AI broke the deterministic contract
Classic interface design assumes the same input gives the same output. You click Save and the file saves. AI features break that contract. The same prompt can produce different answers. A response can take one second or sixty.
As AI moves from answering questions to taking actions, those gaps get more expensive. The patterns below are the ones we reach for when designing AI features. They follow the four phases in Microsoft's Guidelines for Human-AI Interaction (Amershi et al., CHI 2019): before the user starts, during the interaction, when the AI is wrong, and over time. Those 18 guidelines, Google's People + AI Guidebook and Emily Campbell's Shape of AI pattern library are the canon for this work, and each pattern below points back to them.
Before the first interaction: set expectations
Pattern 1: Say what it can do, and how well. The first two HAX guidelines ask products to make clear what the system can do and how well it can do it. Most AI features skip the second. A line like "Drafts replies for common billing questions. Check amounts before sending" sets a more accurate mental model than "Your AI assistant".
Pattern 2: Introduce capability in steps. Start with the lowest-risk interaction, such as a suggestion the user can accept or dismiss. Offer more autonomy as people learn where the system is reliable. PAIR's chapter on mental models recommends this kind of gradual onboarding.
Pattern 3: Show examples instead of a blank box. An empty prompt field asks people to guess what works. Suggested prompts, templates and worked examples, which Shape of AI files under "wayfinders", show the range and the limits at once.
During the interaction: show the real work
Pattern 4: Stream, and let people stop. A spinner for ten seconds feels broken. Streaming output lets people start reading and judge the direction early. Design four states: a short working indicator before the first words, the progressive render, a stop control that keeps partial output, and a clear completion state.
Pattern 5: Show the plan before the action. For agents, show the steps the system intends to take and let the user edit them before it runs. Shape of AI groups this with citations and verification under "governors", the patterns that keep a person in charge of what happens.
Pattern 6: Calibrate trust instead of maximising it. PAIR's Explainability + Trust chapter aims for calibrated trust: people rely on the system when it is reliable and use their own judgment when it is not. Show uncertainty at the point of decision, in words the user can act on. "Two invoices could match this payment. Pick one" helps more than a 62% confidence badge.
Pattern 7: Handle citations and reasoning traces with care. NN/g's research on explainable AI in chat interfaces (Megan Chan, December 2025) found that people rarely click citations, so a list of sources can create false confidence. It also warns that step-by-step "reasoning" shown to users is often produced after the fact. Show the sources the system used and make the key claim quick to check. Put disclaimers in plain, specific language next to the input.
Pattern 8: Keep the tone human-sized. The same NN/g article found that human-like language inflates trust. A tool that says "I'm sure" when it isn't teaches users to stop checking.
When the AI is wrong: design recovery first
AI fails in ways deterministic software does not: confident wrong answers, off-topic output, refusals and timeouts. HAX gives this its own phase, and PAIR has a chapter on errors and graceful failure.
Pattern 9: Make correction cheap. Let people edit the output in place, regenerate with a hint or undo an action. The HAX guidelines call for efficient correction and efficient dismissal.
Pattern 10: Design each failure mode on purpose.
- Wrong or irrelevant output. Offer a retry with a way to add context, and keep the previous version.
- Refusal or guardrail hit. Say what the system can't do and what the user can try instead, without implying they did something wrong.
- Timeout. Keep partial results and offer to continue.
- Low confidence. Hand over to a person with the context attached, so the user doesn't have to repeat themselves.
We cover recovery flows in more depth in AI error handling UX.
Over time: learn, and let people steer
Pattern 11: Feedback that feeds something. Thumbs up and down is the minimum. For important workflows, offer a short set of reasons ("wrong amount", "wrong tone", "missing information") so the team can group failures. Then tell people what their feedback changes. HAX asks products to encourage granular feedback and to convey the consequences of user actions.
Pattern 12: Let people set the autonomy. Give users a visible, reversible control over how much the system does alone: suggest only, draft for approval, or act and report back. Our guide to levels of autonomy for AI agents covers how to choose the default.
Test the patterns with real users
A pattern that works in one product can fail in another, because trust depends on the task and the stakes. Two cheap ways to check:
- Wizard of Oz tests. A person plays the AI behind a prototype, so you can test expectations and error flows before the model exists. NN/g's guide to the method covers the setup.
- Scripted failure sessions. In a usability test, make the AI wrong on purpose at step three. Watch whether people notice, how they recover and whether they trust the next answer. Trust shifts after the first failure, so test for it.
FAQ
What are the most important UX patterns for AI? Setting accurate expectations, showing uncertainty where decisions are made, making correction cheap and designing the failure states. All four map to Microsoft's HAX guidelines and Google's PAIR guidebook.
Should we show confidence scores? Rarely as raw percentages. Most people can't act on "72%". Translate uncertainty into a next step: check this field, pick between these options, or ask a person.
Do citations make AI answers more trustworthy? They can make answers look more trustworthy than they are, and NN/g found people rarely click them. Link each specific claim to its specific source, and make checking fast.
Where do we start with an existing AI feature? Audit it against the 18 HAX guidelines, list the failure modes you haven't designed, and fix the one users hit most often.
For the full picture of trust, transparency and control, read our pillar guide, UX for AI: designing AI products people trust. If you want an AI feature reviewed against these patterns, our UX for AI consulting starts with an audit of every state your users see.