Stop turning every answer into text

A lot of AI is pretty good at giving answers. You ask what to do next, and then it comes back with a few possible directions, explains the tradeoffs, tells you what it thinks is safest, and a lot of the time it's right.

But it gives you all of that as a big block of text, so now your job is to pull out the options, compare them, remember which one is risky, work out what is actually clickable, and turn the answer back into a decision.

Cool. But why is that still your job?

That's the bit I think a lot of AI product work is getting wrong. Models can generate the answer easily now, but the UI still has to decide what shape that answer should take.

The answer is not always the interface.

Sometimes the answer should stay as text, of course. But sometimes it should become options, a status, a queue, a button group, or a normal screen with a clear next action.

The lazy version is to dump the answer into chat because chat is already there. The better version is to ask what the user is meant to do with the answer.

Chat is becoming the default

Chat is useful when someone is exploring stuff. If they don't know what they want yet, a conversation can be the right thing because they can ask a messy question, push back, add context, and work their way towards something.

Plenty of product moments are about acting, though. The user needs to approve something, choose between options, or work out why something failed. A long explanation might feel impressive for a second, but they still have to turn it into choices.

That is the bit the product should be designing.

A good answer still isn't enough

Picture a support or ops tool that spots a failed sync.

The assistant says:

The sync failed because two mapped fields are missing, the API token is close to expiry, and three records need manual review. You can either retry after mapping the fields, ask an admin to refresh access, or skip the three records and continue with the rest.

That is actually pretty good. It's pointed out everything blocking the next attempt, but the user still has to turn it into a set of decisions:

• what is blocking the next attempt?
• which option is safest?
• what can they do without admin access?
• what happens if they skip the three records?

That should probably not be a paragraph. It should be a small decision surface with enough context to make each action safe.

Fix mapping

Ask admin

Skip 3 records

Retry sync

The model has identified the blockers, and the product already has the structured data. So why make the user reconstruct that shape from prose?

This is where "AI UX" gets a bit lazy. Chat is not the problem, but it can become the place teams stop thinking while the actual product decision is still sitting there: what should this answer become?

Answer as text

Next move visible

This is where the issue stops being just a UX taste thing. When everything stays as text, the user has to parse it, but the product often loses structure too. If the assistant gives them three options in a paragraph, the system may not know which one they chose, which one they ignored, which action failed, or what should be remembered later.

When the answer becomes options, controls, states, or a proper surface, the user gets an easier next move and the product gets cleaner signal.

That does not mean the UI is easier to make. It is probably harder, because now you have to think about the options, labels, defaults, safe action, fallback, and what happens if the user changes their mind.

A travel app should not always make someone search, read, and interpret. If the thing they need is your gate changed, boarding starts in 18 minutes, and it is a nine-minute walk, that wants to be a status surface. Maybe a card. Maybe a notification. Maybe a map cue. Not a long answer they have to decode while walking.

The team ships the mode that is easiest to produce, not the shape that makes the user's next move easier.

When text works

Text works well when the user is exploring, learning, writing, comparing ideas, or trying to understand why something happened. Sometimes a paragraph is exactly what they need.

Once the user needs to act, the surface should change with them. That might mean options for a decision, a scannable comparison, a clear status, or a safe way to recover. If the next step is genuinely conversational, keep the chat.

Ask what work the user still has to do after the answer arrives.

Pick one AI or assistant surface in your product, find an answer that currently arrives as text, and ask what the user is meant to do next.

If they are meant to choose, compare, approve, fix, recover or continue, work out what shape would make that next move easier. Dumping the answer into text is easier at the start, but it is often harder for everyone after that.