The best AI features don't need a chat window
Useful AI can fit into familiar workflows while keeping people in control.

You’re writing an email and Gmail suggests the rest of your sentence. If it’s useful, you press Tab. If it isn’t, you keep typing. You can also turn Smart Compose off altogether.1
There’s machine learning behind that suggestion, but you don’t have to think about models, prompts or context windows to use it. You’re still just writing an email.
I think that’s a useful way to think about AI in products. The best use of AI isn’t necessarily the feature with the most impressive demo. Sometimes it’s a small improvement to something people already do, in a place where they already know how to do it.
Start with the task
A chat interface is tempting. It’s flexible: you can type almost anything and get something back. But flexibility comes with work for the user. They have to figure out what to ask, explain the context, check the response and then bring it back into whatever they were doing.
For an open-ended problem, that can be a good trade-off. If I’m exploring an idea or trying to understand a complicated topic, a conversation gives me room to change direction and ask follow-up questions.
But many tasks aren’t open-ended. If I want to remove the background from an image in Figma, I don’t need to explain that to an assistant. Figma offers Remove background as an action in the design workflow.2 The intent is already clear from the image I’ve selected and the action I’ve chosen.
The difference isn’t that one product uses AI and the other doesn’t. Both can. The difference is how much work the interface asks of me before I get to the task.
That doesn’t mean every button should become an AI button. The model can be slow, wrong or unnecessary. The point is to choose the interaction around what the user is trying to do, rather than starting with the assumption that AI needs its own chat window.
Familiar doesn’t mean invisible
There’s a risk in taking this idea too far. If AI fits so neatly into a product that people can’t tell when it’s acting, they may also struggle to understand what changed or correct a mistake.
A suggested sentence in Gmail is visible before you accept it. Google also warns that Smart Compose predictions may not be factually correct.1 That visibility matters: the suggestion can be helpful without being treated as an authority.
A photo edit has different consequences from an email sent to a customer, a change to a company policy or a payment. The more consequential or difficult to reverse an action is, the more important it becomes to show what the system is proposing and give the user a meaningful way to review it.
Microsoft’s guidelines for human-AI interaction include making it clear what a system can do and making it easy to dismiss or correct it.3 I think that’s a better goal than making AI invisible: let the technology stay out of the way, while keeping its suggestions and actions clear enough for people to stay in control.
This also applies when AI shapes what we see instead of producing an obvious edit. A recommendation can feel like an ordinary part of an interface while influencing what gets our attention. Less friction isn’t automatically better if it makes that influence harder to notice or question.
Put AI where the work happens
At ArtifactBridge, the product I’m working on, the idea is for agents to propose changes to documents and for people to review and approve them. The AI can help prepare a change, but the person remains responsible for deciding whether it belongs.
That distinction matters to me because company documents aren’t just text to be generated. They can contain decisions and context that other people depend on. A fluent edit can still be wrong or miss something important. A review step gives someone the chance to catch that before accepting it.
It’s also a different way to think about an AI product. Instead of asking someone to open a separate assistant and start from a blank prompt, start with the work already happening: a document that needs updating, an image that needs editing, an email someone is writing. Then ask where a suggestion or action would help, and what the person needs to see before accepting it.
Not every workflow has a clear answer. Sometimes the user wants to explore, and chat is the right tool. Sometimes a direct action is enough. And sometimes the better product decision is to leave the task alone.
The question I keep coming back to is: what is the user already trying to do, and can we make that easier without making them manage the AI?