Normal software is honest when it fails. A calculator doesn’t give you a nearly right answer, and a form with a missing field refuses to submit. AI features fail politely. They hand you a confident answer that is slightly wrong, and nothing on the screen looks different.
Over nearly three years I’ve designed for three AI features like that. One reads scanned paper forms, one answers questions from a company’s own data, and one is a meal log that guesses what’s on your plate from a photo. They looked nothing alike. They taught me the same thing, and it became one of my working rules: with AI, design the fix before the guess.
Most of it comes down to three questions: where the model is unsure, where the answer came from, and how fast someone can fix it. The examples below are made up, so they stay clear of client work; the real stories are in the case studies linked at the end.
Show where it’s unsure
Say an expense app reads a photo of a receipt. It fills in the shop, the date and the total. The shop and the date came out clean; the total was smudged.
If the app shows all three fields the same way, people do one of two things. They trust everything and approve a wrong total, or they stop trusting the app and check every field by hand, which removes the reason it exists.
The better screen flags the one field the model is unsure about, keeps the photo in view, and highlights where on the photo that number came from. The model doesn’t get the last word here. Its job is to point a person at the one number worth checking.
One warning. If almost every field gets flagged, the flag stops meaning anything. Before you ship it, decide what you’ll count, for example how many flagged fields turned out to be real mistakes.
Show where the answer came from
Say a company adds an assistant to its HR portal. Someone asks how many leave days carry over to next year. A good answer gives the number and names its source, “Leave policy, section 4”, as a link. Now the person can check it in one click.
The same assistant needs a plain answer for questions outside its data. “I can only answer from company policies” is better than a guess from the open web. A few starter questions in the empty state show people what it’s good at before they type.
None of this needs the word AI on the screen. As NN/g puts it, “Powered by AI” on its own doesn’t tell anyone what the feature does for them.
Make the fix cheaper than starting over
If fixing a wrong guess takes longer than doing the job by hand, people go back to doing it by hand. It helps to split the fix by the kind of mistake.
A wrong amount: give a slider to drag or a stepper to tap, not a number to retype.
A missing or wrong item: let people say it in words, the way they’d tell a friend: “add two onions”.
A generated list they mostly like: let them pin what they want to keep, and regenerate only the rest.
Ask before it acts
Booking a meeting or sending an email carries more risk than answering a question. Before an assistant changes anything, show what it is about to do and let the person confirm or edit it first.
Where these patterns go wrong
Each one has a cost.
Flags wear out. Watch how often a flag points at a real mistake.
Mixed fixes need names. A slider for amounts and a text box for items is fast, but only if each control says what it changes.
A scoped assistant disappoints people who expect a general chatbot. Saying no is the price of not making things up.
A quick map by risk
If a wrong guess costs | The AI’s job | Show | Let people fix it by |
|---|---|---|---|
A lot: one wrong field matters | Fill in, then point a person at what to check | Which fields it’s unsure of, beside the source | Editing in place with the source in view; nothing approved on its own |
Some: a wrong number in a report | Answer from known data | The source of every answer, and what it can’t answer | Opening the source in one click |
Little: a list, a plan, a log | Save typing | The result, updating as they change it | Sliders for amounts, words for items, pins for what to keep |
Related case studies: the form reader in the KYC registry, the assistant in the IR portal and the meal log in Burpy.