Without a Design System, AI Just Helps You Make a Mess Faster

Organisations have finally found the thing they always wanted from software: speed. With AI-assisted coding, teams can spin up tools in days that used to take quarters. Internal apps. Customer portals. Dashboards. Little agents that sit in the workflow and remove a painful step.

The productivity is next level, but the hangover can be a bastard too.

Many corporations expect every team to vibe-code new solutions, it's in KPI's and OCR's everywhere. They quickly end up with fifty tools with fifty different user interfaces and user experiences. Fifty slightly different buttons. Fifty ways to interact, fail, and feel stupid. 

A central design system is how you keep the speed and stop the mess.

Think of it as the product version of a brand playbook: colours, type, spacing, voice. Then go one level up into a component library: buttons, forms, navigation, cards, maps, empty states. One source of truth for how software in your company should look and behave.

The old brand PDF lived in a drawer until someone remembered to open it. A design system wired into your AI production setup does the opposite. Agents pull from it. Devs pull from it. New tools inherit it. Change the system once and the change can travel everywhere.

That way you can take control over how solutions look and behave.

When people with no design background generate interfaces at volume, accessibility is usually the first casualty. Contrast too low. Text too small. Labels missing for screen readers. 

That is not a niche complaint. It leaves behind more people than you might think ad it is also regulatory risk. Universal access rules exist for a reason. An organisation that ships AI tools without a shared standard will break those rules again and again. but it's easy to fix. Guess once what the fix is. Yeah you got it:

A central design system encodes the non-negotiables: contrast, type sizes, focus states, readable structure. Then every new app starts compliant, instead of adding to a backlog of risk. And it's just as fast!

Creating new tools is easy, the hard part is usually adoption. We often tell our clients: "It's an adoption game".

If each department ships its own interface language, every new tool is a fresh learning curve. Search works differently. Filters live in a different place. Status colours mean something else. The mental tax stacks up until people revert to how they did things before with the spreadsheet they trust.

A shared component library eases adoption, a lot. Search looks like search. Cards behave like cards. Maps work the same way across products. Staff only have to learn what the new tool does, not a new way of using a tool.

And it's just as fast.

Ask an unconstrained model to design an interface and it reaches for what it knows: free libraries, stock patterns, soft purple, rounded everything, the same empty dashboard with four big numbers. I wrote about this pattern in AI and the Demise of Great UI.

What Claude Code builds when you give it no design system: the shadcn defaults. Four stat boxes across the top, purple sparkles on anything with AI in the name, a card inside a card inside a card, and a green pulsing dot to say all systems are operational. It works. It also looks like every other tool built this year.

Users may never say "this was made by a model." They feel it. Investors feel it. Customers feel it. And they do not trust it.

What Lovable ships out of the box: a gradient headline, an emoji announcement pill, Inter everywhere, an em dash in the subheading and a row of generated client logos. Fine for a demo. Fatal if it is the face of your company.

Your company has a look and feel. AI will not protect it for you. Without a design system that is yours, and operationalised every generated screen slowly erodes the brand into something generic and fairly untrustworthy. With one, the brand is enforced. 

In a world where anyone can ship software, looking deliberate is a competitive advantage.

You do not need a 200-component cathedral on day one.

Start with the essentials that carry trust and usability: colour, type, spacing, buttons, forms, navigation, empty states. Make them specific to you. Put them where your agents and builders actually work, not only in a Figma file nobody opens. Then make that system mandatory for AI-assisted builds the same way source control is mandatory for code.

Our own homepage annotated into a system: a sand and cream palette, a tiny widely spaced mono eyebrow, a huge serif headline, and an underlined text link where a button would be. Four decisions, and every screen after this inherits them.

The pattern is simple:

1. Humans set the standard.
2. AI builds inside it.
3. The organisation ships fast and recognisable.

What good looks like is not complicated. Two products we designed, each with its own palette, type and components, so nothing on the screen could be mistaken for a template.

Orli, an emotional support platform for schools. A warm palette, cards that belong to the brand, and a dashboard that tells a member of staff where to look first.

Resani, hand-hygiene monitoring for hospitals. One chart, three signals, and a visual language that is unmistakably theirs.

That is the difference between AI in charge and AI assisted.

AI in charge gives you velocity and a mess. AI assisted gives you velocity with a face, a floor for accessibility, and a fighting chance at adoption. It is the same discipline we apply to our own builds, and the reason we put the system in place before the first prototype, not after the tenth.

If your company is about to mandate AI-built apps at scale, do not start with another pilot. Start with the system those pilots will inherit. If you want help putting one together, get in touch.

Otherwise you will get exactly what you asked for: more software, faster.

And a workforce that cannot bear to use it.