Your AI assistant just booked a flight, cancelled a meeting, and rescheduled your dentist appointment. It did all of this while you were making coffee.
Helpful? Maybe. Unsettling? Definitely. And this is the design challenge that's creeping up on us — most of us aren't ready for it.
What Happened
Agentic AI is the phrase you're going to hear a lot this year. Unlike the chatbots and image generators we've been tinkering with, agentic AI systems don't just respond to prompts — they take actions. They browse, book, buy, send, and decide on your behalf. And that shift from "tool" to "agent" creates a design problem we haven't really faced before.
The question isn't whether to make these systems transparent. It's when.
A new framework doing the rounds tackles exactly this. The argument: most AI transparency falls into two traps. Either the system is a black box — it does things and you've no idea why — or it drowns you in explanations, surfacing every micro-decision until you stop reading any of them. Neither works.
The better approach? Map every decision point in the AI's workflow, then identify the specific moments where surfacing the system's reasoning actually matters. Not every decision. Not no decisions. The right ones.
The Bigger Picture
This is a genuinely new design challenge, and it's worth taking seriously.
For years, we've designed interfaces where the user is always in control. Click a button, something happens. Fill in a form, submit it, see a result. The feedback loop is tight and predictable. Agentic AI breaks that loop. The system acts autonomously, sometimes across multiple steps, and the user might not even be watching.
So where do you put the "here's what I did and why" moment?
Too early, and you're asking permission for every minor action — which defeats the purpose of an agent. Too late, and the user discovers the AI made a poor decision after the damage is done. Try to explain everything, and you get notification fatigue. People tune it out entirely.
The smart move is to think of transparency as a design material, not a toggle. You wouldn't apply the same typeface at the same weight to every element on a page. Transparency works the same way — it needs hierarchy.
High-stakes decisions (spending money, sharing personal data, contacting someone on your behalf) need upfront clarity. Low-stakes decisions (sorting your inbox, suggesting a playlist) probably don't. The grey area in between? That's where the interesting design work lives.
And here's the bit that makes this urgent: we're not talking about a hypothetical future. Agentic features are already shipping in products designers work on every day. If you're designing anything with AI that takes actions — even simple ones like auto-sending a reply or scheduling a task — you're already in this territory.
A Difficult Footnote
While we're thinking about transparency in our interfaces, it's worth acknowledging what's happening behind them. This week, reporting revealed that tens of thousands of gig workers — many of them credentialed professionals like artists, journalists, and teachers — are scraping personal social media profiles, harvesting copyrighted material, and transcribing deeply uncomfortable content to train AI systems. Many described the work as morally distressing. One artist spoke about the guilt of "contributing directly to the automation of my hopes and dreams."
I don't raise this to be preachy. But if we're going to design transparency into AI products, we should at least be honest about the layers of opacity underneath them. The infrastructure powering these tools isn't clean or simple, and as designers we're closer to it than most people in our organisations.
Transparency isn't just a UI pattern. It's an ethical position.
Tool Spotlight: Decision-Point Mapping
If you're working on any AI-powered feature, here's a practical technique worth trying. Before you design a single screen, map the AI's decision points — every moment where the system makes a choice or takes an action.
Then sort them by stakes. What happens if the AI gets this wrong? If the answer is "not much," the system can probably act quietly. If the answer is "the user loses money, time, or trust" — that's your transparency moment. That's where you need to surface reasoning, offer control, or at minimum provide a clear undo.
It's not complicated. A simple flowchart or decision tree will do. But doing this exercise before you start designing screens will save you from building an interface that either over-explains or hides too much.
This works for any designer touching AI features — you don't need to be building a full autonomous agent. Even an AI that auto-suggests email replies has decision points worth mapping.
Takeaway
Agentic AI is shifting our job from designing interactions to designing trust. And trust isn't built by showing everything or hiding everything — it's built by knowing which moments matter.
Pick one AI feature in your current project. Map its decision points. Ask yourself: if this goes wrong, would the user know why? If the answer is no, that's your first transparency moment to design for.