Skip to content
mh studio Book a free call
← All articles

AI UX

Stop asking people to trust AI. Give them something to review.

Diffs, reasons, replays and a one-second hold: how I designed review moments into two AI concept products.

Mohamed Hassan, Lead Product Designer
· 4 min read

Most AI products I look at ask for trust up front. A prompt box, a sparkle icon, a spinner, and then a result you either accept whole or throw away. When it works, it feels like magic. When it gets one thing wrong, the user learns a lesson that is hard to unlearn: this tool can change things I did not ask it to change.

I think we frame the problem the wrong way. We keep trying to make AI output more trustworthy. That matters, but it is mostly a model problem. The design problem is different: how does a person check the work, quickly, before it touches anything they care about? Answer that well and trust follows. Skip it and no amount of friendly copy will save you.

I explored this in two concept projects on this site: Draftboard, an AI app builder, and Relay, a control room for AI agents that handle refunds and invoices. Different products, same rule underneath.

Make the unit of change small enough to read

The default in AI builders is to regenerate. You ask for a new filter on a table and the whole screen is produced again. Maybe the filter appears. Maybe your column order also changes, a label is reworded and a button moves. The user never asked for those changes, and now they have to spot them.

In Draftboard I rejected regeneration as the default. Every follow-up prompt produces small proposed changes, shown as hunks the way a code review shows them: what was there, what will be there. The user accepts or rejects each one. Two things happen. People can read what changed, so they stop fearing the next prompt. And a rejected hunk becomes a signal: the product knows what the user did not want, so it can stop proposing it.

The trade-off is speed. Reviewing hunks takes longer than accepting a blob. I think that is the right cost for anything that will reach a team on Monday morning. For a throwaway prototype, an "accept all" shortcut is fine. The point is that review is the default, and skipping it is a choice the user makes, not one we make for them.

Explain before you ask

Relay's agents sometimes need a human: a refund over a limit, an invoice that does not match its purchase order. The lazy version of that screen is a question, "Approve refund?", with two buttons. It looks clean. It is also useless, because the person has to leave the screen to find out why the agent wants to do this.

So every approval in Relay leads with the reasoning: what the agent saw, why it decided, how confident it is, and how the team handled similar cases before. The buttons come last. My assumption, and the first one I would test with real ops teams, is that people approve faster and more safely when the evidence sits on the same card as the decision. It is the same rule I use on money screens: terms sit next to claims, never behind an asterisk.

Make the past replayable

When an automated system gets something wrong, the first question is always "what happened?". A log answers that for engineers. Nobody else can read it.

Relay turns each agent run into a replay with three tracks: Saw, Thought and Did. You scrub it like a video. Draftboard does the same for versions: history is a timeline you drag back through, not a chat log you scroll and hope. Rolling back should feel as safe as pressing undo. If undo is scary, people stop experimenting, and an AI tool people are afraid to experiment with is a tool they stop opening.

Show the wait, name the steps

AI work takes time, and a spinner says nothing. In Draftboard, publishing shows four named checks that pass one by one. Waiting is easier when you can see what you are waiting for, and a check that fails tells you exactly where to look. This is also where motion earns its place: each check resolving is a small, honest piece of progress, not decoration.

Add friction where the cost is real

I usually fight friction. But on Relay's mobile app, approving a €480 refund is one tap away from happening by accident in the back of a taxi. So approval is a one-second hold. It is just enough to stop a mistake without adding a second confirmation screen. The rule I use: friction should scale with the cost of being wrong, not with how nervous the room is.

What I would tell a team shipping AI features

  • Decide what the reviewable unit is before you design a single screen. A hunk, a step, a row, a draft.
  • Put the reasoning on the same surface as the decision.
  • Give every action a history that a non-engineer can read and reverse.
  • Name your loading states. "Checking permissions" beats a spinner every time.
  • Measure what people keep after review, not only how fast you generate. A fast result nobody keeps is not fast.

Both products are concepts, so treat the specifics as design hypotheses, not proven results. The principle I am confident about: people do not trust AI because it sounds confident. They trust it because they can check it, correct it and undo it.

Get new articles first.

Leave your email and get one short note when a new article is out. Nothing else, no spam.

One short email when a new article is out. Unsubscribe with one reply. See the Privacy page for details.