summaries, chapters, and tags) are automatically generated after every interview you complete.

Taking the “chat with your data” part to the next level, some companies are creating their own custom “synthetic personas”. If chatting with your data is about bringing insights to life, synthetic personas are about breathing new life into stodgy old personas. By building archetypes on real research data, your stakeholders can run scenarios past your existing data to anticipate how customers might react.

You might be thinking that this sounds a lot like synthetic users from earlier, and while you’re not completely wrong (both use AI to create a replica of real users), the intended use case is different. When using synthetic users, you’re trying to stress-test your research methodology. Interactive personas are intended to be an easier way for your stakeholders (or other colleagues) to interact with your findings easier and more naturally. Once you have interactive personas, you can try feeding your research plans back into them; however, this creates a risk of generation loss and is not recommended.

There is an economic argument hiding in here too. As more of the research pipeline gets automated, the quality of the remaining human judgment matters more, not less. Economists call this O-Ring theory: one weak link breaks the chain. The researchers who thrive will not be the ones who resist AI or the ones who trust it blindly. They will be the ones who direct it well.

You don’t need a PhD or permission. Think of chat as a way to iterate on your prompts, but ultimately you should work toward productizing the workflows that fit your needs. When productizing AI, there are two key things to think about:

You can use Kaleb Loosbrock’s CRAFTe framework as a great way to structure your experiments. At first it will feel like you’re sinking a lot of time into this, but similar to how Research Operations is a flywheel, eventually you’ll start to build momentum, and the long-term time-savings are unmatched.