Questions and answers about the best AI design systems

Among public systems, IBM Carbon, Atlassian, Adobe Spectrum 2 and Ant Design rate Yes on design tokens, agent access, docs and guidance. Which of them suits you depends on your stack; for a product that must look like itself, the best AI design system is your own one made AI-ready.

No, and it says so. It is a Simbolik Reference. The public ratings of AI design systems are kept to the systems’ own pages so anyone can check them.

Each rating of an AI design system shows the day its source was opened. The sources are opened again before each release of this Site, and a rating changes when its system’s own page shows something different. The day then changes with it.

A public AI design system, such as IBM Carbon or Ant Design, gives an agent its maker’s look, and the leading ones ship an agent kit of their own. A custom AI-ready design system is built for one company, so the agent follows that company’s brand instead. The 12 public systems are rated in the comparison, and the custom option is its last row. The guide on how to choose a design system for AI agents has the full steps.

It depends on the scope. The deciding things are whether a design system exists yet, how many components and surfaces it covers, the state of its design tokens, how many stacks it ships to and how deep the checks go. The code of a public AI design system is usually free to use, though some agent tools need an account or a paid key. Simbolik publishes no price list. It prices by the client’s size and how complex the work is, and states the price in a one-page proposal after a 20-minute call. The cost and scope guide lists every line.

Then one is built first. That is design work before it is AI work, because an AI-ready design system needs something to write down. Simbolik’s AI Systems page says the same: with no design system yet, it gets built first. The procurement brief has a line for it, where you write "none".

It depends on how much there is to build. With a design system and design tokens already in place, most of the work on an AI-ready design system is turning them into data and written rules. With none, the build comes first. Simbolik states the timeline in its one-page proposal, agreed before any work starts, and a bounded pilot on one surface is a smaller first step than a full build.

For the code, usually yes. Most component code in the public AI design systems compared here is under MIT or Apache 2.0, which allow reuse when the license text travels with it. Fonts, icons and hosted services carry terms of their own, and that is where the limits are. GOV.UK’s typeface, for one, is only for services on service.gov.uk. The licenses guide lists them, and it is not legal advice.

Each of the 48 ratings of the public AI design systems rests on a public page of the system itself: its docs, its official GitHub, its npm page or its maker’s blog. That page is linked from the system’s entry in the directory, with the day it was opened. A third-party list is never proof. A part that could not be found is never guessed at, and the note says where it was looked for.

Agent access is one of the 5 parts of an AI-ready design system, and an MCP server is the usual form. Static files an agent can fetch, such as a JSON list of components, count as Partly on the checklist. Docs come first. Start with readable docs and written guidance, and add a server when agents still pick the wrong parts. The MCP server guide covers what to check before you connect.

With the checklist. It asks 5 questions, one per part, and gives 2 points for Yes, 1 for Partly and 0 for No. A 10 is an AI-ready design system, and every point you miss is a place where an agent invents something. The comparison rates 4 of the 5 parts from public pages, and the checklist asks all 5 because you can see your own code.