The whole industry answers every AI product question with a chat window. For clinicians who spend their day in conversation, that's the wrong gift, and a case for ambient help over one more thing to talk to.
My son has recently discovered hospitality. He’ll appear at my elbow in the evening and offer turn-down service, a term he learned from a hotel and deploys with total sincerity, or he’ll bring a glass of water to my room because he figured mine might be running low. Nobody asked him. That’s the entire point. Unasked-for help lands in a different part of the heart than the requested kind, and it lands harder.
I’ve been thinking about that glass of water while watching my industry converge on a single answer to every AI product question: add a chat window. Give the user a text box, a blinking cursor, and an invitation to describe what they need. Chat has become the default gesture of AI product design the way the hamburger menu was once the default gesture of mobile, and I believe it’s wrong for the people I design for.
I design for clinicians. Clinicians talk for a living. A therapist finishes a day holding six or seven hours of the most demanding conversation a human can conduct, and our industry proposes, as a productivity gift, one more entity to talk to. The chat window presumes the user has spare articulation lying around, and my users spend theirs on people who need it more than software does.
The future I keep arguing for in rooms this year is ambient. Prompts and nudges that show you what’s possible. Work performed on your behalf and surfaced for review. The system that notices unsigned notes ahead of a billing deadline and gathers them for you, or drafts the routine reply and holds it for approval. Let’s be the helpful partner. Let’s just do the thing.
A chat window outsources the hard problem to the user: figure out what to ask, supply the context, judge the answer. Ambient help keeps the hard problem with us.
Notice what the ambient posture demands of the design team. We have to know the clinician’s day well enough to anticipate it. We have to earn the right to act by being correct, repeatedly, in small ways before large ones. We have to design the disclosure moment so review feels like a glance instead of an audit, and make undo so effortless that accepting help never feels like surrendering control. That’s a harder brief than a text box, and harder briefs are usually the right ones.
Anticipation has failure modes, and anyone ambushed by an over-eager autocorrect knows them. Done badly, it’s presumption. So the craft starts with tasks where being wrong is cheap and being right is frequent, shows the work before it counts, and widens the system’s autonomy at the pace of demonstrated trust, the same way you would with a new colleague. Chat keeps its place for the genuinely open-ended request. My argument is with chat as the default and the entire strategy. Somewhere in your product there’s a glass of water you could be bringing to the room. Go find it.
The teachable part
For users who talk for a living, don’t add another thing to talk to. Find the help you can deliver unasked, show your work, and grow the system’s autonomy at the pace of earned trust.