AI-Native UX
Context
Section titled “Context”Your design team built a prototype: a chat interface for your new AI feature. The CEO thinks it looks “very ChatGPT.” The problem: 80% of your users don’t want to chat. They want results.
As a PM, you need to understand that AI interfaces aren’t just “normal UI plus chat.” AI output is non-deterministic — and that changes everything: how you display results, how you build trust, and which interaction patterns work.
Concept
Section titled “Concept”Non-Deterministic Interfaces
Section titled “Non-Deterministic Interfaces”Traditional UX is deterministic: input X always produces output Y. A button always does the same thing. AI UX is probabilistic: input X produces a distribution of possible outputs.
This has direct consequences for your interface:
- Every AI output must be communicated as a suggestion, not a fact
- Interfaces must surface variability, not hide it
- Users need regeneration options (“Generate Again” buttons)
- Confidence indicators belong in the output, not in a debug view
Chat vs. Structured Output
Section titled “Chat vs. Structured Output”Not every AI interaction needs a chat window. The distinction matters:
| Chat Interface | Structured Output | |
|---|---|---|
| Strength | Exploratory, ambiguous, reasoning | Known inputs, validation, high volume |
| Weakness | Inefficient for repetitive tasks | Inflexible for open-ended questions |
| Example | ”Analyze this contract” | Invoice classification with dropdowns |
The rule: “If it’s a form, it should stay a form” (widely cited UX design principle). Forcing a form into a chat interface makes it worse, not better.
Best practice: Every form should be triggerable conversationally, and every agent should be able to hand off to a form. The pure chat paradigm is already outdated for complex workflows.
Progressive Disclosure for AI
Section titled “Progressive Disclosure for AI”AI outputs contain a lot of information. Not all of it belongs on the first screen:
- Level 1: Core result — the answer the user needs
- Level 2: Confidence scores, sources, reasoning summary
- Level 3: Full chain of thought, raw data, alternatives
Microsoft’s Copilot dynamically adapts its interface based on user behavior: power users see more options and detail, casual users see less.
User Onboarding for AI Features
Section titled “User Onboarding for AI Features”The biggest UX danger with AI features: The demo feels magical, reality disappoints. This “Magic Demo Gap” is the most common reason AI features lose adoption after launch.
- Expectation management: Clearly communicate what the AI can and cannot do — before the user sees the first result
- Progressive trust: Start with simple, reliable AI functions. Only offer more complex ones once trust is established
- Leverage existing patterns: Notion introduced AI through slash commands — a pattern users already knew. New interaction patterns require more onboarding effort
- No feature dump: Gradual introduction beats “here are all 12 AI features at once”
For you as a PM: Plan onboarding as part of the AI feature, not an afterthought. The first AI contact determines whether users trust the feature or ignore it.
Loading and Streaming UX
Section titled “Loading and Streaming UX”AI operations are inherently slower than CRUD. Streaming isn’t optional:
- Users perceive streaming responses as 40-60% faster than identical non-streaming responses
- Streaming signals “AI is working” — not just “loading”
- Accessibility concern: Streaming is problematic for screen readers. Plan for accessible alternatives.
Framework
Section titled “Framework”AI UX pattern decision matrix:
| Task | Pattern | Example Product |
|---|---|---|
| Exploration, open-ended questions | Chat + Streaming | ChatGPT, Claude |
| Content creation | Chat + Side-by-side editor | ChatGPT Canvas, Claude Artifacts |
| Inline code assistance | Ghost text / Autocomplete | GitHub Copilot |
| Augmenting existing tools | Slash commands, inline AI | Notion AI (purple highlights) |
| Background intelligence | No visible AI UI | Linear (triage intelligence) |
| Visual generation | Grid + Variations | Midjourney (4-image grid) |
| Code generation + preview | Chat-to-output + live preview | v0 by Vercel |
Design principles for choosing patterns:
- Match modality to task — not every AI needs chat
- Show AI is working — streaming, skeleton screens, progress indicators
- Layer information — progressive disclosure over information overload
- Accommodate variability — “Generate Again,” variants, sliders
- Respect existing workflows — the Linear approach: AI enhances, instead of bolting on a chatbot
Scenario: Notion AI — Inline Over Chat
Section titled “Scenario: Notion AI — Inline Over Chat”Late 2022. Notion has over 30 million users and is one of the most widely used productivity tools in the world. The AI hype is peaking — ChatGPT just launched, and every SaaS product is announcing “AI features.” Notion is under pressure: the waitlist for Notion AI already has over 2 million signups. Users are expecting something.
The central question for the product team: How should AI show up in Notion?
The facts:
- 2M+ waitlist signups for Notion AI before launch
- Notion is an editor-centric product — users spend their time inside documents
- The market default in 2022/2023: standalone chatbot as a separate feature
- User insight from research: “I don’t want to leave my document to talk to an AI”
The decision — three options:
- Standalone AI chat: A separate chat window next to the editor. Users ask questions, copy results back into the document
- Inline AI: AI directly in the editor — via slash commands (
/ai), purple-highlighted, within the existing document workflow. No new interface - Hybrid: Chat sidebar for complex questions, inline for quick actions
Decide
Section titled “Decide”What did Notion decide — and why?
The decision: Option 2 — Inline AI, fully integrated into the editor.
Notion explicitly rejected the “bolt on a chatbot” approach. Instead: “Write with AI” and “Edit with AI” as native functions inside the document. The /ai slash command fits seamlessly into the existing slash command system that Notion users already know. AI-generated content appears with a purple highlight — right where the user is working.
Why it worked — through the lens of the AI UX pattern matrix:
- Modality matched to task: Notion is a writing tool. The task is “create and edit text” — not “have an exploratory conversation with an AI.” Inline AI fits the task; a chatbot would have forced a new interaction pattern
- Existing workflows respected: Users already knew slash commands. No new interaction model required. Default-on for all users, no separate onboarding needed
- Progressive disclosure applied: Level 1 is the core result directly in the document (purple-highlighted). Level 2 appears on hover or click — options to refine, shorten, rephrase. No information overload
- Variability built in: “Try Again,” tone adjustments, and length options give users control over non-deterministic output
The result: Launched in February 2023. Roughly 50% of active users adopted AI features within six months — an exceptionally high adoption rate for a new feature.
The key insight: Users don’t want to leave their document to talk to an AI. Integrating into the existing workflow eliminated the friction that a separate chat interface would have created.
Reflect
Section titled “Reflect”- The interface must match the task, not the technology. Notion didn’t ask “How do we build a chatbot?” — they asked “How does writing get better?” Chat is an interaction pattern, not the default. The slash command integration worked because the team understood the existing task, not because they centered the technology.
- “Bolt on a chatbot” is this generation’s anti-pattern. Notion explicitly rejected it. A separate chat window would have broken the workflow and created a copy-paste loop. The high adoption rate (~50% within six months) shows that low friction beats feature richness.
- Progressive disclosure determines adoption. Purple highlight as Level 1, options on hover as Level 2 — Notion layered AI output instead of showing everything at once. This reduces overwhelm for casual users while still giving power users control.
- Default-on lowers the activation energy. No separate onboarding, no feature toggle, no “enable AI” button. The AI was simply there — in the same slash command menu users already knew. The best AI UX is the one that requires no new mental model.
Sources: Notion AI Launch Announcement (February 2023), Notion AI Waitlist Data (2022), “Notion AI: Lessons in Building AI-First Features” — Lenny’s Podcast (2023), Nielsen Norman Group “AI UX Patterns” (2024)