What happens when your user's main job isn't pressing buttons — it's watching an AI work?
That's not a hypothetical anymore. And most of us aren't remotely ready to design for it.
What happened
Two things are converging right now that deserve our attention. First, there's growing momentum behind the idea that AI agents are fundamentally reshaping SaaS products — moving away from manual, user-driven workflows toward autonomous systems that act, decide, and adapt on their own. Dashboards and form fields give way to agents handling complex tasks end-to-end with minimal human input.
Second — and more practically useful for us — there's serious new thinking on interface patterns designed specifically for AI transparency. The argument is simple: traditional loading spinners are completely inadequate for agentic AI experiences. We need patterns that reveal process, status, and reasoning in real time. Step-by-step progress indicators, decision rationale displays, status feeds that make the black box a little less black.
Put those two trends together and you've got a design problem that's genuinely new.
The bigger picture
For decades, interface design has been about enabling action. We design buttons, forms, sliders, toggles — all instruments of human intent. Click here. Type this. Drag that. The entire discipline is built around the question: how do I help the user complete their task?
But when the AI is the one acting, the interface's job flips entirely. It becomes about observation, understanding, and trust.
Not "how do I help the user do this?" but "how do I help the user understand what the system is doing, why it chose that path, and whether to step in?"
Think about it like this. A traditional SaaS dashboard is a cockpit — you're flying the plane. An agentic AI interface is more like air traffic control. You're monitoring something that's largely autonomous, intervening only when needed.
That's a profound shift. And it demands completely different patterns.
The spinner — that little rotating circle we've relied on for years — assumes the user is waiting for something they explicitly asked for. But in an agentic system, the AI might be doing things the user didn't directly request. It might be making decisions, evaluating options, discarding paths. A spinner tells you none of that. It just says "wait."
What works better? Patterns that show each stage of the AI's process. Displays that explain why the system chose a particular direction. Status feeds that let users scan what's happening without needing to parse every detail.
These aren't polish. They're trust infrastructure.
There's a connected challenge here too — one that's easy to overlook. AI outputs often stream. They arrive word by word, chunk by chunk. Designing stable layouts for streamed content means handling interruptions gracefully, maintaining keyboard navigability as things shift dynamically, and getting your semantic markup right so screen readers aren't lost in the chaos. Most design systems don't account for any of this yet.
We've been designing for static states and user-initiated actions for so long that streaming, autonomous interfaces feel genuinely alien. But they're coming fast.
Tool spotlight
If you're designing any kind of AI agent experience, I'd recommend building what I think of as a confidence layer — a persistent UI element that communicates the system's current state, its certainty level, and what it's about to do next.
This isn't a loading bar. It's more like a heartbeat monitor for the AI.
Even something as simple as a three-state indicator — thinking / acting / waiting for input — gives users an anchor. It tells them the system hasn't stalled, it hasn't gone rogue, and it knows when to hand control back.
You don't need anything fancy to prototype this. A small persistent panel, a colour-coded status dot, a brief text description of the current step. Prototype it early, test it with real users, and watch how dramatically it changes their comfort level with the experience.
The key insight: people don't need to understand every decision the AI makes. They need to feel oriented. They need to know where they are in the process, even when they're not the one driving.
Takeaway
The next time you're designing an AI-powered feature, resist the urge to start with "what can the user do?" Start with "what does the user need to see?"
That single question — what needs to be visible — will reshape your layout, your component hierarchy, and your entire interaction model. It's the difference between designing a tool and designing a window.
We're moving from interfaces that empower action to interfaces that communicate intent. Not our intent. The machine's. And learning to design for that shift is going to be one of the defining skills of the next few years.