Map the four role variants onto a company org chart and show the reporting lines you would choose.

Cravenwood Care Systems builds two things: HomeWatch, a fall-detection wristband sold straight to families, and FacilityWatch, a dashboard sold to senior-living operators like Grayhollow Manor. Both read the same shared model, Sentrion, to decide when someone has fallen. Four AI PM roles touch it: an applied PM on each product surface, plus a platform PM, an infra PM, and a research PM who all work on Sentrion itself. Leonce Kestwick, hired last month as Cravenwood's first Head of AI Product, has one week to put reporting lines on the org chart before the board asks why a resident at Grayhollow Manor sat on the floor for nine minutes before anyone knew.

How to answer this, stage by stage

Nobody is grading whether you can name four job titles. They're grading whether you can draw a chart, defend the top line, and say what breaks first if you draw it the other way.

Let's learn

What happens when two products need two different things from the exact same model?

Cravenwood Care Systems builds HomeWatch, a wristband and bedside sensor that tells a family when someone has fallen, and FacilityWatch, a dashboard that does the same job for senior-living staff watching forty or more residents at once. Both run on Sentrion, one shared model that decides, from movement alone, whether someone just fell.

Nobody sat down and decided this. HomeWatch shipped first, so Sentrion's small team just kept reporting where it always had.

Sentrion-3 had been running at Grayhollow Manor for over a year. Its own numbers were solid: it caught 97 percent of witnessed slow, sliding falls, the kind common in memory care, where a resident settles down a wall or a chair instead of dropping straight down, within 90 seconds. Facility staff trusted it enough to space their own room checks out to every 15 minutes.

Then FacilityWatch needed more. A contract renewal depended on catching slow, sliding falls even better, and Sentrion-4, already live and tested inside HomeWatch, was built to do exactly that.

Here's the turn. FacilityWatch's engineers didn't build a bad model. They built a hybrid: Sentrion-4's improved detection, running underneath the exact confidence cut line that had been tuned for Sentrion-3. Nobody with a view across both products was in a position to say that combination didn't mean what everyone assumed it meant.

Thaisa and Aurelian both depended on the same model. Neither one reported to anyone whose actual job was to referee between them.

What it cost at its worst: across those 11 days, nine residents in Grayhollow Manor's memory-care wing had a real slow, sliding fall. The system caught three of them on its own. The other six were found the slow way, five by staff on their normal round, one by a night aide nine minutes late. Nobody was seriously hurt. That's luck, not design. Cravenwood's FacilityWatch product serves 85 partner facilities. If even one in ten had quietly built a similar local hybrid to hit their own deadline, since Grayhollow's shortcut was never unique, that's roughly eight facilities running a mismatched threshold at any given time, an estimated two dozen missed slow-sliding falls a month across the platform. And the one dashboard leadership actually watched, the blended flag rate across every fall type, moved from 96 percent to 94. Barely a ripple.

What I would leave alone: Grayhollow Manor's own shift scheduler, the tool that decides which aide covers which hallway on a given night, doesn't touch Sentrion at all. It's pure calendar math. FacilityWatch's applied PM can own every reporting line for it herself, no coordination needed, because nothing about it depends on a model two products share.

The lesson: a shared model with no dedicated line isn't ownerless. It's owned by whoever built it first, for exactly as long as it takes their own deadlines to quietly become the whole platform's default, without anyone deciding that on purpose.

Now here is the same thing as a story

The short version above is what you actually say in the room. Read this one for the eighteen months it actually took for one meeting nobody thought mattered to catch up with a hallway in Grayhollow Manor.

Every Tuesday morning, Aurelian Sandholm pulls up the same dashboard before anyone else at Cravenwood is even at their desk. He has run FacilityWatch, the senior-living side of the business, for two years, and he is good at the part nobody thanks him for: he can read Grayhollow Manor's overnight fall log and tell within a glance which alerts are real and which are the wing's old motion sensor twitching again. He was the one who talked Grayhollow's administrator into a six-month trial in the first place, sitting in the staff break room at 6am with a laptop, showing the night shift exactly how an alert would read on their phones.

For the first year, FacilityWatch and Sentrion-3 were good together. Alerts came in fast enough that staff stopped needing to walk every hallway on the hour; they trusted the wristband instead. Grayhollow's own numbers held: 97 percent of witnessed slow, sliding falls caught inside 90 seconds. The facility's fall-related incident reports, the kind that go to the state, dropped by half.

Then the good habit thinned, in three small ways, none of which felt like anything at the time. Grayhollow's administrator stopped asking for the weekly false-alarm report, because there hadn't been a bad one in months. Aurelian stopped sitting in on the night shift's handover meeting, because his other twelve facilities needed him too. And the small monthly check where Cravenwood's platform team re-ran Sentrion's numbers against Grayhollow's real incident log quietly stopped happening, because nobody's job description said it was theirs to keep doing.

Four points on a line, and not one of them looked like a decision while it was happening.

The trigger was almost nothing. Grayhollow's contract came up for renewal, and the administrator mentioned, almost as an aside, that a competitor's system claimed better recall on "the slow ones," the falls where someone eases down a wall instead of dropping. Aurelian didn't have an answer ready. He knew Sentrion-4 existed. He knew it was built to do exactly this. He also knew the renewal decision was six weeks out.

So he did the sensible thing, given what he could see from where he sat. He asked his two engineers to bring Sentrion-4's improved detection into FacilityWatch fast, and to keep the existing confidence cut line, the one already written into Grayhollow's onboarding paperwork, so nothing else about the integration had to change. Nobody told him this was unsafe. There was nobody whose job it was to tell him anything at all. The people who actually understood how Sentrion-4's score distribution differed from Sentrion-3's sat inside HomeWatch's org, three floors away, working on their own roadmap, for their own product's launch date.

The hybrid shipped on a Thursday. For eleven days, nobody at Cravenwood knew anything had changed in a way that mattered. The blended flag rate the dashboard showed the exec team barely moved, 96 percent one week, 94 the next, the kind of dip a busy person waves off.

On the eleventh night, a resident in the memory-care wing eased down against her own doorframe the way Sentrion-4 was specifically built to catch. The score came back 0.54. The alert fires above 0.62, a number written for a model that scored the exact same motion around 0.7. Nothing rang. A night aide, walking her normal round nine minutes later, found the resident sitting against the frame, unhurt, confused, glad to see someone.

Aurelian found out about the hybrid the next morning, not from a dashboard, from the aide's own incident note, forwarded up because it mentioned the wristband hadn't buzzed.

Here's the meeting I would take back, and it isn't the one where Aurelian's team decided to ship the hybrid under deadline pressure. That decision, given what he could see, was reasonable. It's the meeting eighteen months earlier, when FacilityWatch first launched, and someone at Cravenwood decided the two engineers keeping Sentrion running could keep reporting through HomeWatch's VP, the same as always, because building a whole new line for a two-person team felt like more process than the moment called for. Nobody in that room was wrong about the moment. They were wrong about how long "the moment" would stay true.

Run the same eleven days again, with Leonce already sitting above all three of Sentrion's roles. Aurelian still asks for Sentrion-4's improvement, still under the same six-week deadline. But now there's one person whose actual job is to check whether a cut line built for one version means the same thing under another, and the answer comes back before the hybrid ships, not eleven days and one incident note later. The fix costs Aurelian two extra days of waiting. It costs Grayhollow's memory-care wing nothing.

The two designs, side by side: one hands a shared model to whoever built it first and calls that ownership. The other hands it to the one person accountable to everyone who depends on it, and asks that person to slow a deadline down by two days when the model itself is what's changing underneath two products at once.

What I'd tell myself, looking back at that first meeting: nobody ever asks "should this get a boss" at the moment it's true. They ask it the day it's already too late to be cheap. The chart should have moved the day a second product started depending on the same model, not the day a resident spent nine minutes on a floor before anyone knew.

ORDER, for choosing where the lines go before anyone else draws them for you

FLIPS would fit if this were only about Aurelian's trust in the wristband. But the actual question wants a ranked org chart for four fixed roles, and that's ORDER's job.

The middle box is the one that actually happened at Cravenwood, eighteen months after the first box became true.

One of these you fix by moving a name on a slide. The other took a year to grow and takes longer than that to untangle.

Only three names sit under Leonce on purpose. Thaisa and Aurelian keep their own lines into HomeWatch's and FacilityWatch's leadership.

Three things worth stating directly, since the real judgment sits here. The alternative worth naming and rejecting isn't just the other three roles, it's also "let every applied team own its own model work end to end, no shared line at all." That option loses because it's close to what Cravenwood already had, and it's exactly what let a threshold tuned for one model quietly keep running under a different one. The AI-specific failure worth naming is a threshold carried across a model-version change without being retested against that specific version's own score distribution, a cut line isn't a fixed number, it only means what it meant on the model it was tuned against. The guardrail is a shared eval set that any version change has to clear before it ships to any product surface, checked by someone accountable to every surface, not just the one under deadline. And the trade-off is real and accepted on purpose: that shared gate adds real days to a model change that used to ship the same afternoon it was ready. Slower shipping, on purpose, for a model two products already depend on.

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

Same five letters, garbage trucks instead of hallways, and the thing nobody separated this time is a homeowner's stray bottle cap from a city's real contamination fine.

Emberdale Waste Solutions runs Riftscan, one shared model that reads a photo from a truck-mounted camera and scores whether a bin's contents count as contamination, glass in the recycling, food waste in the yard clippings. Two products read the same score: CurbSight, sold to homeowners as a text alert before pickup, and BinPulse, sold to businesses as a monthly compliance report that keeps them off a city's fine list. CurbSight's applied PM, Eldrin Aldercott, built Riftscan's shared scoring pipeline first, for his own product, the same way Thaisa built Sentrion first at Cravenwood.

Different truck, different city, same missing box on the chart.

When Riftscan's underlying model got a version bump, tuned to cut false alerts that were annoying homeowners over a stray bottle cap, Eldrin's team retested CurbSight and shipped. Nobody retested BinPulse, because nobody owned it jointly. Three weeks later, a municipal contract manager, Osmond Nordholm, called to ask why BinPulse's monthly report had stopped flagging contamination that used to trip it, real contamination, at the exact rate that used to earn his city late fees under their contract with Emberdale.

Same chart, mapped onto Emberdale: applied PMs, Eldrin for CurbSight and his counterpart on BinPulse, stay inside their own product lines. Platform, infra, and research AI PMs working on Riftscan all move under one Head of AI Product, for the same reason as Cravenwood: two products, one model, and only a central line is accountable to both when a version change means two different things to two different customers.

Swap the trigger and it still runs.
Speed: an interviewer caps you at ninety seconds. Skip straight to it: applied stays local, platform, infra, and research report central, because a shared model needs one owner accountable to every product that depends on it, not the one that built it first.
Cost: no budget for a dedicated Head of AI Product hire this quarter. Have your most senior platform PM absorb the coordinating role as one explicit line item, with real authority over version changes, rather than leaving the gap open.
The model got better, for real: say Riftscan's new version is measurably more accurate across the board. Keep the shared, per-product eval gate anyway. A better model can still mean something different to a threshold tuned for the old one, and better was never the same claim as calibrated.

Where people run it wrong.
They let whoever built the shared model first keep it by default, and call that ownership.
They add a coordinating title with no real authority over version changes, so the chart says one thing and Thursday says another.
They watch one blended number that hides exactly the segment where the real damage is happening.

How to use it live. Before you rank anyone, ask yourself out loud: "if this model changes underneath two products at once, who is the one person both product teams can go to?" Whoever answers that question honestly is usually the exact rank an interviewer is listening for.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0