Pitch me an AI feature for a product I use daily, and tell me why it has not been built yet.

Daymark is a calendar app. It blocks a person's day into meetings, holds, and travel time, and tries to keep that day honest when something changes underneath it. Bettany Ozeri has used it every workday for three years. Corran Grieveson, running her final interview for a senior AI PM role, pointed at the tab open on her screen and asked her to pitch it one AI feature, then explain why that feature does not exist yet.

How to answer this, stage by stage

Nobody is grading whether Reflow is a good feature. They are grading whether Bettany can turn a fun thirty-second pitch into a real diagnosis the moment the follow-up lands, instead of just saying "and that would be great."

Let's learn

Daymark is a calendar app. It holds a person's meetings, blocks their focus time, and tries to keep the day honest when something changes underneath it.

For most of the time Bettany has used it, when a meeting ran long, Daymark did exactly one thing well: it drew a red line the second two events overlapped. That part took under a second, and it was never wrong. Everything after the red line, deciding what to actually do about the rest of the day, was hers. On a normal overrun, that meant opening three or four events by hand, guessing which ones could slide, and texting someone to ask if twenty past worked. About twelve minutes, on a meeting that only ran forty minutes long.

Twelve minutes of manual triage, every time a forty-minute overrun hits a packed morning.

That is not rare. On roughly one weekday in five, something on her calendar runs over or falls through, and the same twelve minutes of dragging things around follows it.

It is not just her. Daymark's own public feedback board carries a version of this request every quarter, and it has not been getting quieter.

What it costs at its worst: skip straight to the flashy version, an AI that silently rewrites your day with no confirmation at all, and the first time it is wrong, it moves the one meeting that mattered, the client call, the school pickup block, the interview itself, and it does it without asking. A tool like that does not get a second chance. It also demos beautifully in a room full of executives exactly once, then gets turned off by the first real user it embarrasses.

What I would leave alone: the red-line conflict flag itself. Spotting two things booked at the same time needs zero judgment about anyone's priorities. Wiring an AI onto that would be solving a problem that is already solved, for free, by a clock.

The lesson: a feature that has to guess what matters to you has to earn the right to know what matters to you first. You cannot build your way past that with a smarter model. You have to be given the access, one small, provably safe decision at a time.

Now here is the same thing as a story

Say the short version above out loud if an interviewer cuts you off early. Read this one for the two weeks that actually built the answer, before Corran ever asked the question.

Bettany Ozeri has run her own calendar carefully for longer than Daymark has existed. Colour-coded blocks, a hard stop for lunch, a standing rule that nothing gets booked before her daughter's drop-off. When she started using Daymark three years ago, the conflict flag felt like magic for about a month. It caught things she would have missed herself on a busy morning, and it caught them in under a second.

Then the magic went ordinary, the way useful things do. First she noticed the flag, fixed the clash, and moved on without thinking about it. Then fixing the clash became a small ritual, the same order every time: check the next meeting, check the one after, text whoever needed twenty minutes back. Then, somewhere in the second year, she stopped noticing she was doing it at all. It was just what a busy Tuesday cost.

Two Tuesdays before the interview, it happened again. Her 9 o'clock ran to 9:40. She was mid-drag on the third meeting, phone in one hand, laptop in the other, when a colleague leaned past her desk and said, half-joking, "Doesn't that thing do this for you yet?"

She laughed it off. Then she didn't stop thinking about it.

Five points on the same two years, and the gap that never closed sits right in the middle of them.

Over the next two weeks, in the gaps between real work, Bettany did what she would later do out loud in Corran's office. She checked Daymark's own feedback board and found the trend line climbing, not flat. She checked Daymark's own changelog and found Smart Move, the one-tap suggested reschedule, shipped about a year earlier. And buried three posts down in a product blog nobody reads twice, she found a mention of a ninety-day pilot: a small cohort had tried a version of Smart Move with the confirmation tap removed entirely, silent and automatic.

Four honest guesses, written out before any of them got to be the answer.

She laid out all four honestly, the way she'd have to defend them if pushed. Too hard technically didn't sit right, because Daymark had already shipped harder things, multi-calendar merges, travel-time blocking, timezone handling across forty countries. Already exists somewhere she hadn't found turned out to be half true, and led her straight to Smart Move. Real but low priority was possible, but it didn't explain why the pilot that did exist had been quietly shelved rather than expanded. That left the fourth guess standing alone: to move someone's day well, without asking, a model needs to know which meetings are precious and to whom, and Daymark's data model was never built to hold that.

Everything on the left, Daymark already has. Everything on the right, nobody has ever asked her for.

She could guess at the meeting where Smart Move's always-ask rule got decided, even without sitting in it: a small team, a brand-new feature nobody had earned trust for yet, and a reasonable, cautious call that every move needed a tap. Nobody in that room was wrong. But a year of every tap being logged as "accepted" or "rejected," with no record of which accepts were reluctant and which were instant, meant the one signal that could have justified loosening the rule was never collected in the first place.

Before she trusted the fourth guess, she wrote the check she would need to survive Corran asking "how do you know." One alternative would have been to just call it "an execution gap" and move on, since a competitor calendar app had, a few months earlier, demoed a flashy autonomous version at a launch event to visible applause. She turned that explanation down. A demo works once, in a room full of people who will forgive one wrong guess. A calendar has to be right every single ordinary Tuesday, for a person who will not forgive it twice.

One question, three possible answers, and only one of them matched what Bettany actually found.

The evidence test was the one thing that could have proven her wrong instead of just agreeing with her: does a smaller, lower-trust version already exist, and does it stop exactly where real judgment would have to start? It did. Smart Move suggests, and asks, every time, for a year, without exception, even for the ten-minute nudges nobody would ever contest. That is not what "too hard" looks like. That is not what "low priority" looks like either, since a low-priority feature does not usually get a careful, cautious confirmation rule built into it on purpose. That is exactly what a company being careful with something it does not yet have permission to do on its own looks like.

By the time Corran asked the question, forty minutes into the interview, Bettany did not need to invent an answer. She had one, tested against her own best attempt to prove it wrong, and she said it in under four minutes, ending on the one build decision she would actually make: high-confidence moves under fifteen minutes go automatic, everything else still asks, the same shape Smart Move already uses, just with a real threshold behind it instead of "always ask, forever."

Daymark climbed three rungs over two years. The fourth is the one nobody has been given permission to build.

What I would tell myself, the first time that red line ever saved me four minutes: do not mistake the part that got easy for the part that was actually hard. The hard part had not even been attempted yet.

TRACE, so a pitch survives its own follow-up question

Not a script for sounding clever about scheduling. TRACE is what stops "and that would be great" from passing as an answer, the moment an interviewer asks the one question that actually matters: why doesn't it exist yet?

Three things worth saying plainly, since interviewers push here. The rejected alternative: treating a competitor's flashy, no-confirmation demo as proof the gap is "just execution," which Bettany turned down because a demo only has to survive one room once, and a calendar has to survive every ordinary Tuesday for a person who will not forgive a wrong silent move twice. The AI-specific risk, named by name: an autonomous move made with total confidence and no explanation, moving the one meeting that actually mattered, is a silent failure that looks identical whether it's right or wrong until the damage is already done, so any real build attaches a plain "why I moved this" note to every automatic change, and ships only once a move clears a set bar on a labeled eval set of real move-or-don't-move judgments, not a promise to always get it right. And the trade-off, accepted on purpose: acting automatically only on high-confidence, easily reversible moves is slower to feel magical than full autonomy, and worth it, because the alternative's one wrong guess costs the whole relationship, not just one meeting.

And if you want to be sure it really works, try it somewhere else

Same five letters, a library cataloguing desk instead of a calendar app. This time the trust wall isn't about a person's priorities. It's about what happens when a mistake can't be undone at all.

Cartouche is cataloguing software a public library uses to log, classify, and shelve incoming books. Ashling Marrash catalogues donations there five days a week. Pitched the same question, cold, in her own interview, she names AutoShelf: an AI feature that classifies and shelves an incoming donation end to end, no review, the moment it's scanned.

Mapped onto TRACE: timeline is two years of requests for "just let it file itself" on Cartouche's own forum, and a "suggested subject headings, one click to accept" feature that shipped fourteen months ago and has never been extended past a suggestion. Recut is the same four honest slices: too hard, already exists smaller, known but low priority, or too sensitive to hand over fully. Assume nothing rules out "nobody thought of it," the same way a rising request count did for Reflow. Cause candidates land somewhere adjacent but not identical: classifying an ordinary paperback donation is a solved problem, Cartouche already does it well. Deciding, alone, to shelve a fragile local-history pamphlet or a first edition under the wrong heading is a mistake that might never get found and corrected, because almost nobody browses that shelf by accident. Evidence test: the suggested-headings feature stops exactly at rare and local-history items, where a human still makes the final call every time, and sails right through ordinary donations without a second look.

The item furthest from safe to automate is also the one a library can least afford to get wrong.

Swap the trigger and it still runs.
Speed: an interviewer caps the whole answer at ninety seconds. Skip straight to the pitch and the evidence test; the timeline and the four-way recut are what you'd add back if asked to go deeper.
Cost: there's no feedback board to check live. Say what you'd check instead: "I'd pull the last year of support tickets that mention this by hand, and see if the shape of the ask has changed."
The model got better, for real: say Daymark's scheduling engine gets meaningfully smarter at guessing free time. The evidence test still runs exactly the same way, a better engine changes how good the guesses would be, not whether the model has been given permission to act on them alone.

Where people run it wrong.
They stop at "nobody thought of it," because it's flattering to imagine you found a gap everyone else missed.
They find one real reason and stop looking, instead of naming all four honestly before picking.
They treat a competitor's demo as proof the whole problem is solved, without checking whether it holds up under a real, repeat, no-audience Tuesday.

How to use it live. The moment an interviewer asks "why hasn't this been built," buy two seconds by saying the two-part structure out loud first: "let me pitch it, then actually diagnose that, rather than guess." Saying the plan is often the calmest four seconds in the whole answer, and it's free.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0