AI Design Guidelines: 6 Design System Examples for 2026
Six design systems with AI design guidelines, re-checked in October 2026: what each covers now, which moved or were retired, and which ship an MCP server.
Nezar MansourContent Writer

We first published this list of AI design guidelines in January 2025. On 26 October 2026 we opened all six pages again. Five of the six URLs had moved, one set of guidance had been retired, and three of the systems now publish something the original list never mentioned: an MCP server that lets AI coding agents query the design system directly.
That last change is the interesting one. AI design guidelines tell designers how AI features should look and behave for users. The newer work is about the other direction: making the design system itself usable by the agents that now write a lot of the UI.
The six systems, re-checked on 26 October 2026
| System | AI guidance page | Status on 26 Oct 2026 | Last updated (as shown) | Agent-facing resource |
|---|---|---|---|---|
| Carbon | Carbon for AI | Live, moved under Building blocks | 12 Aug 2026 | Carbon MCP |
| Emplifi Soul | AI guidelines: patterns | Live, moved out of "in progress" | 12 Mar 2025 | None found |
| SAP Fiori | Explainable AI | Deprecated in guideline version 1.130 | Not shown | UI5 and Fiori elements MCP servers |
| PatternFly | AI design principles | Live, old URL returns 404 | Not shown (release 6.7.0) | PatternFly MCP |
| ServiceNow Horizon | Designing for AI | Live, moved | Not shown | None found |
| HAX Toolkit | Guidelines for Human-AI Interaction | Live, same URL | Not shown | None found |
"Not shown" means the page carries no update date, only a copyright year. "None found" means we found no MCP server or agent-facing package on the system's own site.
1. Carbon: an AI label every AI component must carry

Carbon for AI gives AI-generated content in IBM products its own visual identity. The page was last updated on 12 August 2026 and now sits under Building blocks, not Guidelines.
What it covers now:
- The AI label: the primary signal that AI is present, and the trigger for an explainability popover.
- Light-inspired styling: glow and gradients reserved for AI-generated or AI-recommended content, backed by colour and style tokens in the main Carbon themes.
- AI variants of core components: checkbox, data table, date picker, dropdown, form, modal, number input, radio button, select, tag, text input and tile. Each one must embed the AI label and the explainability popover.
- Revert to AI: a way back to the AI's original output after a user has edited it.
What's new since January 2025: Carbon MCP, which connects AI clients to Carbon's components, tokens, icons, guidance and code examples. Its overview page was last updated on 28 September 2026.
2. Emplifi Soul: eight interaction patterns, unchanged

Emplifi's Soul design system still documents eight AI interaction patterns. The page shows a last update of 12 March 2025, so its content hasn't changed since our first version. Only its URL did, from an "in progress" section to the main patterns section.
The eight patterns, as Soul names them:
- User-Initiated AI with Custom Input
- User-Initiated AI with Automatic Output
- AI offers recommendations over content
- Continuous AI Assistance
- AI-driven Automation
- Predictive AI Insights
- Conversational AI Interface
- AI-Enhanced Search and Retrieval
Each pattern is written as a short sequence: what the user does, what the AI returns and what the user can then do with the result. That makes it one of the easiest of the six to turn into a checklist for a new AI feature.
3. SAP Fiori: explainable AI retired, a new set in progress
SAP's Explainable AI page, which we covered in detail in 2025, was deprecated in guideline version 1.130. The page now says its content "is no longer aligned with SAP's latest AI design guidance" and that a refreshed set of AI guidelines is being written. The older content is still available by switching to guideline version 1.120.
What SAP has published since is more interesting. Its TechEd recap describes:
- Six AI experience pillars: design for clarity, earn trust, give control, design fairly, boost flow and fail-safe.
- AI UI elements across platforms: in the design system portal, the web, iOS and Android UI kits, and UI5 Web Components.
- MCP servers for SAPUI5 and SAP Fiori elements: open-source servers that give AI agents SAPUI5 and Fiori elements knowledge, including best practices and project context.
We left the 2025 screenshot out because it shows guidance SAP has withdrawn, and SAP's site blocked our automated screenshot tool.
4. PatternFly: AI guidance became a full section

PatternFly, Red Hat's open-source design system, moved its AI guidance to a new /ai/ section. The old /patternfly-ai/ai-guidelines/ URL now returns a 404. The principles are now three short rules:
- Be transparent: make it clear users are interacting with AI, using labels and visual cues.
- Make it personable, but not human: polite and on-voice, but never pretending to be a person.
- Stay within the Red Hat brand and design language: AI features follow the same standards as everything else.
The section has grown around those principles. It now has pages on legal requirements, transparency notices, iconography, colour, chatbot avatars, animation and conversation design.
It also has an AI-assisted development area. The PatternFly MCP server gives coding agents two tools: searchPatternFlyDocs to find components and guidelines, and usePatternFlyDocs to pull full documentation with JSON schemas for React component props. PatternFly calls it an evolving experiment.
5. ServiceNow Horizon: ten questions before you add AI
ServiceNow's Horizon design system moved its Designing for AI page, and the content has changed too. It's now built on four responsible-AI principles: be human-centred, be inclusive, be transparent and be accountable.
The most usable part is a list of ten questions to ask before adding AI to an experience. Three of them:
- Is AI really needed to solve the problem?
- Do people understand they are interacting with AI?
- Is the experience transparent, traceable, and contestable (TTC)?
TTC is the idea to borrow. Users should be able to see how a decision was made, trace it back to the AI, and challenge it. The page links to ServiceNow's 2026 Responsible AI guidelines handbook for more depth. As with SAP, the site blocked our screenshot tool, so this section has no image.
6. HAX Toolkit: 18 guidelines across four moments

The HAX Toolkit still lists 18 guidelines for human-AI interaction, grouped by when they apply. Our 2025 version split them into three even groups of six. The live page uses four groups of different sizes:
| When | Guidelines |
|---|---|
| Initially | 1. Make clear what the system can do. 2. Make clear how well the system can do what it can do. |
| During interaction | 3. Time services based on context. 4. Show contextually relevant information. 5. Match relevant social norms. 6. Mitigate social biases. |
| When wrong | 7. Support efficient invocation. 8. Support efficient dismissal. 9. Support efficient correction. 10. Scope services when in doubt. 11. Make clear why the system did what it did. |
| Over time | 12. Remember recent interactions. 13. Learn from user behaviour. 14. Update and adapt cautiously. 15. Encourage granular feedback. 16. Convey the consequences of user actions. 17. Provide global controls. 18. Notify users about changes. |
The "when wrong" group is the one most product teams skip. It's also the one users remember.
Other design systems with AI guidance (checked 26 October 2026)
Two more public design systems now publish AI guidance worth knowing:
| System | Page | What it covers |
|---|---|---|
| Cloudscape (AWS) | Generative AI patterns | 3 foundations (visual affordance, colours, iconography) and 19 patterns, from output labels and loading states to agent management and user-authorised actions |
| GitLab Pajamas | AI-human interaction | Guidelines on risk, confidence, disclosure, user control, graceful failure and feedback, plus pages for agents and flows, chat, calls to action and feedback |
Cloudscape's visual affordance page is a good place to start. It says the visual treatment should convey "the degree of intervention" AI has in a process, which is more precise than "label everything".
We only list systems whose AI pages we opened on 26 October 2026. Several large systems we checked have no public AI section.
Building your own AI design guidelines
The six systems above suggest a short framework. Each step should produce something you can link to.
- Set your AI principles. Decide whether you always disclose AI use, whether users can switch it off, how much you explain and what happens when it's wrong. PatternFly's three rules show how short this can be.
- Catalogue the interaction patterns you ship. Use Soul's eight patterns as a starting list and delete the ones you don't use. For each, document when to use it, the visual treatment, user controls, error states and accessibility needs.
- Give AI a visual language. A label, a loading state, a confidence indicator where it helps, and tokens for the styling, as Carbon does.
- Define how you explain. Decide what level of explanation each decision gets and who sees which level. ServiceNow's TTC test is a quick check.
- Plan for failure and bias. Write the "when wrong" behaviour first. HAX's guidelines 7 to 11 are a ready checklist.
- Document implementation. Code examples, design files, a decision tree for picking a pattern, and a launch checklist.
Then ask who will read it. A guideline page is written for designers and engineers to read. The agents now generating UI in your codebase need something they can resolve. "Each AI component must embed the AI label and the explainability popover" is a clear sentence for a person. An agent building a form needs to know which component that is, which variant, which token, and that the popover is required, not optional.
That's what Carbon, PatternFly and SAP added with their MCP servers. The guidance stayed the same and gained a second audience. We cover how the design system MCP servers on the market compare, and what an agentic design system needs beyond documentation.
Where Supernova fits
Supernova's documentation keeps AI guidelines next to the components and tokens they refer to, so the AI label rule and the AI label component live on the same page. The same content is available to agents through Supernova's MCP server, with unlimited lookups on every plan, including Free.
To try it with your own system, connect it to Claude Code, Cursor or VS Code or start free.
What to check before your next AI feature ships
Pick one of your shipped AI features and run it past three tests from this list: HAX's "when wrong" group, ServiceNow's TTC questions and Carbon's rule that every AI component carries a label. Whatever fails is the first page of your AI guidelines.
For a wider view of how public systems are preparing for agents, see how 31 public design systems score on AI readiness.
Give your agents a system worth following
See how Supernova documents, syncs, and serves your system in one place.