The Affordance We Haven't Built Yet
Posted on by Johanna EvansWe took the most human communication primitive we have, stripped out its most essential quality, and called it a feature.
We borrowed the word “conversation” to describe how we interact with AI. But we only borrowed the word. Real conversation has a quality we take for granted: the ability to interject, to redirect, to tap in before the thought is finished. You don’t wait for someone to complete a full monologue before you nod in agreement or offer a contrasting thought. You don’t wait for the driver to reach the wrong destination before you say, take the next left. The affordance is always there: the pause, the interruption, the course correction mid-sentence.
Chat interfaces gave us the aesthetic of it (the back and forth, the message bubbles, the casual tone) without the actual mechanics. And for a while, its absence was fine, because the AI was just answering questions. But now it’s doing things. It’s executing. It’s making decisions in sequence, calling tools, taking steps. And we’re still sitting there, watching it go, unable to say anything until it’s done.
For most of AI’s chat era, the stakes of being unable to interject were low. You asked a question, it answered. If the answer was wrong, you asked again. The cost of waiting was a few seconds and a follow-up prompt. The loop was tight, the damage minimal. But something shifted when AI stopped answering and started acting. Now it’s not composing a response. It’s composing an email, running a search, pulling data, making a decision, moving to the next step. Each tool call is a small commitment. Each step builds on the last. And the further it goes in the wrong direction, the more expensive the correction becomes. Waiting until it’s done is no longer just inefficient. It’s wasteful. Starting over isn’t a minor inconvenience. It’s a tax on every interaction. The interface hasn’t changed but the cost of its missing affordance has compounded enormously.
You know the feeling. Someone is talking and there’s no air in it, no pause, no breath, no opening for you to come in. You have something to say, something that would change the direction entirely. But the door never opens. So you either bulldoze in and break the whole thing, or you wait and let it go somewhere you didn’t want to go. That’s what AI execution feels like right now. You’re present but you’re not participating. The best conversations don’t feel like that. They have a liveness to them, a current running both ways, where each person is shaping what the other says next. That’s the affordance we haven’t cracked for AI. And the longer we wait, the more urgent it becomes. Tasks are getting longer. Agents are getting more capable. The gap between where a task starts and where it lands, without you being able to touch it in between, is only going to grow. We need a word for this. We need a design pattern for it.
I don't want to stop it. I want to steer it.
I don’t want to stop it. I want to steer it. There’s a fundamental difference between those two things that current chat interfaces don’t recognize. A stop is a failure state. Everything halts, context collapses, you start again from zero. But a steer is something else entirely. It’s a light touch. A course correction that preserves momentum. When I see a tool call going in the wrong direction, I don’t need to blow the whole thing up. I need a way to reach in and say, not that path, this one. The affordance I’m describing isn’t an interrupt. It isn’t an approval gate. It’s closer to what happens in the best real-time collaborations, where one person can shift the direction without stopping the work, and the other absorbs it and keeps going. We have the typing indicator. We have the read receipt. We have the live cursor in a shared doc. These are affordances we designed because we understood how people actually work. We haven’t designed the equivalent for conversation yet. The mid-flight steer. The whispered redirect.
The solution doesn’t require a radical reinvention of the interface. It requires something almost insultingly simple: the ability to type while it’s thinking. While it’s executing. While it’s mid-flight. The same way you can speak while someone is still forming their thought. Not to stop them. Not to take over. Just to introduce a new signal that gets folded into where they’re heading. The chat box is already there. The model is already running. The only thing missing is the connection between the two, a way for your words to reach in and adjust the trajectory without pulling the emergency brake.
Every week the tasks get longer. Every month the agents get more capable. And every day we ship interfaces that assume the human is either fully in control or fully out of it, with nothing in between. That gap is where trust erodes. That’s where people override the AI entirely or abandon it mid-task. We didn’t design ourselves into this corner on purpose. But we have to design our way out of it on purpose. The affordance exists in every great conversation you’ve ever had. We just haven’t built it yet.