AI & UX

AI gives companies scale. User research gives it direction. Most teams only invest in one.

Key Takeaways

AI gives companies data, automation, and personalisation at scale. But without UX and user research, these capabilities rarely translate into something people actually want to use. Four problems come up again and again – and they’re all fixable.

There’s a pattern I’ve been watching play out across organisations over the last few years.

A company invests heavily in AI. Impressive things get built – personalisation engines, recommendation systems, predictive models. The demos look great. Then it ships. Adoption is lower than expected. Users don’t engage with it the way anyone hoped. The recommendations feel slightly off. Something’s not working, but it’s hard to say exactly what.

In almost every case, the gap is the same: the team understood what AI can do, but not what their users actually need.

The technology worked. The product didn’t.

What AI Enables – and What It Can’t Do Alone

AI gives organisations four real levers. Data – learning at scale from behaviour you could never manually analyse. Automation – removing the manual effort from tasks that used to require a human every time. Personalisation – adapting to individuals rather than assuming everyone wants the same thing. And prediction – anticipating what users are likely to do next.

None of those guarantee a product that’s useful, trusted, or adopted.

That’s where UX and user research become critical. Not as an optional extra layer, but as the thing that tells you whether you’re optimising for the right thing in the first place.

AI systems optimise for proxies. Click-through rates. Completion rates. Signals that correlate with what you care about, but aren’t the same thing. User research tells you what the user is actually trying to do – and whether the AI’s proxy is tracking that accurately. When these two things are misaligned, you get a technically impressive system that subtly makes the experience worse.

UX as the Missing Layer in Unified AI

When AI is deployed across an organisation without UX research embedded in the process, it optimises for the wrong things. The framework below shows what each side brings – and what customers actually get when both work together.

The left column is achievable without UX research. Data, automation, personalisation, and prediction can all be built on technical capability alone. The middle column – real user problem focus, early validation, trust and usability – is the part most AI projects skip. And without it, the outcomes in the right column – adoption, relevance, and not building the wrong thing – don’t follow.

Four Problems That Keep Coming Up

The same four issues appear across almost every organisation trying to integrate AI and UX.

Research findings that disappear. What users misunderstand, what context they’re missing, what emotional state drives their decisions – this sits in a research report the model team never reads. No handoff process means the AI keeps optimising for things research already knows don’t reflect real user goals.

No shared dashboard. Marketing, product, and data science each look at a different slice. Nobody has a view that connects them. Fragmented data leads to fragmented decisions, regardless of how good the individual sources are.

Confirmation bias in prompting. Most people prompt AI tools in a way that confirms what they already believe. The output feels like validation – but it’s their own assumption reflected back in slightly different words. Without actively challenging rather than confirming, AI becomes an expensive way to stay wrong.

Governance gaps. AI gets deployed in parallel without coordination. Marketing builds a chatbot. Product builds personalisation. Customer service adds automation. The user experiences three different AI interactions in one session, each with a different tone and logic. That’s not innovation – it’s fragmentation dressed up as progress.

Data Harmonisation Across Departments

AI in UX has a precondition most organisations aren’t ready for: the data needs to be consistent before AI can do anything useful with it.

Marketing tracks different metrics than product. Customer service uses different tools than analytics. Everyone has their own version of the user – and those versions often contradict each other. When AI is layered on top of fragmented data, it doesn’t resolve the contradictions. It amplifies them.

Data harmonisation means agreeing on shared definitions across departments: what counts as a conversion, how sessions are measured, how qualitative and quantitative signals connect. Without that foundation, AI models train on noise and produce confident-sounding results that don’t hold up.

The direction matters both ways. UX insights should flow into the data infrastructure – personas, process maps, and usability findings should shape how data is collected, not just how it’s read. And the data should flow back into UX, giving designers access to behavioural signals no usability test alone can surface.

AI doesn’t fix data silos. It makes them more expensive to ignore.

What Good Looks Like

When AI and UX are genuinely integrated, four things are true.

Research findings have a clear path into the systems being built – not as documentation on a shared drive, but as inputs that actually change what the model optimises for. Teams share a view of the full customer journey, not just their own slice. AI features are tested with real users before full deployment – not just for accuracy, but for whether they’re useful and trusted. And someone owns governance: how AI features interact across the product, so three teams don’t accidentally build one terrible experience.

The technology is not the hard part. The hard part is deciding what it should actually be for.

The Scale of What’s Already Here

One thing the “AI is coming” framing misses: it’s not coming. It’s here.

According to Maze’s 2026 Future of User Research report, 78% of UX teams now use AI in some part of their research or design workflow. That’s not early adoption – that’s the majority. The organisations still treating AI as a future consideration are already behind the curve.

In practice practically: the interesting research questions have shifted. They’re no longer “should we use AI here?” They’re “how do users understand what the AI is doing?”, “when does the AI’s confidence make users over-trust it?”, and “how do we design for AI failure?” – because AI systems fail, and the failure experience is often worse than if the feature hadn’t existed at all.

Two methods gaining specific traction in 2026: Predictive Personas – where behavioural data from AI systems feeds into dynamically updated user models rather than static documents; and Continuous Discovery – where AI-assisted analysis of ongoing user interactions replaces the quarterly research cycle with a near-real-time feedback loop. Neither replaces structured research. Both make it faster and more connected to what’s actually happening in the product.

A third method is gaining attention – and more scepticism: AI Synthetic Personas, where language models generate user archetypes from existing behavioural data, research repositories, or product analytics rather than from direct user contact. The appeal is speed: a team can generate testable assumptions about user needs in hours rather than weeks. The risk is significant. Synthetic personas are only as reliable as the data they’re built on, tend to produce overly optimistic outputs, and are prone to generating plausible-sounding but unverifiable user opinions. Nielsen Norman Group’s position is clear: synthetic users are useful for generating better questions and focusing research – not for replacing it. Used that way, they’re a legitimate early-stage tool. Used as a shortcut past real users, they produce confident-looking research that tells you what you already believed.

Shadow AI: The Problem Nobody Talks About Openly

Most companies roll out a basic AI solution and call it done. Employees use it for simple tasks – and then quietly switch to their personal ChatGPT Plus or Claude Pro subscription for the work that actually matters. 78% of employees who use AI bring their own tools, and 68% do so without IT approval.

The problem isn’t the workaround. It’s what goes into it. 48% of employees have uploaded sensitive company or customer data into personal AI tools – for UX teams that means interview transcripts, session notes, and persona documents built on proprietary research, fed into accounts with different data retention and training policies than your organisation agreed to.

Only 37% of organisations have a governance policy covering AI tool usage. The answer isn’t to ban tools – it’s to close the capability gap between what’s provided and what people actually need, and to set clear guidelines on what data can and can’t go into an AI workflow.

The gap between the tools companies provide and the ones employees actually need is where Shadow AI lives. Closing it is a product decision, not a policy one.

When Stakeholders Use AI to Overrule Design

Shadow AI does not only move upward through the organisation. It also moves sideways – into design reviews.

A pattern that is becoming more common: a stakeholder feeds a UX concept into ChatGPT without the research behind it, without the constraints that shaped it, without the alternatives already ruled out – and arrives at a design review with “the AI also flagged this.” It looks like a second opinion. It functions as a veto.

This is the AI-washed HiPPO: the Highest Paid Person’s Opinion, now dressed in the language of data. What changes is that it becomes harder to push back on. Gut feel can be countered with research. “The AI said so” feels objective – even when the AI had none of the context that informed the original decision.

The asymmetry is worth naming. A UX designer would not prompt an AI to critique a marketing campaign and present the output as grounds to override the team’s judgement. Design receives this treatment precisely because its value is still seen as subjective in many organisations – and AI fills that vacuum with confident generality.

What to Do Next

1. Map where AI is already in use in your organisation – including informal use by individual teams. Understanding the current picture is the prerequisite for any governance conversation.
2. Identify one place where a research insight could improve an existing AI feature. What does your team know about users that your models don’t yet reflect?
3. Start a shared feedback loop. Even a simple channel where research findings, support signals, and product analytics are visible together is more valuable than keeping them in separate silos.

Sources & Further Reading

These resources directly informed the thinking in this article – all focused on AI in the context of UX, product design, and organisational practice:

• Google PAIR. People + AI Guidebook. Google’s own framework for designing AI-powered products with users in mind. Covers user needs, mental models, feedback loops, and when AI is the right solution. Essential reading for any team building AI features.
• Google PAIR. User Needs + Defining Success. The chapter on why starting with user needs – not available data or model capabilities – is the foundation of AI products that actually deliver value.
• Nielsen Norman Group. AI Hallucinations: What Designers Need to Know. Directly relevant to the problem of forwarding AI output uncritically. Covers how hallucinations occur, what they look like in practice, and what designers and teams can do about them.
• Nielsen Norman Group. How AI Literacy Shapes GenAI Use. Research on how people’s understanding of AI affects how well they use it – and why confirmatory prompting is so common among teams with low AI literacy.
• Nielsen Norman Group. Generative UI and Outcome-Oriented Design. On how AI-generated interfaces change the design challenge and why user outcomes – not model outputs – need to be the measure of success.
• Nielsen Norman Group. The User Experience of Chatbots. A foundational look at how users experience AI-driven conversational interfaces, including trust, expectation, and failure modes.
• Interaction Design Foundation. Artificial Intelligence in UX Design. IxDF. A comprehensive topic hub covering AI’s impact on design practice, research methods, and product strategy.
• User Interviews. 30+ AI Tools for Every Phase of UX Research. A practical overview of how AI tools are being used across the research workflow – relevant to the section on AI as a UX tool in daily practice.
• Cybernews. The Rise of Shadow AI in the Workplace. Overview of how employees are bringing personal AI subscriptions into work – and the data privacy risks this creates.
• Lasso Security. What is Shadow AI? Risks and Best Practices. Covers the 48% stat on sensitive data uploads and governance gaps organisations face in 2026.
• Nielsen Norman Group. Synthetic Users: If, When, and How to Use AI-Generated Research. NNg’s clear-eyed assessment of where synthetic personas help and where they mislead – the source for the position cited in this article.
• ACM Interactions. The Synthetic Persona Fallacy. Why AI-generated research can undermine UX research when used as a substitute rather than a starting point.

René Manikofski is a Senior UX Designer with 10+ years of experience in e-commerce and digital product design across Europe. All articles are based on personal professional experience and supported by AI in writing.