AI Is a Tool, Not the Goal

We wanted people to upload a document to get a claim approved and paid, or to get reimbursed for an out-of-pocket expense. In either case, we reviewed their document(s), filled in the details, and let them know whether the expense was eligible under their plan.

It was our first AI feature, and the excitement was real. Product, Marketing, and some of our partners fought to put “AI” front and center on the page. It was a milestone, so I understood the enthusiasm.

Before: AI leads, and upload takes two-thirds of the width, promising about 2 minutes against up to 5. Two paths still compete.

Sales decisions aren’t UX decisions

What’s exciting to sell isn’t always exciting, or even necessary, to the person using it. Someone uploading a document isn’t thinking about our milestone. They want this done, and done right. Plus, many people hesitate to let AI touch anything related to their health or money.

AI can help two of our outcomes and hurt the other two

We design for four user outcomes: ease of use, clarity, confidence, and trust.

Used well, AI helps with the first two: less typing, less reading of confusing, jargon-filled rules, fewer mistakes, and less time. Used carelessly, it can seriously undermine the last two.

If people don’t understand what’s going on, or feel pressured, they lose confidence in their decisions and stop trusting us.

Being transparent is about what’s happening, not about the technology

“AI” names the technology (and not even that well… AI is such a broad, overused term that can refer to many different technologies and capabilities). People need to know what happens to their document and trust that it’s being handled right.

In the second version, we used “automated tools,” wording we developed based on guidance from various regulatory bodies and best practices from some of our faves (IBM, NNg, etc.). We also added a small “AI” glimmer icon (a subtle “this is AI” indicator that we would test and refine as we added more AI capabilities).

After: one upload card, “automated tools” in place of “AI”, and manual entry as a link beside the main button.

Then we tested. The AI (automated tool) worked exceedingly well with very few errors, and people used the upload feature more often than not.

Then we listened. Did anyone call or email with questions about the “tools” or how it all works? Did anyone complain? Nothing. Crickets.

Based on additional research (good research is never done after all) and changing user expectations, we dropped all AI/tools wording and the “glimmer” in the latest version and made sure the flow carried the transparency. You see what’s happening at all times. You upload; we tell you we’re processing your document; the details appear; you review them; you edit them if necessary; you submit.

Again. Crickets.

In fintech and healthcare, crickets is exactly what you want. People upload, review, submit, and move on with their lives. Done and done.

Latest: the upload is the page. Manual entry stays easy to find as the secondary action.

Never assume AI will do a perfect job

Whether people upload or type the details themselves, the next step is the same: they review everything and can edit or delete anything before submitting.

When automated tools read the document, people also check a box confirming they’ve reviewed the details. The review screen flags what needs attention, like missing details and possible duplicates, and says plainly that “ready to review” doesn’t mean approved.

The tool does the first draft. The person always gets the final say, and a clear chance to fix anything it got wrong.

Review step: every path ends here. Problems are flagged, everything can be edited or deleted, and the page says plainly that “ready to review” isn’t approval. People confirm they’ve reviewed before they submit.

Ask for feedback. Don’t force it

If the tool didn’t read an item completely or parts are missing, we call it out, and people must edit or delete it before submitting.

If the tool thinks something is a duplicate, people have to edit it, delete it, or mark it “not a duplicate” (which routes it for humans to review).

Whenever someone edits an item or marks it as not a duplicate, we politely ask why they made the change. We tell them we use their answers to improve our tools.

This is where it almost got shady.

Some people wanted to require an answer: if you don’t provide an explanation, you can’t submit. I argued strongly against it.

Forcing people to do something that only helps us isn’t okay. Plenty of people are generous with genuinely helpful, much appreciated feedback.

But in fintech and healthcare, the ideal visit is simple: log in, find what you need or do what you came to do, and get on with your life. Every extra ask adds friction, and you don’t know what the person on the other side is going through.

Maybe this task is one of a dozen to-dos on someone’s list. Maybe they’re trying to get paid back so they can buy groceries. Or maybe they’re nervous about this whole thing, because a denial could mean their child doesn’t get lifesaving medicine or treatment.

Something that’s mildly annoying in another context can set someone off in ours. And we may never know.

They might call and yell about it. They might switch to manual entry every time to avoid the extra step, which costs us money every time. Or they might keep doing it the same way, with their trust in us slightly shaken.

That shaken trust has a cost too. It’s the moment they don’t listen when we suggest things that would help them: maxing out their contributions, investing, getting a debit card for themselves and their family and using it, switching to electronic statements, or adding a beneficiary. Those choices help them, but they also affect our revenue and our partners’.

Make the thing you want people to do the main thing

We want people to upload. It’s usually faster and more accurate for them, and it costs money every time someone enters the details manually.

The original page tried to make that case. Upload took two-thirds of the width and manual entry one-third, and “about 2 minutes” sat against “up to 5 minutes”.

But it still presented two competing paths with too similar visual weight, and it led with AI. The information architecture didn’t match the intent.

Now the page is the upload, and manual entry is the secondary action.

Don’t go shady

Secondary doesn’t mean hidden.

Some people can’t upload: they don’t have a digital copy, or their connection is spotty because they’re filling this out during the boredom of a long layover. Some just don’t want to. And sometimes their account is set up so the automated option isn’t available, or their administrator chose not to use it.

Manual entry stays easy to find and clearly labeled, even though we don’t want them to use it.

Give people every option, and be open about what’s going on.

The lessonS

  • IA 101: Match the visual hierarchy to the information hierarchy, your intent, and the behavior you want.

  • Simplify, clarify, and be honest about what’s happening.

  • Don’t mistake what’s exciting to sell for what the person on the other side of the screen needs.

  • Ask for feedback. Don’t force it.

  • Keep testing and listening, even after launch. Sometimes the best feedback is crickets.

Kimberly Morrow (She/Her)

I’m a builder, creative problem-solver, and AI-native design leader who makes complex decisions clear.

I design experience systems, experiment and prototype in code with AI, lead distributed design teams in regulated fintech and healthcare, and act as a bridge (and sometimes therapist) between Product and Engineering.

Off the clock, I’m an artist, illustrator, and writer who brings a spirit of exploration to everything I do.

See my case studies and connect with me on LinkedIn.

http://www.kimberlymorrow.com