Designing for Uncertainty: The UX Patterns Every AI Product Needs

Daniel Adeeri

AI products do not fail like deterministic software. They misunderstand, improvise, vary, stall, and succeed only partially. Good UX makes those uncertain states visible, manageable, and recoverable.

Traditional interfaces assume predictable behaviour. Tap “Save,” and the file saves. Apply a filter, and matching records appear. Failures usually map to specific errors.

Traditional interfaces assume predictable behaviour. Tap “Save,” and the file saves. Apply a filter, and matching records appear. Failures usually map to specific errors.

AI changes that contract. The same prompt can produce different outputs; a fluent response can be wrong; a task can succeed in four steps and fail in the fifth.

This changes what good UX must do.

An AI interface should help users understand what is uncertain, how much it matters, and what they can do next. “Human in the loop” must become a precise interaction system, not a vague slogan.

A practical way to begin is to design for five uncertain states: interpreting, planning, generating, acting, and failing.

1. Interpreting: what does the system think the user means?

Natural language feels direct, but it is full of unresolved choices.

“Make this more professional” could mean more formal, more concise, less emotional, or more appropriate for a specific audience. “Find the best plan” depends on price, speed, risk, and personal preference. If the product silently chooses, the result may be polished and still wrong.

Pattern: targeted clarification

Ask when an unresolved detail materially changes the outcome. Offer options, defaults, or an editable interpretation. Focus on intent, preference, missing constraints, or authority—not every minor detail.

Pattern: visible assumptions

For moderate-risk work, let the system proceed while showing the assumptions it made: “I’m optimizing for the lowest monthly cost and excluding annual contracts.” Users can correct the assumption without rewriting the request.

Pattern: editable task brief

For complex work, convert the prompt into a structured brief: objective, audience, constraints, sources, exclusions, and definition of done.

2. Planning: how will the system approach the task?

The agent may understand the goal but choose a weak method, miss a dependency, or use an inappropriate tool.

A plan preview can help, but it can also create false confidence. A CHI 2025 study of plan-then-execute agents found that plausible plans could be convincingly wrong; user involvement improved some imperfect plans, but did not automatically calibrate trust. The design lesson is not “always show a plan.” It is “show what the user can meaningfully evaluate.”

Pattern: consequence-based plan preview

Expose steps that affect money, data, permissions, external communication, or irreversible changes. Compress routine internal steps.

A useful plan should identify:

  • the sources and tools it will use;

  • the artifacts it will create or modify;

  • important dependencies;

  • approval points;

  • expected output and stopping condition.

Pattern: editable constraints, not fake reasoning

Let users change the destination, budget, audience, source set, or level of autonomy. Avoid long displays of internal reasoning that look authoritative but are difficult to verify.

Pattern: staged commitment

Separate exploration from execution. The system can research and prepare options first, then request approval before sending, buying, deleting, publishing, or representing the user.

3. Generating: how should variable output be presented?

Generated text, images, code, analysis, and recommendations vary in quality.

A common mistake is presenting generated content with the visual certainty of a database record, implying that it is complete and correct.

Pattern: editable draft by default

If the output is meant to be reviewed or refined, call it a draft and make editing easy. Preserve the user's original input and make regeneration non-destructive.

Useful controls include revising a selection, comparing alternatives, changing constraints, restoring a version, and accepting or rejecting individual suggestions.

Pattern: evidence and provenance

For factual tasks, show where claims came from. Grounding connects generated output to verifiable sources and supports auditability. Citations should lead to the relevant material, not merely a homepage or an impressive-looking list.

Source-level evidence is more useful than a blanket “AI may make mistakes” disclaimer.

Pattern: uncertainty signals with a job to do

Numeric confidence is often difficult to interpret. A percentage can look precise without explaining what was measured or what the user should do.

Prefer actionable uncertainty:

  • “Two sources disagree on the date.”

  • “I could not verify this claim from an approved source.”

  • “The image is too blurred to read the total.”

  • “This recommendation depends on your risk preference.”

Google's People + AI Guidebook recommends considering both AI confidence and user impact when deciding how much explanation or uncertainty to show. The higher the consequence and the weaker the evidence, the more explicit the interface should be.

4. Acting: what happens while the AI uses tools?

During tool use, the product may wait on another system, encounter a changed page, lack permission, or complete only part of a workflow.

Pattern: honest progress states

Replace generic loading with real states: interpreting, planning, searching approved sources, waiting for a service, ready for approval, applying changes, and verifying completion.

Do not show a percentage unless it is based on measurable progress. An indeterminate process with a fabricated “82%” is false precision.

For longer tasks, show a timeline or checklist with queued, active, complete, blocked, skipped, and failed states. Make important status changes available to assistive technologies; WCAG 2.2 includes guidance for programmatically identifying status, progress, success, and error messages without forcing focus changes.

Pattern: point-of-risk confirmation

OpenAI's computer-use guidance recommends confirming when the risky action is about to happen, with an explanation of the action, risk, and intended effect. This is more useful than asking for broad permission before the system has enough context.

Confirmation is especially important before external submissions, financial transactions, deletions, permission changes, sensitive-data transmission, or actions that represent the user.

Pattern: pause, redirect, and resume

Users should be able to stop without losing completed work and see which steps must be repeated after a changed constraint.

Pattern: completion evidence

“Done” should mean verified, not attempted. Link to the sent message, changed file, booking confirmation, created ticket, or published record. State any partial results explicitly.

5. Failing: how does the product create a way forward?

AI can refuse, time out, misunderstand, produce weak output, fail to use a tool, or finish only part of a workflow.

A good error identifies the affected step, preserves successful work, and offers a relevant recovery path.

Pattern: specific failure and preserved progress

Compare “Something went wrong” with: “The draft is saved, but the publishing step failed because your session expired.” The second message tells the user what remains safe and what needs attention.

Pattern: recovery ladder

Offer recovery options in increasing order of effort:

  1. retry the failed step;

  2. revise the input or constraint;

  3. choose an alternative source, tool, or method;

  4. undo or restore;

  5. hand off to a person.

Do not offer regeneration when the problem is missing data, permission, or an external outage.

Pattern: graceful degradation

If the full task cannot be completed, return useful partial work and name the gap. A scheduling assistant that cannot book can still prepare options and a ready-to-send message.

A pattern-selection matrix

Use four dimensions to scale explanation and control.

Dimension

Lower-risk condition

Higher-risk condition

Design response

Consequence

Easy-to-ignore suggestion

Financial, legal, medical, reputational, or access impact

Increase explanation, review, and approval

Uncertainty

Strong evidence and clear intent

Ambiguous request, weak evidence, conflicting sources

Clarify, show assumptions, cite gaps

Reversibility

Cheap, local undo

External or irreversible action

Stage commitment and confirm at risk

User expertise

User can inspect the result easily

User cannot reliably judge correctness

Add provenance, guardrails, and escalation

A reversible writing suggestion may need only accept and reject. A consequential recommendation with uncertain evidence needs sources, limitations, alternatives, and review. A hard-to-reverse action should not hide behind one optimistic button.

False-confidence patterns to avoid

Avoid patterns that add the appearance of certainty without trust:

  • Decorative confidence scores: Numbers without calibration, explanation, or a clear user action.

  • Generic “done” states: Completion labels unsupported by an artifact or system confirmation.

  • Finished-looking generated content: Outputs styled as authoritative records when they still require review.

  • Invisible automation: Actions taken without a clear plan, scope, or record.

  • Blanket consent: Broad permission granted long before the exact risk is known.

  • Endless spinners: No indication of what the system is doing, waiting for, or able to preserve.

  • Regenerate as the only recovery: A new roll of the dice instead of correction, comparison, or escalation.

  • Disclaimers instead of design: “AI can make mistakes” used as a substitute for evidence and control.

The uncertainty QA checklist

Before release, test each major workflow across the five states.

Interpreting

  • Are material assumptions visible?

  • Does ambiguity trigger a focused clarification?

  • Can the user correct the interpreted goal without restarting?

Planning

  • Are consequential steps, tools, and approval points clear?

  • Can users edit meaningful constraints?

  • Does the plan avoid implying certainty it cannot support?

Generating

  • Is generated output clearly reviewable and editable?

  • Are factual claims grounded where appropriate?

  • Are uncertainty signals specific and actionable?

Acting

  • Are status and progress states honest and accessible?

  • Are risky actions confirmed at the point of risk?

  • Can users pause, redirect, or verify completion?

Failing

  • Does the error identify the affected step and preserved work?

  • Are retry, revise, alternative, undo, and handoff paths available?

  • Have partial failure and external-system failure been tested?

Uncertainty is part of the product

Better prompts, data, and models matter, but uncertainty does not disappear. A stronger system takes on more complex tasks and gains more opportunities to act beyond the user's intent. The product still needs a clear interaction contract.

Designing for uncertainty means making the system's interpretation, evidence, status, limits, and recovery paths usable. It means scaling oversight with consequence rather than adding friction everywhere. It means treating correction as normal, not exceptional.

The strongest AI experience is not the one that looks certain at every moment. It is the one that helps people make good decisions even when the system is not.

Stay in the Loop!

Discover insights, updates, and my perspective on technology, design, AI, creativity, and growth straight to your inbox.

Unsubscribe at any time.

Stay in the Loop!

Discover insights, updates, and my perspective on technology, design, AI, creativity, and growth straight to your inbox.

Unsubscribe at any time.

Black and white portrait of a man with a beard and glasses

Daniel Adeeri, PMP®

Design, Product & AI

Contact me

Fill out the form with your enquiries, or reach out directly. I’ll respond within 6 hours.

Let’s chat!

Or send a direct e-mail

© Copyright 2026. All rights Reserved.

Black and white portrait of a man with a beard and glasses

Daniel Adeeri, PMP®

Design, Product & AI

Contact me

Fill out the form with your enquiries, or reach out directly. I’ll respond within 6 hours.

Let’s chat!

Or send a direct e-mail

© Copyright 2026. All rights Reserved.