Design the empty state for an AI feature that cannot answer.

Cambrune is a marketplace for secondhand furniture. Camille Osei is a shopper hunting for a couch for a new apartment with a narrow stairwell. Aviel is the chat assistant built into every listing, meant to answer shoppers' questions before they buy.

How to answer this, stage by stage

The interviewer isn't grading whether you can write a polite error message. They're grading whether "I don't know" becomes a dead end or a door.

Let's learn

The screen says one line: "Sorry, I don't have that information." Then nothing.

Aviel is a chat assistant built into Cambrune, a marketplace for secondhand furniture, and it answers shoppers' questions about a listing before they buy. Before it existed, shoppers messaged sellers directly with sizing questions and waited, on average, about 19 hours for a reply. About 1 in 4 gave up on the purchase entirely while waiting.

Four different ways of guessing, all standing in for one missing answer.

With Aviel, most questions get answered in about 2 seconds. Except when the listing itself is missing the measurement Aviel needs, which happens on roughly 1 in 6 listings. In that gap, the original design just said "I don't have that information" and stopped there.

Here's the turn: the missing answer was never really the problem. Nobody expects an app to know everything about a stranger's couch. The problem is what happened right after: nothing told the shopper what to do next, so she had to invent her own plan, alone, usually by guessing.

At its worst, a silent gap costs more than one bad return. Shoppers who guess wrong pay for large-item return shipping themselves under Cambrune's policy, and the ones who don't quietly guess just abandon the cart and remember Aviel as the app that didn't help.

What I would leave alone: the fast, plain answer Aviel gives for the questions it can actually answer doesn't need touching. Bolting a "want to ask a human too?" offer onto every single confident answer would just slow down the cases already working fine.

The lesson: an empty state isn't a missing feature. It's still a decision, made by whoever built it, whether they meant to make one or not.

Now here is the same thing as a story

The short version above is what you'd say out loud in the interview. Read this one for how the actual bad day played out.

The phone in Camille's hand had eleven tabs open, one of them a green couch listed on Cambrune. Her new apartment had a narrow stairwell with one sharp turn, and she needed to know, before spending anything, whether the couch would make it up.

She typed the question straight to Aviel: "Will this fit through a 30-inch stairwell turn?" For most questions, Aviel answers in about 2 seconds, reading the listing's photos and stated dimensions. This time it came back fast, but empty: "Sorry, I don't have that information."

Nothing else appeared on screen. No offer to ask the seller, no note about what Aviel could see, just the one apologetic line and a blinking cursor. Camille looked at the photos herself, decided it "probably" would clear the turn, and bought it.

Same missing fact. One version stops there. The other keeps going.

Camille paid for return freight herself, since large-item returns aren't covered on Cambrune, and posted about the whole thing in a review thread that a few thousand people read. The product team traced the complaint back to one line of copy that had shipped, unchanged, since the app's first version.

The middle box used to be where the whole interaction just stopped.

The redesign kept Camille's exact situation in mind. Same question, same missing stairwell measurement. This time, Aviel answers: "I don't have the stairwell turn radius for this listing, but I can see it's 84 inches long and 38 inches deep. Want me to ask the seller for the turn measurement? I'll message them now." One tap, and the seller replies in about 40 minutes, since the question arrives specific and pre-written instead of vague.

Four parts, always in the same order. Nothing here is optional.

The seller confirms the couch won't clear the turn. Camille skips the purchase. No return, no half day of wedging furniture into a stairwell, no viral thread. The same missing fact, but this time it led somewhere.

The four decisions SPARK is built fromNot a checklist for writing error copy. SPARK is what turns a dead end into a design decision.

Not every "I don't know" gets the same response. The reason behind the gap decides what happens next.

The greyed-out half isn't a failure to ship it. It's a decision not to, yet.

The recap, one line per letter: situation is Camille checking fit by guesswork before Aviel, payoff is trusting an instant answer enough to skip the direct message, anchor is the three-part honest empty state with one specific next step, risk is the guess-and-return cycle if the gap stays silent, and keep out is holding back auto-measurement until it can be trusted.

And if you want to be sure it really works, try it somewhere elseSame five letters, a plant-identification app instead of a furniture marketplace. This time the wrong guess could be dangerous, not just expensive.

Mossveil is an app that identifies plants from a photo, used by home gardeners like Foster Adeyemi to check whether something growing in his yard is safe to touch or pull. Mossveil's original empty state, when a photo was too blurry or the plant too similar to a lookalike species, just said "Not sure," and stopped, exactly the same shape of dead end Aviel used to leave shoppers with.

Mapped onto SPARK: situation is Foster checking an unfamiliar plant before letting his kids near it. Payoff is trusting Mossveil enough to skip a slower, in-person expert visit. Anchor, the new one, says which features Mossveil could match confidently, which ones it couldn't tell apart, and offers a real botanist review request with the photo already attached, answered within about 6 hours on average. Risk is what happens when a lookalike is genuinely dangerous, like a wild parsnip that resembles a harmless carrot top; a flat "not sure" gives someone no reason to be cautious. Keep out, for Mossveil, is refusing to ever guess a single best answer on a known-dangerous lookalike pair, even under pressure to seem more confident than the photo actually supports.

Same four parts. A furniture question and a plant safety question turn out to need the same shape of honesty.

Swap the trigger and it still runs.
Speed: an interviewer caps you at sixty seconds. Say "never leave a can't-answer silent; say what you know, name the gap, offer one real next step," and stop.
Cost: there's no engineering time to build a live seller-messaging handoff right now. Say so, and ship the honest "here's what I know, here's what I don't" half first; even without the handoff button, naming the gap plainly beats a flat apology.
The model gets better, for real: if Aviel's underlying data improves and the true can't-answer rate drops from 1 in 6 listings to 1 in 20, the same anchor still holds, it just fires less often, which is exactly what success looks like.

Where people run it wrong.
They write a nicer apology instead of a next step, and call the tone fix a redesign.
They route every kind of "can't answer" to the same generic contact-support link.
They never log the real misses, so the empty state never actually gets smaller over time.

How to use it live. When someone asks you to design an empty state, ask back: what's the one thing the person should be able to do the moment they see this screen? If the honest answer is "nothing," the empty state isn't finished yet.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0