Designing for AI is not UX as Usual

Why general best practices don’t translate, nuance does.

For most of my career, “best practices” have been a comfort.

They gave teams shared language, helped justify decisions, and created a sense of craft continuity. You could walk into a room, say “consistency,” “clarity,” or “ease of use,” and everyone more or less knew what you meant.

That contract quietly breaks when AI enters the picture.

Not because UX no longer matters — but because the assumptions behind traditional UX stop holding. Designing for AI is not just product design with a new interaction model. It’s a different category of problem entirely, and pretending otherwise leads to experiences that feel polished on the surface but unstable underneath.

Best practices assume stable systems

Most UX best practices were forged in a world where systems behaved predictably. You clicked a button, something happened, and that behavior didn’t change tomorrow. Errors were identifiable. Success states were clear. Learn something once and you could apply it everywhere.

These practices evolved in that controlled environment.

AI systems don’t live in that world.

They are probabilistic, context-sensitive, and often opaque. The same input can yield different outputs. The same output can be useful in one situation and misleading in another. There often isn’t a single “correct” answer, but rather a range of plausible ones.

When we apply traditional UX heuristics without adjusting for that reality, we end up solving the wrong problem.

Usability vs sensemaking

Teams often default to familiar questions when designing AI experiences:

Is it intuitive?
Is it simple?
Is it easy to use?

These aren’t bad questions. But they’re incomplete.

The friction users experience with AI rarely comes from interaction mechanics. It comes from interpretation. From deciding whether an output is good enough, trustworthy enough, or appropriate enough for the task at hand.

AI doesn’t just ask users to use a system.

It asks them to exercise judgment alongside it.

Optimizing for speed or simplicity alone can actually make AI products worse by hiding uncertainty, flattening nuance, or encouraging misplaced confidence. Research from Microsoft has already shown that when AI systems feel too seamless, people tend to disengage cognitively, outsourcing judgment instead of exercising it. (The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers, https://dl.acm.org/doi/10.1145/3706598.3713778)

Ease of use matters. But ease of sensemaking matters more.

Why classic heuristics start to wobble

Usability heuristics assume a few things that no longer reliably apply:

• That the system knows what success looks like

• That errors are discrete and preventable

• That consistency equals reliability

• That clarity comes from constraint

With AI, outputs are rarely “wrong” in a traditional sense. They’re misaligned, under-specified, overconfident, or contextually inappropriate. Consistency can reduce usefulness. And constraint often limits exploration rather than enabling it.

In these systems the primary goal becomes judgement, and less about correctness.

Which means the designer’s role shifts. We’re no longer just preventing errors. We’re helping users develop better instincts about when to trust, question, refine, or even walk away from what the system produces.

Designing a relationship, not an interface

Traditional UX focuses on the boundary between human and system: inputs, outputs, flows, states.

AI design is about shaping an ongoing relationship.

That relationship includes:

• How intent is formed, expressed, and refined

• How disagreement and correction happen

• How trust builds, erodes, and recalibrates

• How users learn what the system is good or bad at over time

These dynamics don’t fit neatly into existing best practices because they’re not static. They evolve with use. They’re contextual. And they require design decisions that acknowledge uncertainty rather than masking it.

Intent becomes part of the design surface

In AI systems, the quality of the outcome is deeply tied to the quality of the intent provided. But most users don’t arrive knowing how to specify what they want — especially when they don’t yet understand what the system is capable of.

If your product assumes users already know how to prompt, constrain, or refine effectively, you’re designing for a small, expert audience.

Designing for AI means helping people:

• Discover what’s possible

• Articulate what they actually mean

• Iterate toward clarity rather than aiming for a perfect first attempt

But this is now the core experience, it cannot be treated like the edge case.

Trust is no longer something you “establish”

In traditional products, trust was often front-loaded. Brand, polish, and reliability did most of the work.

With AI, trust is fragile and dynamic. It shifts with every output.

Good AI experiences calibrate trust. They help users understand when the system is likely to be helpful, when it needs oversight, and when it should be ignored entirely.

When design fails to support that calibration, users swing between over-reliance and abandonment (often blaming themselves in the process).

Why this matters beyond design craft

Treating AI design as “UX as usual” is both a design mistake and a strategic one.

It leads teams to underestimate effort, overpromise timelines, and measure success with the wrong metrics. It produces products that demo well but struggle in real-world use. And it creates organizational confusion when adoption doesn’t match expectations despite “good UX.”

It’s important to note here that the nuance isn’t academic, especially when measured through its affects on how teams plan, staff, research, and evaluate success.

We’re early…and that’s the point

Best practices don’t appear overnight. They emerge from repetition, stability, and shared understanding. AI doesn’t have those yet.

We’re in a pre–best practice moment (aka the Wild West similar to early web or early mobile) where overconfidence in old rules can be more damaging than admitting uncertainty.

The teams that will do this well aren’t the ones applying familiar heuristics faster. They’re the ones willing to question which heuristics still apply at all.

Designing for AI isn’t about making systems easier to use.

It’s about making them harder to misunderstand.

And that shift, albeit subtle, changes everything.