Describe a method for finding AI opportunities inside an existing product without starting from the technology.

Corvasel Labs makes Gallerine, an app that turns a museum's floor plan into a personalized audio tour on a visitor's own phone. Ayodele Kindleworth is the AI PM who owns Gallerine's roadmap. Ottone Wayknell, Corvasel's CEO, wants a camera feature to match a rival's demo. Guillermina Abelove runs visitor experience at the Ashbury Museum of Natural History, one of Gallerine's oldest accounts, and has been quietly routing visitors around the same broken stretch of tour by hand for six years.

How to answer this, stage by stage

Nobody is grading whether you can describe "user research." They are grading whether you can name three real sources you'd actually pull, and the one exact question you'd run them through, instead of reaching for whatever capability just made headlines.

Let's learn

Gallerine reads a museum's floor plan and turns it into a personalized audio tour on a visitor's own phone. Ashbury's version, called the Discovery Path, runs a fixed order: Entrance Hall, then Fossil Hall, then Marine Life, then the East Wing, home to Ashbury's Pre-Columbian ceramics collection. Before Gallerine, visitors carried a paper map and mostly guessed, and Guillermina's team fielded a steady stream of "where's the pottery" questions at the front desk every single day.

This is the workaround the audit was actually built to find. Not a lack of insight from the front desk. A lack of anyone ever turning it into evidence.

Gallerine's own telemetry shows a pattern nobody had pulled together in one place until this quarter. Track completion by stop: Entrance Hall 92 percent, Fossil Hall 87 percent, Marine Life 81 percent. Then the East Wing track: 59 percent. That's not a smooth fade from tour fatigue, it's a 22-point drop where the first three stops only lost about six points each.

For fourteen straight months, Corvasel's support queue held the same recurring ticket, about seventeen a month: "how do I skip ahead to the pottery room." Nobody had ever pulled that count next to the telemetry drop. The two were sitting in different systems, filed by different teams, and neither one was labeled as an AI opportunity, because neither one mentioned AI at all.

Without a method, Corvasel had both failure modes running at once. A camera pilot chasing a rival's demo, and a real ticket pattern nobody had ever pulled.

Here's the turn. The extra tickets and the telemetry drop were never the actual problem. The real problem was what Corvasel did next with that evidence. Three months before this, worried about a rival called Museovo demoing a camera feature called Farview, Corvasel had already shipped its own opt-in "point your phone, get commentary" beta at Ashbury. Only 12 percent of visitors ever turned it on, and the East Wing skip rate barely moved, 41 percent down to 39 percent. Two engineering quarters spent chasing a capability, while the actual friction sat exactly where it had always been.

The whole method in one picture. Three sources that already existed, run through one specific question, before anyone names a capability.

What Ayodele built instead: a standing audit that pulls support tickets, drop-off telemetry, and staff workarounds side by side, then asks one specific question of all three, where is a person doing slow, pattern-based judgment, over and over, that a model could plausibly help with. She also considered running a team ideation workshop, brainstorm a list of AI feature ideas, score them on a rubric, and rejected it. Brainstorming still starts from imagination, not evidence, it just takes longer to arrive at the same technology-first guess.

What it costs at its worst: Ashbury's ceramics wing sits under a state cultural-heritage grant worth $210,000 a year, tied to real visitor engagement with the collection. The year the East Wing skip rate held at 41 percent was the first year Ashbury's own reporting came in under the funder's target, putting that renewal in real question, not because the pottery wasn't worth seeing, but because a fixed-order script kept deciding four in ten visitors would never reach it.

What I would leave alone: the Entrance Hall track doesn't need any of this. It completes at 92 percent, nobody skips it, nobody's filed a ticket about it. Reordering something nobody is struggling with just adds a re-ranking step with no real friction underneath it to fix.

The lesson: a model can do a lot of things. That was never the actual question. The question was always which of those things Ashbury's own visitors, and Ashbury's own staff, were already showing us they needed, whether or not anyone had opened a ticket that used the word "AI."

Now here is the same thing as a story

The short version above is what you'd actually say in the room. Read this one for what six years of a paper insert really cost, and what it would have kept costing.

Every Tuesday morning, Guillermina Abelove stands at Ashbury's front desk for the first hour, because that's when families with the whole day ahead of them show up. She's run visitor experience there for nine years, long enough to read a family's whole visit off two questions. "Are you here mostly for the dinosaurs, or did someone come for the pottery?" Ask that, watch the parents glance at each other, and she already knows which door to point them toward.

Six years ago she started handing pottery-curious visitors a torn paper insert with a shortcut route straight to the East Wing, skipping Fossil Hall and Marine Life. Nobody told her to. She'd simply noticed, season after season, that visitors who came in asking about the ceramics almost never made it there on the museum's normal path, and she had a stack of blank inserts and a pen. It worked. It also never once showed up on anyone's dashboard, because Guillermina wasn't measuring anything. She was just doing her job.

Nobody decided to ignore the gap. It just never got translated into anything a roadmap review would notice.

Gallerine arrived three years ago, and for most visitors it was a genuine improvement. Entrance Hall, then Fossil Hall, then Marine Life, all completing north of 80 percent. The East Wing track quietly sat at 59 percent the whole time, and because nobody had ever built a report that put Guillermina's paper insert next to that number, the two facts, a shortcut staff still handed out by hand and a track a real chunk of visitors never reached, lived in separate worlds.

Then, the same week Ayodele started digging into Ashbury's support queue for an unrelated reason, Corvasel's CEO, Ottone Wayknell, forwarded a press writeup of Museovo's new feature, Farview: point your phone at anything, and it talks you through it, live. Ottone wanted something like it shipped for the next museum conference. Corvasel had already tried something close, three months earlier, an opt-in camera beta built to preempt exactly this kind of ask. Twelve percent of Ashbury's visitors ever turned it on. The East Wing skip rate, the thing everyone actually should have been watching, moved from 41 percent to 39 percent and stalled there.

I want to say the problem was that the camera feature underperformed. It did. But that's not really the story. The real story is that Corvasel had no standing method for deciding what to build next, so every ask got evaluated the same way, whichever executive was loudest, or whichever rival had just demoed something shiny. Nobody had ever asked Ayodele's team to sit down with Ashbury's own tickets, Ashbury's own telemetry, and Ashbury's own front desk before greenlighting anything.

So here's the decision I'd take back. A year before this, when the first camera pilot got approved, the unwritten rule at Corvasel was simple: a strong enough demo, or a loud enough exec, was proof enough on its own. That rule made sense when Corvasel shipped one AI feature a year on a small product, and getting it wrong cost a quiet quarter nobody noticed. It stopped making sense the moment the AI backlog became most of the roadmap, and a wrong guess started costing real engineering time and a state grant's renewal.

The fix had to survive its own first stop, where there's no behavior to read yet. This is what that survival looks like.

I would put a real audit in its place instead. Not a bigger camera pilot. A different shape of evidence: pull the tickets, pull the drop-off numbers, pull Guillermina's paper insert, and ask one question of all three, where is a person doing slow, pattern-based judgment that a model could plausibly help with. The answer wasn't a camera. It was letting Gallerine re-rank the rest of a visitor's tour after each stop, based on how long they actually lingered and what they skipped, the exact judgment Guillermina had been making by ear for six years. Stop one always keeps the default order, there's no signal yet to trust. From stop two on, Gallerine only reorders once its read on interest clears a real confidence bar, otherwise the script stays as written.

Three months after it shipped, the East Wing skip rate fell from 41 percent to 14 percent. Guillermina still keeps a short stack of paper inserts behind the desk, for the handful of visitors who never open the app at all, but she hands out a fraction of what she used to, because the app is now doing by algorithm roughly what she'd been doing by ear.

What I'd tell myself, the day the first camera pilot got the green light off a demo alone: the mistake wasn't chasing a rival's feature. It was never having a standing method that would have made the chase optional, one we could have pointed at Guillermina's own desk, six years before a rival ever built anything.

SPARK, or how to spot a real opportunity before picking a modelNot a script for sounding thorough. SPARK is what forces you to name the one concrete audit, and prove it survives the first day it has no signal to work from.

The line Ayodele drew, and held. Camera recognition might still ship one day, once real evidence says visitors want it.

The recap, one line per letter: situation is a team with no designed way to decide what's worth building, payoff is one shared habit, name the friction before the model, anchor is the standing three-source audit run through one specific question, risk is either extreme costing a real quarter or a real grant, and keep out draws the line at the starting question, never at the eventual capability.

And if you want to be sure it really works, try it somewhere elseSame five letters, a public library instead of a museum, and this time the blind spot isn't a wing of pottery. It's a room of local history nobody's suggested path ever reaches.

Yewtree Public Library runs Stackfinder, an app that recommends which shelf to visit next based on what a patron just checked out or searched for. Evanthia Nkrumah owns its roadmap. Yewtree's leadership formed its picture of Stackfinder the same ungrounded way Corvasel's did: the branch director wanted to announce a "smart shelf" camera kiosk after reading about a neighboring system's rollout, while the collections committee had quietly left Stackfinder out of two straight budget cycles, calling it "too experimental to plan around."

Mapped onto SPARK: the situation is a library forming its picture of Stackfinder from outside headlines and its own avoidance, never from a real patron path. The payoff is the same shared habit, name the friction before naming the fix. The anchor is the same three-source audit run again: checkout records, browse drop-off by section, and a front desk workaround Evanthia found the same way Ayodele found Guillermina's paper insert, staff who already walk genealogy-curious patrons back to the Local History Room by hand, because Stackfinder's own suggested path sends them there dead last, if at all. Patrons who follow Stackfinder's path browse that room only 18 percent of the time. Patrons the front desk personally walks back there browse it 61 percent of the time. Risk runs the same both ways: a camera kiosk copying the neighboring system's demo would cost real budget and likely move nothing, and staying silent leaves patrons quietly getting less local history than the desk already knows how to give them. Keep out draws the same line: no pitch for a bigger vendor contract, no walkthrough of a recommendation model's internals in front of the budget committee.

Same three sources, same one question, a different building entirely. The method didn't change. Only what it found did.

Swap the trigger and it still runs.
Speed: an interviewer caps you at ninety seconds. Skip straight to the anchor, the three sources and the one question.
Cost: no budget for a standing quarterly audit. Run it once, tied to the next real capability request that lands, rather than a fixed schedule, the method still holds, it just isn't refreshed automatically.
The model got better, for real: say a future Gallerine release makes the reorder model noticeably more accurate on its own. The audit still matters, because "is this the room people are actually skipping" is a question about the product's users, not about how good any one model happens to be that quarter.

Where people run it wrong.
They run the audit once, find nothing dramatic, and conclude there's no opportunity anywhere in the product, rather than treating a quiet quarter as real information worth revisiting later.
They let "a competitor just shipped this" skip straight past the audit, on the theory that matching a rival is its own justification.
They stop at the drop-off number alone and never check it against a real workaround, so the audit finds a symptom without ever confirming a person is already doing the judgment work by hand.

How to use it live. Before answering a "how would you find an AI opportunity" question cold, ask yourself one thing: what's the one real workaround, something a person on your own team already does by hand, you'd actually point to. Naming that, not a phrase like "we'd do user research," is usually exactly what the question is listening for.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0