AI UX research for AI products people adopt

Ask a product team why an AI feature failed to catch on and the answer usually points to the model. The truer answer is that no one studied what a real person would do the first time the system answered with confidence and got it wrong. Lazarev.agency has shipped enough AI products to see the pattern: the constraint is the distance between what the model can do and what a user expects it to do, and AI UX research is how that distance gets measured and closed.

This is a working guide to how we approach AI UX research at Lazarev.agency: the methods, the sequence, and seven projects in which research determined the outcome. The through-line is simple. When you are designing something probabilistic, you cannot design your way to trust from the whiteboard. You have to go find it in the data and in the people.

Key takeaways

• Research for AI products tests trust on top of usability. The question goes from "can they finish the task" to "can they tell when the AI is wrong, correct it, and keep going."
• The best returns come from a product discovery process that puts research before the build. Every case below where research led the engagement produced a fundable, adoptable product. Several raised seven and eight figures afterward.
• Continuous research beats one-time research for AI. Model behavior drifts, so instrumentation and analytics keep the research live after launch.

What AI UX research means

AI UX research is the study of how people understand and act on a system whose output is probabilistic. Classic UX research asks whether an interface is usable. AI UX research keeps that question and adds harder ones: does the person know what the system can and cannot do, do they calibrate their trust correctly, and can they recover gracefully when the model is confidently mistaken? 

Side note: Google's People + AI Guidebook and Microsoft's Human-AI Interaction (HAX) guidelines both organize around these exact problems.

The distinction matters because it changes what you test and when. The table below captures the core difference between UX research for a conventional product and AI in UX research.

Keep that right-hand column in mind. It is the reason a chatbot with solid demos can still die in the hands of real users, and the reason research earns its place at the front of an AI project.

The numbers behind using AI in UX

Two forces are pushing AI into the research process at once: AI has become genuinely useful for knowledge work, and it has become cheap enough to use at scale. The verified figures below frame both the opportunity and the caution.

• AI lifted knowledge-worker throughput by 66% across three controlled studies, according to Nielsen Norman Group's 2023 analysis of generative AI and productivity.
  Takeaway: It’s the practical case for AI-assisted synthesis and analysis in research work. 
• 88% of organizations report regular AI use in at least one business function in 2025, up from 78% a year earlier, per McKinsey's State of AI.
  Takeaway: AI is now the baseline inside product and design orgs.
• The cost to run a model at GPT-3.5-level quality fell more than 280-fold between late 2022 and late 2024, from about $20 to $0.07 per million tokens, according to the Stanford HAI 2025 AI Index.
  Takeaway: This shift is why AI tooling for research is viable at scale.
• The generative AI market is projected to surpass $324 billion by 2033, per Grand View Research.
  Takeaway: Read it as the commercial tailwind behind AI-for-design tooling; it sizes the broad market, and no clean AI-UX-research figure exists yet.
• Gartner predicts that by 2027, 60% of data and analytics leaders will hit critical failures managing synthetic data, risking governance, accuracy, and compliance.
  Takeaway: AI-generated research needs governance, the same way any synthetic data does.

The pattern across these sources is consistent. Use AI to move faster through the mechanical parts of research, and keep real humans in the loop for the parts that decide whether an AI product earns trust.

The AI UX research playbook in one view

Here is how we sequence research across an AI product engagement, with the question each practice answers and the project where it carried the work.

The rest of this guide walks through each practice with the project that proves it out.

1. Generative research at scale finds the segments before the UI

Generative research is the practice of studying real people broadly enough to discover who your users actually are before you design a single screen. It is the least glamorous and most decisive part of AI UX research, because everything downstream inherits its assumptions.

Practical insight: On Dollet Wallet, a Web3 finance app serving both first-time holders and professional traders, the product kept failing to speak to either group. We ran global UX research with 1,600 participants across three distinct user groups before touching the interface. That scale let us see that "crypto user" was three different people with three different tolerances for complexity, and design one crypto wallet app that scaled from newcomer to pro without dumbing down the professional tools. The research produced the segmentation; the segmentation produced a single scalable product.

Lesson #1: The same logic governs AI products. Before you can design an AI product users understand, you need to know whether your user is an expert who wants terse output or a novice who needs the model to show its work. Research answers that. 

2. A three-pillar research strategy tells you whether to build at all

A three-pillar research strategy pairs market research, user research, and problem validation so a team learns whether a product should exist before it commits to how it will look. For AI products, where enthusiasm often outruns evidence, this discipline is the difference between an AI product roadmap and a wish list.

Practical insight: HiTA, an AI learning platform that gives students subject-aware hints rather than direct answers, had real pedagogical value and unclear direction. Instead of jumping to a redesign, we built a research framework on three pillars: market research to map the category, user research to understand students and educators, and problem validation to confirm the pain was real. 

The research produced a focused product roadmap and a landing page that reads as a strategic sales asset, positioning an AI tool around outcomes its audience cares about, teachers cutting administrative load and students working more efficiently. The research did not just inform the design. It set the product's direction.

Lesson #2: This is what "how to use AI for design" looks like at the strategy layer. The AI is the subject of the research, and the research keeps the AI honest about the value it claims to deliver.

3. Needs research pushes an AI product past the chatbot default

Needs research studies the actual tasks and workflows of a user so an AI product can do something more useful than answer questions in a chat box. Left unresearched, almost every AI product collapses into the same default: a text field and a hopeful prompt. 

Practical insight: Accern, a leading U.S. NLP company, asked us to design Rhea, an AI research tool for financial analysts, VC investors, and ESG specialists. The brief was explicit about going beyond chatbot behavior, and the only way to honor that was to study how analysts actually build a piece of research. That work produced a hybrid interface that pairs prompt-driven AI with dynamic widgets in a split-screen research-to-report workflow, a Report Creator mode, and adaptive questioning that asks clarifying questions the way a good analyst would. 

Rhea helped propel Accern from Series B to an eight-figure acquisition, with $40M+ raised across the partnership. The AI UX patterns it introduced are now visible in other AI products. None of that comes from a chat window. It comes from understanding the job.

Lesson #3: If your AI UX research stops at "users want to ask it things," you will ship the default. If it maps the real workflow, you get AI-native product design.

4. Evaluative research and UX audits find where live AI loses people

Evaluative research is the study of an existing product to locate exactly where it fails users, usually through a heuristic UX audit plus usability testing. For AI products already in market, this is the fastest path to a fundable turnaround, because the problems are real and observable.

Practical insight #1: Fieldstream, a B2B platform that uses an AI-driven algorithm to guide marketing budget decisions, had a solid backend and a product layer that got in the way. Non-technical marketers could not move from analysis to planning, so they exported everything and worked outside the tool. We ran a UX audit, found where the flow broke, and rebuilt the dashboard into a structured decision-making surface that makes the AI's value visible, with a three-stage planning flow. After launch, customers reported an average 30% increase in marketing ROI, and Fieldstream secured $3.8M in pre-seed funding.

Practical insight #2: Suits.ai tells the same story in the AI-agent space. Enterprise customers were signing on, but a long, unclear activation path was costing deals. Research reshaped the journey into a guided onboarding sequence and a single assistant dashboard. The company raised $1M after the MVP launch. Evaluative research does not need a grand hypothesis. It needs an honest look at where people give up.

5. Behavioral analytics turn research into a continuous signal

Behavioral analytics is the practice of instrumenting a live product so research continues after launch, reading what users actually do rather than what they say they will do. AI products need this more than most, because model behavior and user expectations both move.

Practical insight #1: On Bluwalk, a gig-economy platform, we stood up a data-driven growth program on Amplitude so every design change was measured against activation, engagement, and retention. Monthly optimization cycles turned research into a habit, with gains readable within weeks of each change. 

Practical insight #2: The same instinct shaped VTnews.ai, an AI news aggregator built with Patrick Bet-David to expose media bias. There, we designed a reader-analytics dashboard that shows users the political orientation of what they consume, then studied behavior against it. The platform onboarded 85,000 users in its first month, and 90% of them confirmed it helped them escape information bubbles. Analytics closed the loop between what the AI surfaced and how people responded to it.

How to use AI for UX research without losing rigor

Using AI for UX research pays off most in the repetitive middle of the process, the stretch between gathering the data and interpreting it, where speed matters and judgment has not come into play yet. Our rule is simple: AI takes the first pass, and a researcher makes the decision. The table below shows how we divide the work.

‍Where rigor has to hold. Language models are agreeable by design, which is why NN/g found synthetic users overstating task success, and why Gartner expects most organizations to stumble on synthetic-data governance by 2027. Treat AI-generated research as a draft a human validates, the same way you would treat any synthetic data. 

The Shape of AI pattern library is a useful companion here, since it catalogs the interaction patterns behind AI transparency and explainability, like confidence indicators and graceful failure, that your research should be probing in the first place.

Where it is heading next. Using AI for UX design is creeping into the making side of the craft. UX prototype AI tools can turn a research insight into a clickable flow in minutes, which shortens the loop between what you learn and what you test. The speed is real, and it is also where teams get careless, shipping a polished prototype built on an unvalidated assumption. The prototype is only as good as the research under it.

Make research your competitive moat

“The future of UX design with AI will not reward teams with the best models. Every serious player has access to comparable models. It will reward teams that understand their users well enough to shape those models into something people trust. Research is the part competitors cannot copy off your marketing site, which is exactly why it is worth building into your process deliberately.” 

{{Kirill Lazarev}}

A few practices our design leads hold to on AI engagements:

1. Sequence research before the prototype. The cheapest time to learn your assistant confuses people is before you have built it.
2. Study failure as closely as success. How a person recovers when the AI is wrong tells you more than ten clean demos.
3. Instrument for drift. Set up analytics so post-launch behavior keeps teaching you, because AI products change under your feet.
4. Let AI do the synthesis, never the deciding. Speed on the mechanical parts, human judgment on the conclusions.
5. Design for calibrated trust. The goal is a user who trusts the system exactly as much as it deserves, no more.

Get those right, and research becomes the reason your AI product is the one people keep.

Work with a team that researches AI products for a living

Lazarev.agency designs AI products where research leads and outcomes follow, from Rhea's eight-figure exit to Fieldstream's $3.8M raise. If you are building something AI-native and want the research done before the guesswork sets in, start a conversation with our team. We will help you find the trust before you build the interface.