Design Systems and AI

The design system is becoming infrastructure for machines

A design system is not a component library.

It is a shared system for making decisions consistently.

It should define principles, patterns, components, behaviours, accessibility requirements, content rules, tokens and the reasoning required to use them correctly.

Increasingly, AI systems need to understand those rules too.

That changes what good documentation means.

Make the rules explicit

Humans are good at filling gaps.

Machines are good at confidently filling gaps with the wrong answer.

"Use this component where appropriate" is weak documentation if nobody has defined what appropriate means.

Document when to use a pattern.

Document when not to use it.

Document allowed states.

Document tokens.

Document behaviour.

Document accessibility requirements.

Document content expectations.

Document relationships between components.

Document the source of truth.

Reduce the number of decisions an AI system has to invent.

Retrieval before invention

If an AI tool repeatedly creates components that already exist, misuses tokens or invents behaviours, do not assume the answer is simply a better prompt.

The system may not have reliable access to the design system.

Give machines structured access to real documentation and real implementation.

Prefer retrieval from the source of truth over model memory.

Design AI workflows that look things up before they create something new.

Evaluate, do not merely verify

Checking that generated output works is not enough.

Ask whether it should exist.

Does it solve the actual problem?

Does it follow the system?

Does it create unnecessary variation?

Is it accessible?

Is the content true?

Is the interaction appropriate?

Does it introduce technical debt?

Could a competent designer explain why the decision was made?

AI generates options.

Professional designers evaluate them.

Build with AI when building answers the question

The gap between design and implementation is narrowing.

Designers can increasingly generate functioning prototypes directly.

Use that capability.

Test responsive behaviour in a browser.

Test interactions as interactions.

Explore real content.

Evaluate states.

Work alongside engineering earlier.

Do not preserve handoff simply because handoff used to define the workflow.

Disclose material AI use

Not every use of AI requires an announcement.

Autocomplete finishing a sentence does not need a governance meeting.

But substantial use may affect confidentiality, ownership, provenance or trust.

If AI materially produces client-facing imagery, research synthesis, strategic recommendations, code or other consequential work, understand whether disclosure is required.

Professionalism includes knowing when the method matters to the person paying for the result.

Protect information

Do not paste sensitive client information into systems you do not understand.

Know the provider.

Know the account settings.

Know the organisation's policy.

Know whether data can be retained or used elsewhere.

Convenience does not remove confidentiality.

The junior problem

The work most easily automated is often the work through which junior designers learned.

If organisations automate all entry-level production, they need another way to create experienced designers.

Give junior people access to real problems.

Bring them into research.

Let them use AI.

Then require them to explain what it produced.

Teach evaluation.

Teach judgement.

Teach why.

A profession that automates its apprenticeship eventually runs out of professionals.

Responsibility remains human

AI can create the work.

It cannot accept professional responsibility for it.

If you approve it, you own the decision.