Adres
Playground/ docs
AIAI Design

Interaction Models

Six models for integrating AI into products, from embedded assistance to fully autonomous agents.

Interaction Models

Six distinct models for integrating AI into products. Each has different autonomy levels, user expectations, and design requirements. Choose based on task complexity, reversibility, and how much context-switching users can tolerate.

Comparison overview

ModelWhen to useAutonomyKey riskReal example
Embedded AIUser is mid-task, no context switchLow — explicit invokeIntrusive if unsolicitedInline rewrite on selection
Copilot UXGuidance alongside retained user controlLow–Med — suggests onlyPanel ignored over timeSide-panel property insights
Agentic UXMulti-step, complex, well-defined goalsHigh — executes goalsLoss of visibility + controlResearch & report generator
Artifact-CentricRich documents, designs, or structured dataMed — direct manipulationUnexpected scope changesAI edits a listing document
Ambient / ProactiveContext + timing signals are reliableMed–High — autonomousSurveillance feelingPrice drop alert, smart sort
MultimodalModality flexibility is task-criticalVariesModality mismatch errorsVoice search → visual results

Embedded AI — AI inside the interface surface

AI integrated into existing UI surfaces with no context switch. The user never leaves their primary task. Best suited for text transformation, smart fill, semantic search, and contextual actions on selected content.

Inline suggestions

Ghost-text completions appear as user types. Accept with Tab. Works for: search, code, email composition.

Selection actions

AI options surface on text or element selection: Improve, Summarise, Translate, Expand. Zero friction.

Smart autofill

AI completes form fields from context: previous inputs, user profile, document content. Show confidence.

Contextual actions

AI options based on current view or item context, not text. 'Suggest similar listings' on a property page.

Semantic search

Search interprets meaning and intent, not just keywords. Returns conceptual matches, not string matches only.

AI-assisted navigation

AI suggests next steps or related content based on current context and usage patterns. Adaptive pathfinding.

Copilot UX — persistent AI alongside the workflow

A persistent AI panel that lives alongside the primary interface — not a popup, not a separate page. Copilot maintains session-wide context awareness. Users retain full control; the AI offers, never imposes.

Side panel

AI assistance panel that opens alongside content, maintaining full context of the current view and recent actions.

Contextual copilot

AI adapts suggestions and capabilities to what the user is currently viewing or editing. A listing page gets different AI than a dashboard.

Command layer

AI accessible via ⌘K or command palette, accepting natural language commands that operate on the entire interface.

Workspace-aware AI

AI has memory of the session — open items, recent actions, user preferences — so suggestions are always relevant to the current work context.

Smart assistance

Proactive suggestions offered in the panel without interrupting the main flow. User can dismiss, accept, or explore without leaving their task.

Keyboard-first

Copilot features must be fully accessible via keyboard. Tab navigation, arrow key selection, Escape to dismiss. Mouse is optional.

Agentic UX — AI executing multi-step goals autonomously

The most complex AI model. An agent executes sequences of actions — searching, writing, calling APIs, modifying data — with significant autonomy. This demands the most careful design. Visibility, checkpoints, and interruption controls are non-negotiable.

Autonomy spectrum — start at level 1, earn higher levels through demonstrated trust

L1

Suggest

AI proposes, human does

L2

Confirm

AI does, step-by-step approval

L3

Delegate

AI does, checkpoint approval

L4

Automate

AI does, notification only

L5

Autonomous

AI does, no human involvement

Planning visibility

Show the agent's intended steps before execution. Users must be able to inspect and edit the plan before it runs.

Execution transparency

Real-time progress view: which step is running, what it found, what decision it made. Never a spinner with no context.

Approval checkpoints

At defined stakes thresholds, pause and surface a human decision point. Design these as meaningful choices, not dismissible alerts.

Interruption controls

Pause, stop, and undo must be always available and visually prominent. Agents that can't be stopped are a design failure.

Delegation patterns

Clear scope definition: what domains can the agent act in? What data can it access? Explicit permission grants, not broad defaults.

Recovery states

When an agent fails mid-task, show exactly where it stopped, what succeeded, and what needs human completion. Never silent failure.

Artifact-Centric UX — AI operating on user-owned objects

AI operates directly on documents, designs, spreadsheets, timelines, and codebases. The artifact is the interface — AI is an operator on it. This is fundamentally different from chat AI: the user never leaves the artifact.

Artifact types + relevant AI

DocumentsRewrite · Summarise · Expand · Translate · Structure
SpreadsheetsFormula generation · Data cleanup · Chart suggestion
DesignsComponent fit · Spacing critique · Copy improvement
TimelinesTask suggestion · Dependency mapping · Risk analysis
CodebasesRefactor · Explain · Test gen · Bug detection
WorkflowsStep optimisation · Automation suggestion · Gap analysis

Artifact AI principles

Changes visible in context

Always in the artifact — never in a separate result pane

Diff patterns mandatory

Before/after, track changes, per-item accept/reject

Scope always explicit

"You selected 3 paragraphs. I will only affect those."

Preserve direct manipulation

Keyboard shortcuts for accept/reject all must exist

Reversibility guaranteed

Every artifact AI action must be undoable instantly

Ambient & Proactive AI — acting without explicit prompting

The highest-risk model. AI acts on context signals without user trigger. Gets consent and transparency right or it feels like surveillance, not assistance. Applied carefully, it delivers the highest perceived intelligence.

Ambient AI types

Recommendations

Surfaces relevant content or actions based on inferred intent from behavioural signals.

Event-triggered

Activates in response to external events: new listing match, price change, deadline approaching.

Context-aware automation

Performs low-stakes actions when contextual signals are highly reliable and action is easily reversible.

Adaptive defaults

Interface and AI defaults shift based on learned usage patterns without explicit user instruction.

Trust & consent requirements

Show why AI acted

"I flagged this listing because you search this area weekly."

Action log

Persistent record of all ambient AI actions — timestamped and attributed.

Opt-out by default

Consequential actions are opt-in. Notifications are opt-out. Never assume consent.

Batch notifications

Notification fatigue destroys ambient AI trust faster than any other failure mode.

Multimodal AI — spanning input and output modes

AI that accepts and produces content across multiple modalities: text, image, voice, sketch, structured data, and generated UI. Mode-switching must feel natural, not technically imposed.

InputOutputPatternAdres example
VoiceText + visualSpoken search"Find 2-bed in Marina" → listings grid
ImageTextVisual analysisProperty photo → listing description draft
TextImageText-to-visual"Show me a sea-view layout" → generated floor plan
SketchStructured dataIntent extractionDrawn floor plan → structured spec
TextGenerated UIDynamic interface"Compare these 3 listings" → comparison table
Gesture + mapActionSpatial AIMap draw → AI filter by polygon

Mode confirmation

Always show what modality AI interpreted the input as. Don't assume voice = search.

Fallback clarity

When AI can't process a modality, say so immediately with a concrete alternative.

Accuracy signalling

Confidence in modality parsing must be visible — especially for voice and sketch.

Smooth mode switching

Users switch modes mid-interaction. The interface must support this without reset.