Patterns Cheat Sheet

Reference intermediate

The Four Core AI UX Patterns

Each pattern defines how AI relates to the user's workflow. Choosing the wrong pattern is the most common strategic mistake in AI product design.

Pattern	Definition	Canonical Examples
Chat	User initiates an open-ended conversation; AI responds in natural language	Customer support bots, general-purpose assistants, research tools
Inline Suggestions	AI offers completions or edits inside the user's existing workflow without breaking context	Writing assistants, code completion, smart compose in email
Autonomous Agent	AI executes multi-step tasks independently and reports back; human reviews outcomes	Scheduling agents, research automation, document drafting pipelines
Ambient Intelligence	AI operates silently in the background; user never directly interacts with the model	Spam filters, content ranking, anomaly detection, auto-tagging

Pattern Decision Table

Use this table to select the right pattern before you start designing.

Criteria	Chat	Inline Suggestions	Autonomous Agent	Ambient
Task complexity	High — open-ended, exploratory	Medium — single-step assist	High — multi-step, goal-directed	Low to medium — classification, ranking
User involvement needed	High — user drives the dialogue	Medium — user accepts or rejects	Low — user reviews final output	None — no user action required
Tolerance for errors	Medium — user can rephrase	Medium-low — wrong suggestions break flow	Low — errors compound across steps	Low — silent errors go unnoticed
Latency sensitivity	Low — conversational pace expected	High — must feel instant	Low — background task	Very high — must not add perceptible lag
Explainability requirement	Medium — user can ask follow-up	Low to medium	High — user needs audit trail	High — regulatory or fairness need
Best starting point for new teams	Yes — lowest infra complexity	Yes — high user value, limited scope	Only with strong fallback design	Only with robust monitoring in place

User Expectations by Pattern

Understanding what users expect prevents the most common trust failures.

Pattern	What Users Expect	Where Expectations Break Down
Chat	Accurate, current information; ability to ask follow-ups; clear indication of limitations	Hallucinated facts presented with confidence; no admission of uncertainty
Inline Suggestions	Speed (< 200 ms feels instant); relevance to current context; easy dismissal	Suggestions that interrupt rather than assist; no way to turn them off
Autonomous Agent	Transparent progress updates; ability to pause or cancel; summary of what was done and why	Silent failures; irreversible actions taken without confirmation
Ambient	Reliable over time; fair and consistent; explainable on demand	Invisible errors that compound; perceived unfairness or bias; no override path

Risk Profile by Pattern

Pattern	Primary Risk	Secondary Risk	Mitigation Approach
Chat	Hallucination accepted as fact	Over-reliance on AI judgment	Hedging language; source attribution; "verify before acting" prompts
Inline Suggestions	Flow disruption; suggestion quality degrades on unusual input	Users accept suggestions without reading	Easy keyboard dismiss; acceptance rate monitoring
Autonomous Agent	Irreversible action taken incorrectly	Scope creep beyond user intent	Confirmation gates for destructive actions; detailed audit log
Ambient	Silent bias or discrimination at scale	User has no recourse	Fairness monitoring; "why did this happen?" explanation; manual override

Trust-Building Checklist

Apply these before shipping any AI feature, regardless of pattern.

[ ] Confidence indicators — Does the UI signal when the AI is uncertain? (Hedging phrases, probability labels, "may be inaccurate" notices)
[ ] Source attribution — Can the user see where the AI's information came from? Especially important for factual claims.
[ ] Explainability — Can the user understand, at a high level, why the AI produced this specific output?
[ ] User control — Can the user adjust, override, correct, or disable the AI? Is that path obvious?
[ ] Appropriate confidence calibration — Does the AI admit what it does not know, or does it always sound certain?
[ ] Disclosure — Is it clear to the user that they are interacting with AI, not a human?
[ ] No deceptive framing — Does the interface avoid implying AI outputs are authoritative or verified when they are not?
[ ] Recourse path — If the AI is wrong, is there an obvious way for the user to fix or report it?

Error Handling Checklist

AI errors are inevitable. These patterns make failures recoverable.

[ ] Graceful degradation — If the AI fails completely, does the product still provide value (fallback to search, manual entry, cached result)?
[ ] Hedging language — For uncertain outputs, does copy say "suggested," "estimated," or "may include errors" rather than presenting results as fact?
[ ] Easy correction flow — Can users edit, regenerate, or request alternatives in one or two actions?
[ ] Failure acknowledgment — When the AI cannot complete a task, does it say so clearly rather than producing a low-quality result silently?
[ ] Error categorisation — Do you distinguish between "AI was wrong" and "AI was right but the user changed their mind"? These require different responses.
[ ] No silent compounding — For autonomous agents, does an early error halt the task or trigger human review rather than cascading into further incorrect actions?
[ ] Undo / rollback — For actions the agent takes on the user's behalf, is there an undo path?

Feedback Collection Comparison

Mechanism	Signal Quality	Participation Rate	Best For
Thumbs up / thumbs down	Medium — binary, limited context	Low — adds friction	Quick sentiment baseline; prioritising what to improve
Star rating	Medium	Low	Overall satisfaction; comparisons across versions
Correction / edit	High — reveals exactly what was wrong	Very low — only motivated users	Ground truth for model improvement; identifying systematic errors
Regenerate click	Medium-high — clear dissatisfaction signal	Medium	Measuring output quality; surfacing pattern failures
Acceptance rate (inline)	Medium — user accepted but may not have read	High — implicit, no friction	Measuring suggestion relevance at scale
Edit distance after acceptance	High — shows how much user had to fix	High — implicit	Quality of inline suggestions; how much trust users place in outputs
Click-through / dwell time	Low — noisy, correlates with many things	High — implicit	Engagement; not a substitute for task completion

Principle: Implicit feedback is abundant but noisy. Explicit feedback is sparse but high-signal. Use both and triangulate. Show users the impact of their feedback ("your rating helped improve this") to increase participation.

Personalization Boundary Guide

Personalization increases value but can destroy trust if it crosses the wrong line.

Personalization Type	Generally Welcomed	Potentially Uncomfortable	Actively Harmful
Tone and complexity adaptation	Adjusting reading level based on past interaction	Changing persuasion style based on inferred emotional state	Exploiting detected vulnerability
Content recommendations	Surfacing topics the user engages with	Narrowing to only familiar content (filter bubble)	Amplifying harmful or addictive content for engagement
Workflow shortcuts	Remembering user preferences and defaults	Removing options the user "never uses" without warning	Locking user into a workflow they cannot change
Inferred attributes	Adapting to observed behaviour	Acting on inferred demographics without user knowledge	Discrimination based on protected characteristics

Rule of thumb: If the user would feel surveilled rather than helped by a personalisation, it has crossed the line. Offer transparency ("why am I seeing this?") and user control for every personalisation dimension.

Metrics Framework at a Glance

Traditional product metrics often mislead for AI features. Use these alongside standard engagement data.

Metric	What It Measures	Watch Out For
Task completion rate	Did the AI help users accomplish their goal?	Users abandoning rather than failing — counts as incomplete regardless
Time-to-value	How quickly does the user get useful output?	Fast but wrong output is worse than slow but accurate
Error correction rate	How often do users fix AI outputs?	High correction rate = low quality; zero correction rate may mean users aren't reading
Feature adoption rate	Do users try the AI feature at all?	High adoption does not mean high satisfaction — check retention at 30 days
Feature retention (30 / 90 day)	Do users keep coming back to the AI feature?	The most honest signal of real value
Acceptance rate (suggestions)	What share of AI suggestions are accepted without edit?	Gaming risk if users accept without reading; pair with edit-distance
Satisfaction (CSAT / NPS)	How do users feel about the AI interaction?	Survey fatigue; compare AI-assisted vs. unassisted cohorts
Guardrail metric: error reports	How often do users report AI errors?	Lagging indicator; combine with proactive quality sampling

Set a baseline. Before shipping, measure the same task with no AI assistance. Your AI feature must outperform the baseline on task completion and time-to-value, not just on engagement.

Ethics Review Checklist

Run this review before any AI feature ships to production.

[ ] Transparency — Do users know they are interacting with AI? Is that disclosed at the right moment (not buried in a privacy policy)?
[ ] Accuracy disclosure — Are the AI's limitations communicated clearly and early, not just in edge-case failure messages?
[ ] Informed consent — Do users understand how their data is used to power or improve the AI feature?
[ ] Autonomy preservation — Does the feature augment user judgment, or does it replace it? Can the user make meaningful decisions without depending on the AI?
[ ] Equitable access and outcome — Have you tested whether the AI performs consistently across different user groups, languages, and accessibility needs?
[ ] Dark pattern check — Does the feature use AI to nudge users in ways that serve business goals at the expense of user wellbeing? (urgency manipulation, variable reward loops, obscured opt-out)
[ ] Harm at scale — If this feature is wrong 1% of the time, and you have one million users, what is the real-world impact of that 1%?
[ ] Reversibility — Can the team disable or roll back the feature quickly if a harm pattern is discovered post-launch?

Prototyping Stage Checklist

Stage	Method	Validates	When to Move On
Concept	Wizard of Oz (human simulates AI)	Does the concept solve a real user problem?	Users find the simulated output useful; qualitative signal is strong
Design	Prompt prototype (live LLM demo, no product infrastructure)	Is output quality good enough? Does the UX feel right?	Acceptance rate > threshold in unmoderated testing
Alpha	AI-in-the-loop (AI suggests, human approves before output shows)	Is the AI safe enough to show users?	Human reviewers approve > 90% of AI outputs without edits
Beta	Feature-flagged rollout (5–10% of users)	Does the metric framework hold at scale?	Task completion and retention targets met; no safety incidents
General availability	Full rollout with monitoring	Long-term value and trust	30-day and 90-day retention stable; error correction rate plateauing or declining

Key Reference Sources

Source	What It Covers	Why It Matters
Google PAIR People + AI Guidebook (pair.withgoogle.com)	End-to-end design process for AI features; patterns, exercises, and worked examples	Free, comprehensive, written by practitioners who shipped products at scale
Microsoft Human-AI Interaction Guidelines (Amershi et al.)	18 empirically-validated design guidelines across four phases of AI interaction	Research-backed; covers guidelines that generic UX principles miss
Nielsen Norman Group "UX for AI" series	Usability research specific to AI interfaces; what users actually do vs. what designers assume	Evidence-based; covers chat, inline suggestions, and ambient patterns

Quick links: Course Home · UX Patterns deep dive · Rules of thumb · Glossary · Troubleshooting guide