Design the question, not just the answer: AI health interfaces that don't hand you a blank chat box

Walk into an exam room and the doctor does not sit you down and ask you to state your own diagnosis. They ask the next question, and the one after that, narrowing from your vague sense that something is off toward something you can act on. That sequence is the value. It is also the exact thing a blank chat box refuses to do: a health product that opens on an empty prompt and a friendly ask me anything hands the patient the clinician's job and keeps the easy part for itself.

The blank chat box has become the default front door for AI in health products. For most of what people come to do, it is the wrong one. Conversation isn't the problem. The empty prompt is. It asks a person to name what they need before the product has helped them work out what that is, and in health that is often the one thing they can't do. A good interface does the opposite job. It designs the question the user should be asking, and then the answer follows: it lays out what is worth asking, shows the next useful move, and keeps plain language on hand for the moments a user does know exactly what they want. More conversational does not mean more usable.

In brief

• A blank "ask me anything" box hands the patient the clinician's job; a good interface asks the next question instead.
• In the research, menu-based interfaces beat chatbots on cognitive load and sense of control, even though chat feels more natural.
• Keep free text for stating intent and reading back an answer; carry the exploring and comparing on visible structure.

If the open chat box is the wrong front door, which is the case I made in the companion piece to this one, then the constructive question is the one left on the table. What does the right front door look like?

Isn't a chat box the most natural way in?

It feels natural, and we mistake that for evidence that it works. General-purpose assistants have trained everyone to expect a box: something like 230 million people now bring a health question to ChatGPT every week. Your product is measured against that muscle memory, and the tempting conclusion is that giving people the same box will make your interface feel familiar.

But familiar and easy to use are different things. The clinician from the opening is the counter-example: the value was in being asked the next question, not in being handed a blank page and told to describe what was wrong. Familiarity makes the blank box feel friendly. It doesn't make it the right tool for a person who doesn't yet know what to ask.

Is conversation really easier for people?

The evidence says often not. A 2021 study in the journal Computers in Human Behavior compared chatbot interfaces with plain menu-based ones and found the chatbots produced lower perceived autonomy, meaning the sense of being in control of your own choices, and higher cognitive load, meaning the mental effort of holding several things in your head at once. Satisfaction was lower too. The work is grounded in self-determination theory, the well-tested idea that people engage more when they feel capable and in control. A good menu, it turns out, can offer more of both than an open conversation.

A 2026 arXiv paper called "The Keyhole Effect" ran a controlled experiment and reached the same result: chatbots produced higher cognitive load and lower perceived autonomy than menu-based interfaces, even while feeling natural to use. Its authors explain why chat struggles once a task has more than one step. Each new message pushes the previous one out of view, so the user loses their place and their memory of what came before. The details they have to track pile up faster than working memory can hold them. Chat is good for two things, the paper argues: stating what you want, and reading back a summary. It is weak at exploring options, comparing them side by side, and drilling into the reasoning behind an answer. In a health product, that middle part is most of the job.

What does designing the question look like in a health product?

The same paper proposes a set of hybrid patterns that keep natural language for stating intent but build visible structure around it. Several translate cleanly to health software.

Mise en place, borrowed from the way a cook lays every prepped ingredient out before service, means putting the relevant pieces in front of the user before they type a word: their recent readings and the lab value that moved since last time. They can see what is there to ask about instead of guessing.

A state rail is a panel that keeps the current context, what the user is looking at and what they have selected, fixed on screen so it doesn't scroll away with the conversation. Semantic zoom lets someone move from an overview down to a single data point and back without losing their place, the way a map lets you zoom in without forgetting where you are.

Generative UI is the interface assembling the right controls on the fly to fit what the user is doing, a chart when the answer is a trend, a short form when it needs more input, rather than returning another paragraph of prose. This is a shipping idea, not a whiteboard one. Vercel open-sourced a library called json-render in January 2026 that lets a model build an interface from an approved catalogue of components, and Thesys sells an API, C1, that returns interactive UI instead of text. The deeper version of this argument, and how it applies to metabolic health in particular, is a longer piece in its own right.

If you are handing this to a team, a couple of commitments carry most of the weight. When you already know something useful about the user, never open on a blank prompt; lead with what you can show them. And for every answer the product gives, offer the next question worth asking, so nobody is left staring at a cursor wondering what to do with it.

Where does the chat box still belong?

It has a real place, and ripping it out would be its own mistake. Free text is the best tool we have for letting someone state intent in their own words, especially the messy, specific intent a menu could never anticipate: "why does my morning reading spike when I eat the same breakfast every day?" It is also good for reading back a synthesised answer once the product has gathered what it needs. Keep it for those. The mistake is making the conversation carry the whole interaction, from the first blank second onward, when most of it would travel better on structure.

The teams that earn the trust of a nervous user will not be the ones with the most natural-sounding chat. They will be the ones whose product already knew which question to ask. That is the whole difference, and it is a design choice you make long before anyone opens the app.