Problem introduction:

Design systems are primarily built for human use. Designers and developers interpret guidelines, patterns, and documentation by relying on context, experience, and intuition. AI agents, however, do not operate in the same way. They depend on explicit, structured, and machine-readable inputs.

When design tokens, components, and documentation are not clearly structured, consistent, and accessible as structured, machine-readable data, AI agents are unable to reliably interpret the system. Instead, they fill gaps using statistical assumptions based on their training data.

This leads to a growing divergence between intended and implemented design, where inconsistencies accumulate rapidly across prototypes and code. With AI accelerating the pace of generation, this “visual debt” is introduced faster than teams can detect and correct it.

In practice, the problem is not simply about adopting design tokens or using a design system together with AI. It is about whether the design system functions as a reliable, machine-readable contract. Without this, AI agents cannot consistently apply design decisions, and will instead generate outputs that drift away from the system over time.

Other questions:

For whom are we solving the issue of designsystemet+AI?

• for the core design system team?
• for the various organisations own design system team?
• for the design system and token consumers?

For what purpose are we looking into solving for the target group?

• easy/quick prototyping, getting functioning prototypes that look and feel good enough to test various concepts? - more flexibility, allow divergence if DS not enough.
• for quicker production of final products? - need strict contracts and validation
• for governance - need traceablity and control

How should AI access the design system?

• prompt-based, manual context - simple, but need precise description and may more often use assumptions
• structured data/API with tokens and components service with structured contracts?
• stricter tool integration, limits the tool it self and embed into the developer workflow
• hybrid ^^

How should deviation be handled?
How much and when is deviation allowed?

Solutions?

• What solutions should the design system deliver
• Who is responsible? centralised, distributed, ?
• Can guidance replace the need of an AI-product?
• Is it scalable and adjustable - personalised needs

What do we need?

• Sructured metadata? (contracts, skills, json)?
• server-repo?