How do you present a shipped AI project when you cannot share the internal metrics?

Marisol Quiroga spent two years as the AI PM for fraud detection at Thriftbridge, a secondhand goods marketplace. Her signed offer letter bars her from sharing exact internal metrics after she leaves. Dwight Colson is interviewing her for a similar role at a different marketplace, and the first thing he asks is "so what did it actually do."

How to answer this, stage by stage

This isn't about remembering a rule. It's about staying calm the moment someone asks for a number you're not allowed to give, and having a real answer ready instead of freezing.

Let's learn

Say you built something that genuinely worked, at a company that will never let you say exactly how well.

Before Marisol settled on a way to talk about it, an early answer in a practice interview went: "we reduced fraud significantly, it was a big win for the team." No number, no method, nothing anyone could hold onto.

After she reworked it, the answer became specific about everything except the one number she legally couldn't share: the category, the six-month window, the before/after comparison, and a rough direction, all defensible, none of it a violation.

At its worst: a candidate quotes a specific figure from memory, gets a sharp follow-up about methodology in a technical screen, can't answer it cleanly, and every other claim on their resume gets read with new suspicion from that point on.

What I would leave alone: when a project genuinely has no NDA at all, exact numbers are still the strongest thing to share. This isn't an argument against real precision, only against faking it when you don't have the right to give it.

The lesson: an NDA doesn't erase your evidence. It just moves the evidence from the number itself to the method behind it, and the method was always yours to keep.

Now here is the same thing as a story

The short version above is what you'd say in the room. Read this one for how Marisol actually built the answer.

Marisol could tell a genuine return request from a staged one just by reading the photos and the timing, two years before any model helped her do it faster.

The fraud detector she built at Thriftbridge worked well enough that leadership quoted its impact in an all-hands twice, and she watched the real percentage drop appear on a dashboard she wasn't allowed to screenshot for anything outside the company.

One box looks weak and stays weak. The other looks strong right up until it doesn't.

Her first practice run leaned toward saying nothing numeric at all, since the alternative felt like a legal risk she didn't want to take.

A friend doing a mock interview with her pushed back with one plain question: "so it just... worked? That's it?" Nothing dramatic, just a flat, unconvinced pause.

The exact internal number and the vague adjective both fail, for opposite reasons. The middle ground is where the real answer lives.

She spent a weekend rebuilding a small, honest version: the same detection logic, run against a public dataset of marketplace scam listings she found through a research paper's release, with real numbers she could show anyone, no NDA attached.

Four different answers for four different situations. Most of Marisol's real story sat in the second branch.

With both pieces ready, the redesigned answer names the caveat first, gives the rounded direction next, describes the exact measurement method after that, and then offers the public rebuild as something Dwight can actually run himself, live, in the interview.

The order matters as much as the words. The caveat comes second, right after naming the project, never as an afterthought.

Three things are always safe to share. One thing never is, and Marisol learned to know the line by heart.

The old answer asked Dwight to either trust a vague claim or catch her in a risky exact one. The new one hands him a real method, a defensible direction, and a public demo he can check himself, all without touching the locked number at all.

I almost went with saying nothing, believing that silence was the safest, most careful choice. It took one friend's flat, unconvinced pause to show me that silence reads as having nothing, when the truth was I had a great deal, just not the one number I'd been trained to think was the whole story.

PICK, on the one metric she can't sayNot silence. Not a fake number. The one honest position between them.

The recap, one line per letter: position is rounded and directional, always; impact is a cost on both silence and fabrication; cost asymmetry is the hidden, expensive error of fake precision beating the visible, cheap error of no number; and kill criteria is dropping the project entirely if even the direction can't be shared.

And if you want to be sure it really works, try it somewhere elseSame four letters, a parcel-routing project instead of a marketplace. A different candidate, and this time the locked number is a cost figure, not a fraud rate.

Solveig Haugen built a route-scoring model at a regional logistics company, flagging delivery routes likely to miss a promised window before the truck even leaves the depot. The exact cost savings are locked under a client contract, not just an internal NDA. Applied to PICK: position is still rounded and directional, never exact, never silent. Impact is that a hiring panel of operations people will specifically want a magnitude, in hours or dollars, not just "it helped." Cost asymmetry is the same shape: a fabricated dollar figure collapses under one clarifying question from anyone who's actually run a fleet before. Kill criteria: since even a rounded percentage was barred by the client contract, Solveig switched to describing the trend shape only, "the gap between promised and actual delivery time narrowed steadily over the pilot's ten weeks," with no number at all, which held up better than guessing at one would have.

The same four parts hold up even when Solveig couldn't use a number at all, just the shape of the trend.

Swap the trigger and it still runs.
Speed: an interviewer wants an answer in ten seconds. Say "meaningful drop, exact number's under NDA, happy to walk through the method," and stop there.
Cost: you genuinely have zero usable public data to rebuild a demo with. Describe the method in enough detail that a technical panel could estimate the plausible range themselves.
The project gets better, for real: if the model's real impact grows after you leave, the same rounded, directional approach still holds, a bigger number under the same NDA is still a number you can't state exactly.

Where people run it wrong.
They either overshare a number they were never allowed to give, or undershare and leave a hiring manager with nothing at all.
They let the number "sharpen" under a follow-up question, which is the clearest tell that it wasn't real to begin with.
They forget that the method, not the metric, is the part an NDA can never actually lock away.

How to use it live. The moment someone asks for a number you can't give, say the caveat out loud first, calmly, before you say anything numeric at all. That one sentence buys you the room to give a rounded, honest answer instead of freezing or overcorrecting.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0