AI products are being roadmapped with dazzling demos and bold claims. Specifically, MIT finds that 95% of these AI pilots fail. Worse, we’re seeing a resurgence of a tech-led approach to AI, which, as an approach, has a history of failure. See Red Flag: No AI without UX

Yes, the tech is important in AI, and the experience is largely unseen. But, UX is still often an afterthought in AI development. Why? Because good UX is often invisible. And AI teams, focused on models and accuracy, can forget that users don’t interact with models: they interact with an experience.

In AI projects globally, something critical is missing: thoughtful, intentional UX.

Why this matters: Gaining user trust with AI experience is key. Trust drops when systems behave unpredictably, or unexpectedly. People abandon tools that feel confusing or cold.

The problem: AI-first, user-last

Most AI projects start with: “We have AI on our roadmap and a language model. At some point, we’ll wrap an interface around it.”

This tired thinking treats UX as decoration, not a foundation.

Specifically, it ignores:

• Friction in real-world use, such as barriers to constructing Prompts 

• Cognitive load from unclear AI behavior

• Mistrust of opaque systems

• Frustration when users can’t recover from errors

AI UX fails when the human experience is layered on after the tech is built. Seems familiar, right? That’s because when ‘AI as a feature’ drives, practices like inclusive user research, which aim to reduce harm and increase safety, are often afterthoughts. Some risky product teams, in the past few years have even (Meta, Grok, Google) have even launched and then rolled back: The public shame AI launch strategy.

Why UX gets sidelined

1) It’s invisible when it works well
Seamless AI UX hides complexity. But that makes it harder to justify up front…until users complain.

In May 2025, TechSpot reported that 50% of companies like Air Canada, H&M, and Klarna are rolling back their efforts to replace human customer service with AI-only solutions due to poor customer satisfaction with bots.

2) AI teams prioritize model performance
Don’t get me wrong, precision, recall, and latency are critical. But what about understanding, clarity, and recoverability? We need greater focus on ‘alignment‘, namely, focusing on user values and goals.

Learn from Microsoft’s anti-pattern: Privacy or Privacy by Design (an established set of UX principles) is a user value Microsoft missed with its Recall feature (AI desktop search). Dubbed a ‘privacy nightmare‘, the feature seems a great solution to desktop search. However, taking saved screenshots every few seconds caused immediate concern. It seems user control over what’s being snapshotted was not factored in. For example, off the top of my head: Recall could allow users to exclude sensitive applications like email,  or offer a ‘redacted’ mode (encrypted screenshot mode requiring a passkey to ‘view’ the screenshot later)…or creating a privacy bot that manages Recall, with strict instructions to not include this or that (trained with that users guardrails, and GDPR).

3) UX isn’t baked into training data
The AI data pipeline is rarely informed by lived experience. Stress cases? Accessibility? Emotional tone? Cultural context? Yep, they are often overlooked.

4) No clear owner for AI behavior design
Who decides how the AI communicates uncertainty? Or how to display confidence levels? These questions fall through the cracks. Uncertainty is key to understanding AI hallucinations, according to a new paper by OpenAI researchers.

The fix: Bake UX in, to de-risk tech gaps

1. Design from the outside-in
   Start with the user context and needs. Map the emotional journey. Then decide if and how AI should help. Bring in underrepresented users up front to socialise lived experience issues with your AI team.

2. Prototype UX before models
   Use Wizard-of-Oz t(WoZ) tests. Role-play and improvise AI interactions. Note- there are a lot of new AI tools that allow you to use AI to prototype AI experiences. Without fully intending to, I built two agents in a couple of hours last week, which allowed me to evaluate parameters of an AI-assisted experience. These WoZ tests can help you prepare for “evals” (rigorous AI quality evaluations that “are a critical step in going from prototype to production with generative AI.”). WoZ prototyping lets you see what users expect, need, and fear even before writing a single line of code.

   Want to play in a live applied learning lab? Join me for Redesign it with AI, a regular series where we redesign popular apps with an AI lens (a perk of UX Inner Circle membership).

3. Make AI behavior visible
   Explain what the AI is doing. Show confidence levels. Let users inspect and correct outputs. This is especially critical as we define Agentic UX.

4. Design for edge cases and failure
   AI is probabilistic, not perfect. Design for what happens when it guesses wrong. Build graceful fallbacks. Prioritize inclusion, harm reduction and safety.

5. Give users control
   Let them undo, edit, opt out, and escalate. Choice = trust. Again, critical to designing Agentic UX. Why? Because a tech-led mindset means ‘take control away from users’: Trust in Automation.

Final thought: Good UX makes AI usable. Great UX makes AI trusted.

When UX is invisible, it’s often because it was done well. But when it’s absent, the user feels it in an instant.

We need to stop treating UX as a coat of paint. For AI to work in the wild, UX must shape the foundation. No more afterthoughts. Design with users mitigating risks, from the start.

Learn more: Join my (FREE) Vision & Values conversation on Nov 6th re. AI, product and UX with Christian Crumlish

Learn by doing: join me quarterly event, coming up on Oct 22nd ($99 month membership perk) for Redesign it with AI (Duolingo wrap up session)