How do we stop users over-trusting the system?

Make verification cheap rather than making warnings louder. Users over-trust when checking is expensive. Cited evidence that opens in place turns verification from a task into a glance, which is a far more reliable control than a disclaimer.

Is conversational UX ever right?

When the task is genuinely exploratory and user intent cannot be anticipated. For structured, repeated professional work, triage, review, drafting, a structured surface almost always outperforms a chat box.

Expertise / AI UX Design

Interaction design for systems that are sometimes wrong.

Conventional UX assumes the system does what it is told, fails visibly, and has a bounded scope. AI products violate all three, and most patterns we rely on break with them.

Start a conversationAI & Product Design in full

Who this page is for

For product and design teams designing AI features and unsure which conventions still apply.

Related capability

Product design
SaaS design
Mobile app design
Enterprise UX
AI product strategy
AI experience design
AI UX

What actually changes

Less than the discourse suggests, and in specific places. Information architecture still matters. Typography, density, empty states and accessibility still matter: accessibility more, because generated content is harder to make navigable.

Three things genuinely change: output is not predictable, so mental models built by repetition become unreliable; failure is invisible, because a wrong answer looks exactly like a right one; and scope is unbounded, so an open input communicates no affordance at all.

The patterns that work

Provenance on every finding, opening the source in place. Uncertainty expressed in language rather than percentages: 'the policy wording is ambiguous on flood exclusion' tells a user where to look; '73% confident' tells them how to feel.

Correction inline on the specific finding, cheaper than ignoring it. Autonomy tiered by what an error costs and shown explicitly in the interface. And honest limits: a product that states what it cannot do is trusted more than one that implies generality.

Conversational is usually the wrong default

A chat box is the easiest AI interface to build and frequently the worst one to use. It hides capability, provides no affordances, makes output hard to scan and puts the burden of knowing what to ask entirely on the user.

Most valuable AI features are better expressed as structured surfaces, an organised, cited view of a document; a ranked queue with reasons; a draft the user edits. Conversation earns its place when the task is genuinely exploratory and the user's intent cannot be anticipated.

Signals

When this work is worth doing

If two or more of these are true, the conversation is probably overdue.

Users check the work manually anywayThe output cannot be verified cheaply, so the feature is overhead.
Nobody corrects wrong outputCorrecting is slower than ignoring, so you are blind to your failure modes.
One autonomy level across all casesToo cautious for routine work, too aggressive for consequential work.
It is a chat box because that was fastestThe interface hides what the system can do and puts the work on the user.

How we scope it

Three ways this usually runs

Durations are typical rather than fixed. We scope after the first conversation, not before.

AI UX review
2 to 3 weeks
Evaluation of an existing AI feature against provenance, autonomy, uncertainty and correction.
You receive
- Findings & severity
- Pattern recommendations
- Redesign direction
AI experience design
8 to 12 weeks
Autonomy model, interaction design and a prototype tested against real failure cases.
You receive
- Autonomy model
- Interaction patterns
- Tested prototype
AI pattern library
8 to 12 weeks
Reusable AI interaction patterns integrated with your existing design system.
You receive
- Pattern library
- Content & tone guidance
- Integration with design system

AI UX Design

Questions we hear before we start

Rarely. Users cannot calibrate a percentage, systems are frequently not calibrated either, and one number collapses different kinds of uncertainty. Say what is uncertain and why, per finding, in language.

The thinking behind this work

All insights

AI28 August 20269 min read

Why AI products need a different UX framework

Conventional interaction design assumes the system does what it is told. AI products break that assumption, and most of the patterns we rely on break with it.

Read the thinking