Emerging themes

Themes were retained only when supported by more than one independent source. Maturity labels assess the theme in the field, not any organization.

1. AI disclosure is moving from principle to component state

What: Users need recognizable signals for AI-generated or AI-modified content, often linked to layered explanation.

Why relevant: This is the clearest place where ordinary DS assets have measurably changed.

Independent support: IBM’s stable AI label and AI presence variants; SAP’s explainable AI and indicator levels; Atlassian’s Rovo Dialect requires recognizable AI moments.

State: Mature-to-establishing. Production evidence exists, but visual language and granularity are not standardized.

Open questions: When is disclosure useful versus noisy? What must the explanation disclose? How does labeling work when AI and humans co-edit content?

2. Existing components are often extended rather than replaced

What: AI is represented as states, modifiers and behaviors on tables, inputs, forms and content—not only as net-new chat widgets.

Support: Carbon’s AI variants retain ordinary behavior and add label/explainability/revert; SAP applies AI patterns to existing Fiori controls; Microsoft tells agent designers to start with Fluent components.

State: Established practice in leading systems. Chat is a visible addition, but not the whole field.

Open questions: Which changes deserve component APIs versus pattern guidance or content rules?

3. Trust is operationalized through control, provenance and recovery

What: “Trust” becomes editable output, source/context visibility, feedback, dismiss/regenerate, revert/undo, confirmation and recovery.

Support: Carbon revert and explainability; SAP trust/feedback/XAI; Microsoft HAX and agent lifecycle; OpenAI production confirmations; Smashing’s action audit/undo.

State: Mature as principle; uneven as reusable implementation. Agent-specific audit and rollback are emerging.

Open questions: How should risk change confirmation frequency? How can systems prevent approval fatigue? What belongs in DS patterns versus platform security controls?

4. Documentation is becoming an executable context surface

What: DS documentation is being exposed to agents through MCP, manifests, code mappings, structured files and repository instructions.

Support: Carbon MCP, WordPress MCP, Storybook MCP, Figma Code Connect/MCP, zeroheight MCP, Atlassian context engine.

State: Rapidly emerging with real implementations. Transport mechanisms exist; content schemas, evaluation and governance are not standardized.

Examples: Storybook tells agents to verify every property and run story tests; WordPress exposes component docs/examples/tokens because agents cannot reliably consume its rendered Storybook; Figma injects real implementation details and rules from Code Connect.

Open questions: Minimum useful schema? Freshness/versioning? Access control? Which semantics cannot be inferred from code?

5. Component lookup is insufficient; composition intent must be explicit

What: Agents can select valid parts and still produce an incoherent whole. Sources increasingly call for relationships, valid combinations, purpose, context and anti-patterns.

Support: SAP’s Compositional Design System; Atlassian’s semantics/design intent/context files; current Into Design Systems and Reddit discussions; Figma Code Connect’s custom instructions.

State: Emerging. SAP claims production use, but public schemas and cross-organization conventions are scarce.

Open questions: Can composition rules be machine-checked? How much is universal versus product-specific? How are exceptions represented?

6. AI exposes design-system debt

What: Inconsistent names, Figma/code drift, stale docs and underspecified behavior cause agents to invent or misuse components.

Support: Figma states MCP cannot replace judgment or guarantee accurate code from a messy system; community reports repeatedly identify fake component instances and stale guidance; Atlassian says making rules legible to LLMs clarifies them for humans.

State: Strong field observation, supported mostly by practitioner evidence and vendor implementation rationale rather than controlled studies.

Open questions: Which defects most strongly predict agent failure? Can AI-readiness be measured independently of ordinary DS health?

7. Evaluation and constraint are becoming part of the delivery mechanism

What: Grounding alone is insufficient; outputs need story tests, accessibility checks, component-reuse checks, visual comparison and human review.

Support: Storybook MCP includes story tests and anti-hallucination instructions; SGDS explicitly retains accessibility/interaction/content review; Carbon publishes accessibility status; practitioner conference material discusses benchmarks.

State: Emerging but concrete. Individual mechanisms ship; shared success metrics do not.

Open questions: What should a benchmark score: valid imports, semantic fitness, accessibility, appearance, task success, or maintenance cost?

8. Agentic software is a lifecycle, not a chat surface

What: Action-taking agents require expectation setting, activity visibility, scoped permissions, interruption, correction, escalation and recovery across time.

Support: Microsoft’s 2026 agent lifecycle guidance; OpenAI confirmations/watch mode; 2026 permissions research; Smashing/ServiceNow practitioner patterns.

State: Emerging guidance with some production safeguards. It is not yet widely absorbed into formal design systems.

Open questions: How to represent plans and background work accessibly? How should long-running tasks report progress? How are identity, memory and delegated authority surfaced?

9. Generative UI creates a separate composition problem

What: UI can be assembled dynamically around intent and context. This requires guardrails and semantics beyond static screens.

Support: SAP’s Joule Work implementation claim; MCP Apps’ standardized interactive UI delivery; Microsoft’s adaptive card/agent response guidance; current research and practitioner discourse.

State: Experimental-to-emerging. Real products exist, but evidence about usability, governance and maintainability remains limited.

Open questions: What stays stable for learnability? How are generated interfaces tested before display? What is the accessibility model for novel compositions?

10. The field resists a single “AI-native” destination

What: Organizations adapt according to different needs: Carbon formalizes AI presence; Microsoft spans components/content/agent lifecycle; Atlassian builds a context engine; SAP models composition; WordPress exposes documentation.

Support: The Tier 1 set itself, plus zeroheight’s survey showing experimentation far exceeds embedded practice.

State: Established observation. Convergence is around problems and mechanisms, not one target architecture or label.

Open questions: Which adaptations are broadly required, and which depend on whether a system’s users build AI products, use coding agents, or ship autonomous workflows?