AI UX design patterns for shipping AI features people actually trust

An AI UX design guide for SaaS teams: when to use AI, confidence, sources, undo, streaming, errors, feedback and the EU AI Act rules from August 2026.

Summarize with
A laptop with lines of code appearing and a gear turning beside it

AI UX design is the work of making a probabilistic feature dependable enough to rely on. AI features work most of the time, and nobody can tell in advance when they will not. The design job is to make that visible and cheap.

People trust an AI feature when they know what it can do, can see how sure it is and where an answer came from, can undo a wrong answer in a click, and are told plainly that it is AI. In the EU that last point is law: since 2 August 2026, Article 50 of the AI Act requires it, with fines of up to EUR 15 million or 3% of worldwide turnover, whichever is higher.

This guide is for product teams adding assistants, generation, recommendations or automation to a SaaS product: when to use AI, the trust patterns, a checklist, how to measure trust and the EU rules. If your product will also live inside ChatGPT or Claude, see designing apps for AI assistants.

When AI is the right tool, and when it is not

Google's People + AI Guidebook (PAIR) starts with a question many teams skip: would a simple rule do? Rules are easier to build, explain and debug. PAIR finds AI is probably better for personalised recommendations, prediction, natural language and spotting rare events that change over time, such as fraud. It is probably not better when predictability is the point, errors are costly, people need full transparency, or users have said they do not want the task automated.

PAIR also suggests automating tasks that are tedious or need scale, and augmenting tasks people enjoy or where there is no single right answer. Apple adds: offer generative features only where they bring clear value, and keep a non-AI path where you can.

Then ask two questions: what does a wrong answer cost, and can the user spot it quickly?

Cost of a mistakeEasy to check?PatternExample
LowYesAct, with undoTagging support tickets
LowNoSuggest, and say whyNext step in a CRM
HighYesDraft; the user commitsA reply to a customer
HighNoAdvisory with sources, or no AIA payment amount
Two questions that pick how much an AI feature should do on its own.
Two questions that pick how much an AI feature should do on its own.

Set expectations and onboard in stages

Microsoft's 18 Guidelines for Human-AI Interaction, tested by 49 design practitioners against 20 AI products, open with this: make clear what the system can do (G1) and how well it can do it (G2).

  • Name the job, not the technology: "Drafts replies from this ticket's history", not "Powered by AI".
  • Put limits next to the feature. Apple advises stating known limitations up front and explaining why weak results happen.
  • Seed open inputs with example requests. An empty prompt box teaches nothing.

PAIR recommends onboarding in stages: what it can do, what it cannot, how it may change and how to improve it. Show each when it becomes relevant, not in one tour.

Give the assistant a function, not a personality. Nielsen Norman Group (NN/g) reports a study of 410 US participants in which perceived intelligence raised advice-taking from ChatGPT and perceived emotion lowered it. Its advice: no human names or fake avatars.

Show confidence, uncertainty and sources

The goal is calibrated trust. PAIR puts it plainly: users should not trust an AI system completely; they should know when to trust it and when to apply their own judgement.

Apple's warning on confidence comes first: check that your scores track real quality, and if you are unsure, do not show them. When you do:

  • Use categories with actions (high, medium, low), with carefully chosen cut-offs.
  • Prefer reasons to numbers. Apple's example: "Because you listen to pop music" beats "97% match".
  • Let ranking imply confidence, and show alternatives when the top answer may be wrong.
  • Hedge in the first person: NN/g recommends "I'm not sure" over "It's not clear".

On explanations, NN/g's December 2025 article notes that chat citations are sometimes hallucinated and rarely clicked, yet still raise confidence, and step-by-step "reasoning" can be a rationalisation written after the fact. Put each source beside the claim it supports and link to the passage. Ground answers in the customer's own records and name them. Explain right after a user action, when PAIR says people learn fastest.

Human in the loop: review, editing and undo

Make mistakes cheap: suggestions instead of silent changes, drafts instead of sent messages, undo right where the action happened. People try more when trying is safe. Apple advises against automating destructive or hard-to-undo actions, such as deleting photos or making purchases, and recommends confirmation before significant actions.

What the AI doesControl pattern
Reversible, low impactAct, then show undo and a log entry
Visible to other peopleDraft and preview; the user sends
Changes many recordsDiff preview, one-click revert
Spends money or deletesConfirmation that names the consequence
Runs as an agentPlan first, step log, pause and stop

For editing, Microsoft asks for efficient dismissal and correction (G8 and G9), and Apple recommends Edit, Undo, Retry and Adjust beside generated content. Let people regenerate one part without losing their edits, and persist corrections. Apple's sharpest line: never rely on corrections to make up for low-quality results. NN/g adds that chat nudges people on to the next question, which discourages checking, so place review before the next prompt.

From my work: RoomMind AI, an iOS app that redesigns a room from one photo, that I designed.
From my work: RoomMind AI, an iOS app that redesigns a room from one photo, that I designed.
From my work: CareUp, an AI healthcare app for finding a doctor and booking, that I designed.
From my work: CareUp, an AI healthcare app for finding a doctor and booking, that I designed.

Latency, streaming, errors and refusals

Jakob Nielsen's response-time limits still hold: 0.1 seconds feels instant, 1 second keeps the flow of thought, and after 10 seconds people need progress and a way to interrupt. Generative models often run longer, so:

  • Stream text people read top to bottom, with a Stop button that keeps the partial output.
  • Replace the spinner with specific steps. Apple's example: "Finding substitutions for ingredients", not "Processing...".
  • Run long jobs in the background and notify on completion.
  • Do not stream structured output you validate later; show a skeleton, then fill it.

PAIR notes that what counts as an error depends on expectations: a recommender useful 60% of the time can be a success to one person and a failure to another. Give every failure a path forward:

  • Timeout: keep the input, offer Retry and the manual route.
  • Low confidence: ask a clarifying question (Microsoft's G10).
  • Refusal: say so plainly, as Apple's Image Playground does ("Unable to use that description"), and suggest a request that works.
  • Reported mistake: fix it in place and remember the correction.

Feedback loops and privacy disclosures

PAIR separates implicit feedback (accepts, edits) from explicit feedback (ratings), and says to tell people what you collect and how to opt out. Apple wants feedback voluntary and unobtrusive, such as thumbs up and down, and Microsoft's G15 asks for granular feedback on parts of an answer. Be honest about impact: if a thumbs-down goes to a review queue, say so.

For privacy, answer four questions one tap from the feature. Is data sent to a third-party model provider? Is it used for training? How long is it kept? Can an admin switch the feature off (Microsoft's G17)? Apple's guidance is to ask permission, use the minimum and say whether data trains models.

The EU AI Act transparency rules that apply from August 2026

I am a designer, not a lawyer, and this is not legal advice: it is my reading of the official text and the European Commission's guidance as of 2 October 2026.

Article 50 of the AI Act (Regulation (EU) 2024/1689) applies from 2 August 2026. The AI Omnibus, in force since 27 July 2026, delayed the high-risk rules but kept that date, adding one grace period: generative systems on the market before 2 August 2026 have until 2 December 2026 for machine-readable marking. Providers outside the EU are covered when they place a system on the EU market or its output is used there.

RuleWhoIn the product
50(1) Say it is AI, unless obviousProviderDisclosure at the first interaction
50(2) Machine-readable marking of generated contentProviderMetadata or watermarks
50(3) Emotion recognition, biometric categorisationDeployerNotice at exposure
50(4) Deepfakes, public-interest textDeployerVisible label (text is exempt after human editorial review)
50(5) Clear, distinguishable, accessibleBothNever only in the terms

The Commission's guidelines of 20 July 2026 are non-binding but practical:

  • Use plain-language labels and a first-turn line in chatbots, optionally with a persistent badge, near the input or output field.
  • One prominent notice is likely enough in most cases; financial, legal and health advice, vulnerable users and AI agents are likely to need periodic reminders.
  • Not enough alone: disclosure only in the terms, a vague "assistant" label, "Services on this website use AI" or "this system uses LLMs".
  • The system must disclose its nature whenever someone asks.

Recommender systems, translation and autocompletion fall outside the interaction rule; AI summaries need marking, grammar fixes do not. A voluntary Code of Practice on marking and labelling was published on 10 June 2026, with about 190 signatories by the end of July. For SMEs and small mid-caps, the lower fine cap applies.

Write the labels in every language you ship; for German and English products, see my UI/UX design for Berlin startups. Outside the EU, Texas has expected a plain AI disclosure in health care and government services since 1 January 2026, which I plan for as a SaaS product designer in Austin.

AI UX design patterns at a glance

PatternWhen to use itExample
Capability lineEvery AI entry point"Drafts replies from this ticket's history"
AI disclosureAny direct interactionA badge plus a first-turn line
Confidence categoryScores you have checked"Needs review" on extracted fields
Reason attributionRecommendations"Because this deal went quiet after the demo"
Inline sourcesFactual answersA citation chip that opens the passage
Specific progressWaits over a second"Reading the attachments"
Coached refusalBlocked requestsA suggested request that works
Global controlsAdminsA workspace switch to turn it off

How to ship an AI feature: a pre-launch checklist

  1. Write the job and the limits in one sentence each, and test both with real users.
  2. Map the cost of a wrong answer for each output and pick the pattern.
  3. Build an evaluation set from real inputs, including the vague, out-of-scope and sensitive requests Apple recommends testing.
  4. Show confidence only after checking scores against that set.
  5. Design every state: loading, streaming, low confidence, refusal, timeout and undone.
  6. Put the AI disclosure and data notice in the interface.
  7. Build undo, history and global controls before launch.
  8. Test the disclosure and streamed text with screen readers.
  9. Re-test before model updates and tell users when behaviour changes (Microsoft's G14 and G18).

How to measure trust after launch

Trust is not usage. NN/g advises judging a time-saving feature by whether it saves time, not by adoption. Track behaviour and a short question after key tasks on a dashboard your team will open.

SignalWatch out for
Acceptance without editsHigh rates on risky tasks may mean nobody checks
Edits before useA rise after a model change
Undo and override rateSpikes in one feature
Retries and rephrasingLong chains before people give up
Source opens, review timeNear zero on high-stakes outputs

NN/g notes that people often treat confident output as finished work, so over-trust is as real a risk as distrust.

Common mistakes to avoid

  • A chatbot where a button would do.
  • A generic disclaimer instead of design; NN/g notes such warnings lose impact through repetition.
  • Unchecked confidence scores, and citations that do not support the claim.
  • Silent actions with no undo or log.
  • A human name and face for the assistant.
  • Errors and refusals with no next step.
  • Swapping models quietly.

Frequently asked questions

What is AI UX design?

AI UX design is designing the experience of features driven by machine learning or generative models: assistants, generated content, recommendations and automation. Beyond the usual UX work, it handles outputs that vary, confidence, explanations, undoing what the AI did, long waits, refusals and disclosing AI use. The aim is calibrated trust: people rely on the feature when it is right and notice when it is not.

How do you design AI features people trust?

Say clearly what the feature can and cannot do, show confidence only when it is reliable, and put sources next to claims. Keep people in control with drafts, previews and undo. Design the wait with specific progress and a Stop button, give every error a next step, and explain how data is used.

Should I show AI confidence scores to users?

Only if you have checked that the scores track real quality; Apple's guidance is not to show confidence when you are unsure what the values mean. When you do, translate them into categories with clear actions, reasons such as "because you listen to pop music", or a ranked list of options. Raw percentages suit places where people expect numbers, such as forecasts.

Does the EU AI Act require labelling AI features?

From 2 August 2026, providers must tell people when they are interacting directly with an AI system unless it is obvious, and generative AI outputs must be marked in a machine-readable way. Deployers must label deepfakes and some AI text on matters of public interest. Providers outside the EU can be covered too. This is general information, not legal advice.

Is a chat interface the best way to add AI to a product?

Often not. Chat suits open-ended questions, but many SaaS jobs are better served by AI inside the existing workflow: a suggested field value, a drafted reply, a summary on a record or a one-click action with undo. These are easier to discover, check and reverse. Use chat when requests are truly open, and seed it with example prompts.

How do you measure trust in an AI feature?

Combine behaviour with what people say. Track acceptance without edits, edits before use, undo and override rates, retries, and whether people open sources or review outputs on risky tasks. Compare task time and quality with and without AI, and ask after key tasks whether people knew when to rely on the result. High acceptance with no checking is a warning sign.

My recommendation

Start with one AI feature, not an AI strategy. Map the cost of a wrong answer, pick the pattern, and design the disclosure, failure and undo states before the happy path. Sources, approvals and undo are also among the interface trends I expect in 2027.

If you want help, see how I approach product design for SaaS teams, and how I build these states in React and Next.js development. Owning both the design and the code of states like these is what a design engineer does.

Sources and method

All sources were opened on 2 October 2026. EUR-Lex blocked automated access, so the AI Act text was read on the Commission's AI Act Service Desk (EUR-Lex consolidated version as at 27 July 2026). Independent data on which trust patterns work best in SaaS is thin; this guide relies on design guidelines and research reviews.

Tags: AI, AI UX design, Product design, SaaS, EU AI Act

Keep reading

Let’s look at your product.

A free 30 minute call. You leave with a clear next step, whether we work together or not.

Book a free intro call