Explain why responsible AI work is a product requirement and not a compliance checkbox.

Look, I'll just tell you straight how I'd answer this one out loud. Marrow & Finch is a retail chain. Yusuf Demola is the PM who owns their AI recommendation engine, the thing that decides what shows up under "you might also like" on every product page.

How to answer this, stage by stageSix moves. Say it plainly, the way you'd actually say it in the room, not like you're reading a policy.

Let's learn

Here's the object that actually tells this story: a printed one-pager Yusuf still keeps, folded in his laptop bag, from the meeting where legal signed off on the recommendation engine.

Marrow & Finch's engine looks at what a shopper's browsed and bought, and quietly decides what to push at them next, on every product page, in every email.

At launch, the engine went through a real fairness review. Legal signed off. Everyone moved on to the next quarter's roadmap.

At its worst: the engine, left alone for two quarters, began pushing higher-margin, harder-to-return items disproportionately at customers whose order history flagged them as recently bereaved, a pattern nobody designed and nobody was watching for.

What I would leave alone: the legal sign-off itself is genuinely good practice, and I wouldn't skip it. The problem was never the review. It was treating the review as a finish line instead of the first lap.

The lesson: a review that only happens once is telling you about a version of the model that may not exist anymore by the time anyone reads the report.

Now here is the same thing as a story

The short version above is what I'd actually say in an interview. Read this one if you want the fuller picture of how it played out.

Yusuf's good at reading a launch review room. He can tell within the first five minutes whether legal's going to push back hard or wave something through, just from how many questions get asked in the first slide.

Three of these four boxes happened in one week. The fourth one, the quiet one, ran for two full quarters.

The recommendation engine's launch review went smoothly. Legal asked good questions, the team had good answers, and the sign-off memo went into a shared drive nobody would open again for a long while.

Same scale, same weights. The only difference is whether anyone steps up to it again after the first reading.

There was no single bad moment. Nobody made a careless call. The model just kept retraining on fresh purchase data every few weeks, the way it was designed to, and the version running in June had drifted a fair distance from the version legal actually reviewed in January.

One milestone got a signature. The next two happened with nobody in the room at all.

A customer service lead noticed it first, not through any dashboard, but because three separate complaints in one week mentioned the same thing: pushy upsells right after a customer had marked an order as a gift for someone who'd passed away.

With a recurring review and a named owner in place, the same kind of drift now gets caught inside a monthly check, not a customer service inbox: a quarter-over-quarter comparison flags when the model's recommendations start skewing hard by a sensitive purchase signal, and Yusuf's team has an actual kill switch to pull the behavior back before it reaches month three.

The launch review had none of these four. It had a signature, which is a different thing entirely.

The old process asked "did this pass review." The new one asks "is this still passing, this month, on the version actually running."

I told the team the model was approved because that's what the meeting produced, a clean approval. It took watching a customer service inbox catch a real harm two quarters later, something no dashboard was built to catch, to see that "approved" was a fact about January, not a fact about June.

PICK, said out loudNot a compliance framework. PICK is what forces you to actually commit to a position instead of listing pros and cons forever.

Both panels look like reasonable boxes on a slide. Only one of them is actually cheap.

The recap, one line per letter: position is responsible AI as a roadmap requirement, impact is the team's one bad day versus the customer's two bad quarters, cost asymmetry is 340,000 dollars against 60,000, and kill criteria is a metric with no owner that's never actually worked anywhere I've seen.

And if you want to be sure it really works, try it somewhere elseSame four letters, a veterinary clinic instead of a storefront. A completely different industry, and here the cheap-looking miss and the expensive one are almost reversed.

Thornbell Veterinary uses an AI scribe that listens to an exam room visit and drafts the medical notes and treatment summary. Petra Solvang manages the tool, and vets review the draft on a wall-mounted terminal between appointments.

Mapped onto PICK: position is that the scribe's accuracy needs a recurring owner too, not just a one-time clinical validation study before rollout. Impact is that a vet catches an obviously wrong note in the room, in seconds, but a subtly wrong dosage note that slips through review can affect a treatment plan weeks later, once the vet's trust in the draft has grown and their own re-reading has thinned out. Cost asymmetry says the validation study was cheap and visible, a few weeks of work, a clean report. The ongoing risk is a quietly worsening draft accuracy on rare conditions, the kind the original study, built on common cases, never had enough examples to catch. Kill criteria: if the clinic had unlimited vet time to double-check every note forever, none of this would matter. They don't, so the recurring check is the only thing standing between "validated once" and "still accurate now."

Swap "legal sign-off" for "clinical validation study" and this same list still describes Thornbell's scribe exactly.

Swap the trigger and it still runs.
Speed: an interviewer caps you at sixty seconds. Say "a compliance review is a photo of one day, a product requirement is a habit that keeps checking," and stop.
Cost: there's no budget for a dedicated fairness team this quarter. Say so honestly, and start with one existing PM adding a monthly ten-minute check to their own recurring rituals, since a small habit beats a big process nobody funds.
The model gets better, for real: if the recommendation engine's overall click-through improves, that's still not proof it's fair, a model can get better on average while getting worse for one specific group nobody's watching.

Where people run it wrong.
They treat the launch review as the hard part and the ongoing check as optional extra credit, when it's actually the reverse.
They build a metric with no owner, so it sits on a dashboard nobody's job depends on and quietly stops getting checked.
They wait for a customer complaint to be the detection mechanism, which means the harm already happened before anyone official noticed.

How to use it live. If someone asks you this question, don't reach for the word "ethics." Reach for a calendar. Ask yourself: does this check have a date on it that repeats, and a name attached to who does it. If the honest answer is "it happened once, in a meeting, months ago," you've just described a checkbox, not a requirement.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0