AI Experience Principles

Nine principles for designing AI-powered product experiences across all integration models.

Nine principles for designing AI-powered product experiences. They apply across all integration models — embedded, copilot, agentic, ambient — regardless of modality or output type.

Augment, don't interrupt

AI integrates — it doesn't demand attention

AI should make existing workflows faster and smarter without restructuring how users work. The best AI assistance is available on demand and invisible when not needed.

Do

Inline rewrite button appears on text selection — user stays in flow

Don't

Chat popup opens mid-task and forces a context switch to a separate window

Gate AI behind explicit invocation — hover, selection, or shortcut. Never use passive interrupts or unsolicited popups.

Operate on artifacts

Work where the user works

AI should manipulate the actual materials a user is engaged with — documents, designs, data. Forcing users to copy-paste content into a chat window externalises the work and breaks the creative flow.

Do

AI-proposed edits preview directly in the document with diff highlighting

Don't

"Copy your listing description and paste it into this chat window"

Build AI as first-class operations on the document model. Side-channel patterns that separate AI from the artifact are always a last resort.

Progressive autonomy

Earn trust before acting independently

Begin conservative. Require explicit approval for novel or consequential actions. Let users consciously grant more autonomy as experience builds — never assume broad permission upfront.

Do

AI drafts replies for review; auto-send only unlocks after user confirms 3 rounds

Don't

Full inbox automation from first interaction with no approval or preview stage

Design explicit escalation: suggest → confirm → automate. Each stage requires conscious user opt-in. Never skip levels.

Make AI intent visible

Plan first, execute second

Users need to understand what the AI is about to do before it acts — especially in agentic flows. Surface the plan, the scope, and the expected outcome. The result alone is not enough.

Do

"I will search 3 sources, extract key data, and draft a report — shall I proceed?"

Don't

AI silently completes 8 steps; user sees only the finished state with no trail

Show AI plans as inspectable, editable step lists before execution. Agentic flows need planning UI, not just output UI.

Build in human review

Automate inverse to stakes

Low-stakes, reversible actions can be automated. High-stakes or irreversible actions must require meaningful human review — not a dismissible confirmation modal.

Do

Diff view for AI document edits with per-line approve / reject controls

Don't

"AI optimised your email campaign" — changes already live when the user reads this

Define stakes thresholds per action type. Match review UX to stakes: notification → approval → explicit sign-off.

Reduce cognitive load

Replace effort, not decisions

AI should reduce the work of getting started and staying on task. Shifting complexity from the task itself to an AI interface (excessive prompting, configuration) is a failure mode.

Do

Smart defaults handle 90% of cases; optional configuration is progressive

Don't

12-parameter prompt form required before AI will attempt the simplest task

Start with zero-config AI. Expose controls only when users hit the limits of defaults. Track time-to-first-value.

Support iteration

Outputs are starting points, not finals

AI outputs are drafts. Design for regeneration, variation, branching, and comparison. A single output with no way to steer or refine is a dead end that trains users to distrust AI results.

Do

"Here are 3 variations — select one to refine further"

Don't

Single AI output, no regenerate, no history, no way to branch or steer

Every AI output needs: copy · regenerate · branch · rate · history. Never ship an AI feature without these five.

Expose confidence

AI is probabilistic — the UI should be too

AI is not deterministic. High-confidence and uncertain outputs should look and behave differently. Presenting all AI outputs with equal visual authority destroys calibration and erodes trust.

Do

Low-confidence answers include source links, hedging language, and a verify nudge

Don't

Facts and hallucinations rendered identically — users cannot distinguish them

Define confidence bands (high / medium / low). Design distinct visual treatments for each. Wire UI to model confidence scores.

Give users steering

Controls beyond the prompt box

Beyond free-text input, provide mechanical controls: tone sliders, style presets, scope constraints, example-based guidance. Users need repeatable, predictable ways to direct AI behavior without re-prompting from scratch.

Do

Tone (Professional ↔ Casual) and Length sliders that persist across all outputs

Don't

Rephrase the prompt and hope for different results — no repeatability, no control

Every major AI feature needs 3+ steering controls beyond text input. Persist controls as user preferences across sessions.