JUSTIN HOWELL

Product, design & technology

Product · 5 min read

The review step is part of the product

When AI proposes something a person must stand behind, checking and correcting it deserve as much design attention as generating it.

During an acceptance walk of Process Studio in September, one extracted value had nearly everything we wanted: a quoted source passage, marked high confidence. What it didn’t have was a way to open the document the passage came from. A reviewer could read the evidence but couldn’t check it in context. We added a link that opens the source with the exact text highlighted and scrolled into view, and it survives a reload.

An AI demo can move very quickly. A document goes in. A structured answer comes out. Someone says, ‘And then the user reviews it.’ That last sentence can conceal half the product.

What exactly are they reviewing? Where do they look when a value seems wrong? What happens to their correction? If the answer is to open the original document and start hunting, the system has handed them a second job along with the result.

My work on document extraction and review at Coralis Bio has made me pay much closer attention to that handoff. Generating a plausible record is one part of the workflow. Giving a scientist the means to check, correct, and approve it is where the product has to earn its place.

Start with the decision someone has to make

Take a familiar example: a system extracts a duration from a procedure. Showing the number is straightforward. A reviewer also needs to know which operation it belongs to, which document version supplied it, and whether the passage actually supports the interpretation. A perfectly formatted value can still be attached to the wrong thing.

That changes the design brief. The screen needs to help someone decide whether this particular value belongs in this particular record. Evidence, context, and the available actions have to meet at that point.

I’d sketch that decision before polishing the generation experience. Put a questionable result in front of the team and follow the work required to resolve it. The awkward parts of that exercise are useful product requirements.

One of my own layouts got this wrong. A single column tried to answer where an item came from and where it went next, so provenance and process flow were tangled together and the flow was buried in actions. We split them: the source now sits under each item’s name, and “Feeds into” has its own column.

Bring the evidence to the question

A citation is useful when it gets someone to the relevant evidence. Linking to a long document leaves a substantial search task behind. A source passage, its surrounding context, and the document revision give the person something more concrete to examine.

The interface also needs to describe its own limits accurately. A link to a likely section and an exact supporting quotation carry different weight. Labeling both as verified would ask the typography to settle a question the system has not answered.

In the Coralis project note, I describe the connection between extracted fields and source passages. The broader design principle is to put the path back to the evidence where the person is deciding what to do. Making document extraction reviewable.

Correction is ordinary work

Review needs more than an approve button. Someone may need to change a value, reject a suggestion, leave a question unresolved, or return when better evidence exists. Those are ordinary working states. Design them with the same care as the successful path.

After a correction, show what changed and what remains pending. If a new extraction runs, decide how it interacts with work someone has already checked. A person should not have to keep a private list of corrections just to defend them from the next refresh.

This concern predates today’s generative tools. The 2019 Guidelines for Human-AI Interaction include helping people correct a system and understand its behavior. I find that a useful reminder to bring established interaction-design questions into AI work early. Guidelines for Human-AI Interaction.

Measure the work left for the person

It is reasonable to want fewer interruptions. A low-consequence suggestion that is easy to undo may need very little ceremony. A change someone must later explain deserves more context. The useful design choice depends on the consequence of being wrong and the effort needed to recover.

When evaluating the workflow, I’d track how long people spend finding evidence, correcting errors, and reaching a decision. Count the cases they cannot resolve. Compare that work with the process the feature is meant to improve. Generation speed alone cannot tell you whether the whole task became easier.

Put the difficult case in the next product review. Ask someone to resolve it using only what the interface provides. Watch where they hesitate. That is where the next design decision belongs.