AI Product Design Principles

Rule 1: Use AI Where It Adds Genuine Value

AI should make users more capable, not more dependent. Apply AI to tasks that are inherently ambiguous, personal, or variable. If the right answer is deterministic, a rule or a search index outperforms a language model every time.

Decision test: Could a rule handle 95% of cases reliably? If yes, start there. Use AI where rules break down or where variety is the point.

Rule 2: Set Honest Expectations

Users form mental models of AI the moment they first interact with it. Over-promise and they will over-trust; under-promise and they won't engage. Aim for calibrated confidence.

When the model is uncertain, the product should reflect it — not smooth uncertainty over with polished prose.

Rule 3: Design for the Wrong Answer

AI will be wrong. Treating failure as a special case guarantees a bad experience. Three tiers need different design responses:

Before/after: A travel assistant showing one itinerary gives no recourse when it lists a closed restaurant. Redesigned with three options, items flagged "not verified in real time," and per-item swapping — the wrong answer becomes a minor edit.

Rule 4: Keep the User in Control

Autonomy is the line between AI that empowers and AI that patronises.

Start with AI suggesting, humans approving. Move toward automation only after strong evidence of trust, and always keep a path back to manual control.

Rule 5: Make Feedback Effortless

Friction is the biggest barrier to feedback. Design the smallest interaction that is still actionable.

A thumbs-down with a category picker ("Wrong / Unhelpful / Too long") beats a star rating for actionability and an open text field for participation. Show impact: "Your feedback improved this for thousands of users."

Rule 6: Personalise Transparently

AI personalisation can feel like magic or surveillance. The difference is transparency. Progressive personalisation: (1) Generic defaults for new users. (2) Observed adaptation from what users actually do. (3) Explicit preferences — users see and edit what the AI knows. (4) Persistent profile reflecting real working style.

The creepiness line: personalisation reveals data the user didn't know was tracked, or narrows options rather than expanding them.

Rule 7: Measure Real Value, Not Usage

High usage is a vanity metric. A user who regenerates a summary five times before abandoning it has high engagement and zero value.

Track guardrail metrics too: error rates, harmful outputs flagged, support tickets. Rising guardrail numbers should halt rollout even when headline metrics look healthy.

Rule 8: Disclose AI-Generated Content

Users have a right to know when they are reading AI-generated content. The label should be clear without disrupting reading flow.

Do not use AI-generated content where users assume human authorship — expert opinions, testimonials, personal messages.

Rule 9: Preserve Human Judgment for High-Stakes Decisions

Some decisions should not be delegated to AI even when accuracy is high. The question is accountability, consent, and the reversibility of the consequence.

Require a human in the loop when a decision has hard-to-reverse consequences, the affected person has a right to human review, or the AI's error mode harms specific groups disproportionately. Automation should increase human capability, not remove accountability.

Rule 10: Build Ethically Throughout the Lifecycle

Ethics reviews at launch are too late. By that point the feature is built and changing core behaviour is expensive.

Watch for dark patterns: AI that manufactures uncertainty or dependency erodes trust faster than it builds short-term metrics.

Further Study

Related: Course home · UX patterns · Glossary