Design the UI for a feature that is 70 percent confident in its answer.

Thistle Grove Primary Care is a family clinic. Amara Solanke is a nurse practitioner there, managing a panel of about two hundred patients. LabLens is the AI tool that reads each patient's lab history and flags concerning trends, like a slow decline in kidney function, before Amara opens the chart herself.

How to answer this, stage by stage

Nobody's grading whether you know that 70 percent is "medium confidence." They're grading whether your design still works the day it's wrong in either direction.

Let's learn

Here's what a 70 percent case actually needs, and what happens when it doesn't get it.

Before LabLens, Amara reviewed lab trends by pulling up each patient's history and comparing values by eye, catching a slow decline only when she had time to look closely, about fifteen minutes per chart on the ones she flagged for review. LabLens now scans every result overnight and surfaces likely trends in seconds, cutting her review time on flagged charts to about four minutes.

The second box is where a busy morning quietly drops a borderline case. That's exactly the case LabLens should be built to protect.

Here's the turn: the extra speed was never the risk. The risk showed up the day LabLens's launch design put every trend, clear or borderline, into the exact same red banner, with nothing in the screen to tell Amara's team which ones actually deserved a slower look.

Four parts on one card. The raw trend chart is the part that does the real work.

At its worst, a middle-confidence case buried under an identical banner to every confirmed case doesn't just get skipped once. It teaches a clinician to stop trusting the whole pile, confirmed cases included.

What I would leave alone: LabLens's confirmed, high-confidence flags don't need a redesign. They already get opened and acted on fast, because there's little ambiguity left to sit with.

The lesson: the case you're least sure about is the one that most needs its own design, not a smaller font on the same banner as everything else.

Now here is the same thing as a story

The short version above is what you'd say pitching this redesign to Thistle Grove's clinical lead. Read this one for how the near miss actually happened.

Amara Solanke had worked as a nurse practitioner for six years, and could read a lab trend cold before her coffee finished brewing. A colleague two doors down had joined the same panel-review process a year into LabLens's rollout, when flag volume was already climbing.

By her third month, that colleague had started treating the flag pile the way most people treat a full inbox: skim the top few, let the rest wait for a quieter day. Every flag, mild or serious, looked exactly the same, one red banner reading "Trend detected: review recommended."

One skipped banner belonged to a patient whose kidney function had been declining slowly across three visits, a pattern the model had flagged at 70 percent confidence, mixed enough that a careful look would have caught it clearly. It sat unopened for eight weeks until a routine follow-up visit caught the same trend, now further along than it needed to be.

Whichever way the design fails, the raw evidence sitting on the card is what gives someone a chance to catch it anyway.

The redesign that followed gave the 70 percent case its own tier, called "worth a look," showing the two actual lab values that triggered it right on the card. Amara's team could glance at real numbers instead of trusting a badge, and open rates on the middle tier held steady for months after.

The redesign shipped one month after the drop was first noticed. Six months later, the open rate had held, not just recovered briefly.

SPARK, in one screenNot a mockup of every possible screen state. SPARK is what tells you which one decision actually matters.

Three outcomes, not a sliding scale. A clinician can act on three tiers fast; a percentage asks her to do the sorting herself.

Restraint here is what keeps the anchor simple enough for a clinician to act on between patients.

The recap, one line per letter: situation is Amara reading trends by eye today, payoff is a clinician who checks borderline cases as carefully as confirmed ones, anchor is a separate "worth a look" tier with raw evidence on the card, risk is the same evidence catching both over-trust and abandonment, and keep out is no slider and no patient-facing score, not yet.

And if you want to be sure it really works, try it somewhere elseSame five letters, a warehouse's return-fraud flag instead of a lab trend. A different day-it's-wrong breaks the second story.

Falkirk Distribution Center uses ReturnLens, an AI tool that flags customer returns likely to be fraudulent, before a processor opens the package. Marisol Vega is the returns-processing lead who reviews the flagged queue each shift. Mapped onto SPARK: situation is Marisol currently opening every returned package by hand to check it against the order; payoff is a team that spends real inspection time on the returns that deserve it, not an average inspection time that looks efficient on a dashboard.

The anchor carries over almost unchanged: a "worth a look" tier showing the actual evidence, photos of the returned item's condition next to the original order description, with one tap to clear and one tap to hold for manual review. The risk works differently here. At Thistle Grove, the worst case was a missed medical trend. At Falkirk, the worst case is a genuine customer, not a fraud attempt at all, getting held and refunded late over a middle-confidence flag that a five-second look at the photos would have cleared.

Same four parts. "The raw trend chart" becomes "the item photo and order history" on this product.

Swap the trigger and it still runs.
Speed: an interviewer caps you at sixty seconds. Say "a separate 'worth a look' tier, raw evidence on the card, one tap to snooze or escalate," and stop.
Cost: there's no time this sprint to build a full evidence view. Say so honestly, and ship the three-tier labeling alone first, since the tier split matters more than the evidence detail on day one.
The model gets better, for real: if LabLens's accuracy genuinely improves, the middle tier should shrink in volume over time, not disappear. A better model still meets a genuinely ambiguous case sometimes, and that case still deserves its own design.

Where people run it wrong.
They design one screen for "confident" and call anything else an edge case, instead of designing the middle case on purpose.
They show a raw percentage and assume it explains itself, instead of showing the evidence behind it.
They skip testing whether the middle tier still gets opened months later, and find out only after a real miss.

How to use it live. When someone asks you to design for a specific confidence number, ask yourself: what does this case look like if the person using it stops trusting it, or starts trusting it too much? Design so the card survives both.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0