Describe how you would respond to a regulator's inquiry about your AI feature.

Birchwell Invest's model recommends how to rebalance a client's portfolio and can execute the trade automatically. Renata Duarte is the product manager for that rebalancing model. She got the call on a Tuesday: a state regulator wanted to know why a retired client's "conservative" account held forty percent in a single stock.

How to answer this, stage by stage

Seven stages. A regulator's inquiry rewards exactly the opposite instinct from a normal incident review: slower, more documented, less defensive.

Let's learn

Say a robo-advisor manages tens of thousands of client accounts, rebalancing each one automatically to match the client's stated risk tolerance.

Eighteen months ago, Birchwell's compliance team reviewed a full daily sample of fifty accounts, pulled at random, against every account then under management, about forty thousand of them. That sample felt thorough at the time, since fifty was a real, meaningful slice of forty thousand.

Account volume more than tripled over those eighteen months, but the daily review sample stayed fixed at fifty accounts, a number nobody ever revisited. As total volume grew, that same fifty represented a shrinking slice of the whole, and conservative-tier accounts, always a minority of the total, got squeezed out of the sample almost entirely.

At its worst: a retired client with a conservative risk profile ends up holding forty percent of her account in a single technology stock, discovers it herself while checking her statement, and files a complaint that reaches a state regulator before it ever reaches Birchwell's own compliance desk.

What I would leave alone: the aggressive-tier review process didn't need to change. That segment is large enough that even a shrinking fixed sample still caught real problems there. The gap was specific to a smaller segment, not universal.

The lesson: a review sample that felt thorough on day one can quietly stop being thorough as volume grows, with no single moment marking when it happened.

Now here is the same thing as a story

The short version above is what you'd say drafting the regulator response. Read this one for how Renata actually traced it.

Renata Duarte has managed Birchwell's rebalancing model for three years, long enough to remember when the daily fifty-account review genuinely felt like a real check on the whole system.

There was no single bad day that broke it. Account volume climbed steadily, quarter after quarter, while the fifty-account sample stayed exactly the same size, a number set once and never revisited as the business grew around it.

Five steps, and the third one is where a fixed number quietly stopped meaning what it used to.

Then the regulator's letter arrived: a formal inquiry, prompted by a retired client's complaint, asking Birchwell to explain how a conservative account ended up forty percent concentrated in a single stock, and to produce records for a sample of similar accounts.

The real start was eight months before anyone outside the company noticed anything was wrong.

Renata's first move wasn't to explain. It was to rebuild the timeline properly, and then slice the data by segment instead of trusting the aggregate number the team had been watching all along.

The number on the review sample never changed. What it actually represented did, completely.

Before pointing at the model, she ruled out the easier explanation: maybe the dashboard itself was miscounting drift, not the recommendations. A direct check of the actual recommended trades, bypassing the dashboard entirely, confirmed the drift was real, not a reporting artifact.

Three named suspects, and only one of them survived a direct look at the real recommendations.

She sorted every client segment by two things: how big a share of the total book it was, and how much of the daily review sample it actually got.

The smallest segment got the smallest share of review, exactly the opposite of where the real risk was building.

With the root cause confirmed, direct check, not a guess, Renata assembled the response the regulator actually needed.

Four parts, and a response missing any one of them reads as either incomplete or evasive.

The old review asked whether fifty accounts, sampled the same way every day, still looked fine. The new one scales the sample with total volume and stratifies it by segment, so a minority segment can't quietly fall out of view again.

I set the review sample at fifty because it was a real, defensible number the day I set it. It took a regulator's letter, not our own dashboard, to see that a fixed number stops representing the same thing once everything around it triples.

TRACE, for a regulator's letterNot a defense. TRACE is what forces you to find the real cause before you ever put anything in writing to someone who can hold you to it.

The recap, one line per letter: timeline is the eight-month gap between the real start and the inquiry, recut is conservative-tier drift far outpacing the aggregate, assume nothing is ruling out a dashboard bug first, cause candidates is naming three real suspects, and evidence test is checking the actual trades directly to confirm the true cause.

And if you want to be sure it really works, try it somewhere elseSame five letters, a food-delivery marketplace instead of a robo-advisor. This time the regulator asking questions is a labor regulator, not a financial one.

Cartwheel Eats runs an AI system that scores delivery drivers and adjusts how often they're offered high-paying orders. Merritt Sloane manages that scoring model.

Mapped onto TRACE: timeline is when a routing-efficiency update quietly began weighting delivery speed more heavily, three months before a labor regulator's inquiry into pay disparities. Recut is finding that drivers in one specific city, with older phone hardware that reported location less precisely, scored lower on the speed metric and got fewer high-paying offers, while the citywide average barely moved. Assume nothing is checking whether the location data itself was the problem, a hardware and reporting issue, rather than assuming the scoring model was unfairly targeting anyone. Cause candidates are the routing update, a hardware-driven data quality gap, or a genuine behavioral difference in that city's drivers. Evidence test is comparing scores for drivers with newer phones in that same city against drivers with older phones, which isolated the hardware-driven data gap as the real, confirmed cause.

A different platform, a different regulator, and the same shape of blind spot in a different corner.

Swap the trigger and it still runs.
Speed: an interviewer caps you at a minute. Say "timeline first, then segment, rule out instrumentation, then one test to confirm the real cause," and stop.
Cost: legal wants a response sent within twenty-four hours. Send an honest interim letter naming what's confirmed so far and when the full findings will follow, rather than guessing at a cause to look responsive.
The model gets better, for real: if the rebalancing model's overall accuracy improves, that doesn't mean a review process built for a much smaller book is still adequate. A better model with a stale review process can still miss the same segment.

Where people run it wrong.
They respond with the aggregate number, which is exactly the number that made the problem invisible in the first place.
They assume the model must be the cause without checking whether a dashboard or reporting bug could look identical.
They treat the regulator's specific question as the whole scope, instead of checking whether the same gap exists elsewhere too.

How to use it live. When this question comes up, say the word "segment" out loud before you say anything about the model. That's the habit that actually catches what an average hides.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0