Describe the compliance difference between an AI feature in hiring and one in marketing.

Larkspur Retail Group runs a regional grocery chain with two AI tools built by two different teams: a resume screener that shortlists job applicants, and an ad-targeting engine that decides which promotions each shopper sees. Zanele Okoye leads responsible AI across both, and used to review them against the exact same checklist.

How to answer this, stage by stage

Nobody is grading whether you can name every applicable law. They're grading whether you know why the same company treats two AI tools so differently.

Let's learn

Larkspur's resume screener reads applications for store positions and ranks candidates before a recruiter looks at a single resume. Its ad-targeting engine decides which weekly promotions land in each shopper's app.

Before anyone compared them side by side, both tools went through the same lightweight AI governance checklist: a short form, a privacy review, and a launch sign-off. Nobody had asked whether one checklist could really fit both.

Now, the turn: the checklist wasn't wrong for marketing. It was wrong for hiring, because it had never actually been built with hiring's stakes in mind at all.

What I would leave alone: Larkspur's inventory-forecasting AI, which predicts how much produce to order per store, needs no personal-fairness review at all. It never makes a decision about a person, so the entire GUARD framework doesn't apply to it in the first place.

The lesson: "it's an AI feature" isn't a stakes level. A job applicant and a shopper are not carrying the same risk into the same interaction, and the governance has to follow the person, not the presence of a model.

Now here is the same thing as a story

The short version above is what you'd say defending Larkspur's two-tier governance to its board. Read this one for how the gap actually got noticed.

Zanele Okoye has led responsible AI at Larkspur for two years, building the company's very first shared governance checklist herself, back when there were only two AI features to review at all.

A new hire on the marketing team, previously at a hiring-technology company, sat in her first cross-team review and asked an ordinary question: don't we need a bias audit for the ad engine too, the way we do for the resume screener?

Zanele had never actually drawn this picture before. Once she did, the two tools clearly did not belong on the same checklist.

Zanele realized, answering out loud for the first time, that Larkspur didn't actually have a bias audit for either tool yet, only the same shared checklist applied uniformly to both.

The recruiter always had a lever. The applicant never did, and the shared checklist had never asked about that gap at all.

She commissioned a real disparate-impact audit of the resume screener that quarter, the first one Larkspur had ever run.

Three steps ran the way they were built to. The missing third one, an appeal, was the one an applicant actually needed.

The audit found the 0.63 selection ratio for applicants 50 and over, a real gap that had been running quietly since the screener launched, invisible to a checklist built for a different kind of risk entirely.

The resume screener and the ad engine took different branches from the very first question. The old shared checklist had never asked it.

Zanele split the governance in two. The resume screener got a full audit process, a human override step, and a written appeal path for rejected applicants.

Four real obligations, sized for a decision that can end a stranger's shot at a job.

The ad engine kept a much lighter version: honoring opt-outs, screening for protected-trait proxies in its targeting logic, and a spot-check twice a year rather than a full audit.

Real guardrails, just sized to a decision a shopper can walk away from with one tap.

A full year, moving the ratio from a real problem to safely past the fairness threshold.

The old approach treated every algorithm as equally deserving of the same light review. The new one asks what happens to the person on the other end first, and lets that answer decide how much governance follows.

I ran one shared checklist because it kept a small team's review process simple, and for a while both tools genuinely were low-stakes and low-volume. It took a new hire's plain, undramatic question to see that the checklist had quietly become the ceiling for hiring's real risk, not just a floor.

GUARD, sorted by who's at riskNot a policy comparison. GUARD is what forces you to name who has the lever in each tool, and who doesn't.

The recap, one line per letter: groups is the recruiter and applicant versus the marketer and shopper, unequal is a 0.63 selection ratio versus a minor mistargeted ad, ability to contest is no appeal path versus a simple opt-out, reduce is a full audit process versus a lighter policy, and detect is the four-fifths rule versus a complaint-rate check.

And if you want to be sure it really works, try it somewhere elseSame five letters, a university instead of a retail chain. This time the two tools being compared are scholarship screening and course recommendations.

Birchwood University uses an AI tool to pre-screen scholarship applications, and a separate AI tool to recommend elective courses to enrolled students. Tomas Adebayo-Lindqvist runs both product lines and faced the exact same comparison question from his provost.

Groups: for scholarships, the review committee, who can override a recommendation, and the applicant, who often cannot see why they weren't shortlisted. For course recommendations, the academic advisor, and the student, who can simply ignore a suggestion and pick a different elective. Unequal: applicants from under-resourced high schools, whose transcripts look unfamiliar to a model trained mostly on well-resourced school formats, get screened out at a noticeably higher rate. A student who ignores a bad course recommendation loses nothing but a slightly less optimal schedule. Ability to contest: a scholarship applicant has no visibility into why they were screened out, while a student can simply choose a different course with zero friction. Reduce: the scholarship tool gets a full review process and an appeal path; the course tool gets a simple "why am I seeing this" explanation and an easy override button, nothing heavier. Detect: scholarship screening rate by school-resource level, tracked against a fairness threshold; course-recommendation override rate, tracked as a simple usage metric.

Swap the trigger and it still runs.
Speed: an interviewer caps you at sixty seconds. Say "match the governance to whether the person has another way in, hiring usually doesn't, marketing usually does," and stop.
Cost: there's no budget this quarter for a full disparate-impact audit. Say so honestly, and start with a simple selection-rate breakdown by group, since even a rough number beats no visibility at all into hiring's real risk.
The model gets better, for real: if the resume screener's overall accuracy improves, that alone says nothing about whether its selection rate across groups has also improved. Track both, because a better average can still hide a widening gap.

Where people run it wrong.
They apply one governance checklist to every AI feature, calibrated for whichever tool got reviewed first.
They assume "it's regulated" and "it isn't" is the whole story, missing that the real driver is whether the affected person has another path or a real way to push back.
They build a heavy audit process for hiring but never actually check whether it produces a working appeal path, mistaking paperwork for protection.

How to use it live. When asked to compare compliance across two AI features, don't start with which laws apply to which. Start with what happens to the person on the receiving end if each one gets it wrong, and let that answer decide the weight.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0