List four responsibilities an AI PM holds that a traditional PM does not.

Vouchwell is Quillstack's tool for reading invoices and receipts and posting them straight into a finance team's books. Ingerlise Doone owns what Vouchwell gets right and wrong. Her head of product, Bryndis Quennec, is deciding whether Vouchwell even needs a dedicated AI PM, or whether a regular PM could run it just as well, and she wants four real answers, not a job title.

How to answer this, stage by stage

Nobody is grading whether you can name four things fast. They are grading whether you can rank them, and defend the top one, when someone who runs a whole product org is deciding if your job is even real.

Let's learn

Vouchwell reads a vendor's invoice or receipt, pulls out the vendor, the amount, the line items, and the due date, and posts it straight into a finance team's books.

Four jobs, in the order this whole answer defends. Miss the first one and the other three have nothing real to check themselves against.

Before Vouchwell, five clerks on Vesterholt Foods' accounts payable team keyed about 2,100 invoices a month by hand, six minutes each, and an invoice sat nine days on average before anyone posted it. Vouchwell reads most invoices in about 12 seconds now. Anything under $5,000 that scores 96 percent or higher on every field, vendor name, invoice number, line items, total, auto-posts straight to the ledger with nobody looking at it. About seventy four percent of invoices clear that gate on their own.

Here is the turn. The extra mistakes an AI PM watches for were never really about the model reading a number wrong now and then. The real mistake was that nobody had ever checked whether Vouchwell had seen a European style invoice before it was allowed to post one on its own.

Three small gaps, none of them a bug in the usual sense. Stacked together, they let a $30,450 invoice post for thirty dollars and forty five cents.

What I would leave alone: a forty dollar office supply invoice from a vendor Vesterholt has used for six years does not need this level of scrutiny. The format has been the same forty times running, and being wrong there costs an afternoon, not a vendor relationship. Auto-post it and move on. Widening the plausibility check does mean a slice of invoices take ninety seconds of a person's time instead of twelve seconds of the model's. That is a trade Vesterholt accepts on purpose, so a number off by a thousand times never gets the chance to post itself again.

The lesson: a responsibility that only exists at launch is not a responsibility, it is a checklist item that gets crossed off once. The eval set, the threshold, the drift watch, and the fairness check are jobs that keep running long after the ship date. The moment an AI PM stops treating them that way is the moment a wrong number gets through.

Now here is the same thing as a story

The short version is above, the part you'd actually say out loud. Read on if you want to feel why a decimal point mattered more than the whole rest of a clean invoice.

Ingerlise Doone has run Vouchwell for two years, and she built one habit for herself early on: any time a new vendor got onboarded, she pulled that vendor's first three invoices herself, by hand, before letting Vouchwell auto-post anything from them. It caught small things. A vendor who wrote dates day first instead of month first. A vendor whose invoice number sat where the total usually did.

For most of that first year, this habit cost her almost nothing. Vesterholt Foods bought from domestic suppliers, two or three new vendors a month, easy to check by Friday. Vouchwell's auto-post rate climbed steadily, and every quarterly review looked cleaner than the one before it.

Then Vesterholt started buying internationally. Lumber and packaging from Germany. Specialty ingredients from three countries in South America. New vendors went from three a month to fifteen. Ingerlise's personal check thinned the way habits do when nothing punishes them for thinning. First she checked every new vendor. Then just the ones billing over some rough size in her head. Then a random handful, on the weeks she remembered. During a quarter with two other launches stacked on top of Vouchwell, she let the onboarding team add vendors on their own, trusting the 96 percent confidence gate to catch whatever mattered.

Nothing about that felt reckless at the time. The gate had never let anything obviously wrong through. It just had never been asked to read a European style invoice yet.

Thrune Millwork, a lumber and packaging supplier outside Stuttgart, sent its usual shipment invoice: thirty thousand four hundred fifty dollars, written the way German invoices write it, a period marking the thousands and a comma marking the cents. Vouchwell's character reading was flawless. It saw every digit correctly, 98 percent sure of the whole read. Its number parsing step, built and tested almost entirely on US style invoices, took the period as a decimal point instead. Thirty thousand four hundred fifty dollars became thirty dollars and forty five cents. Under five thousand dollars. Every field scoring comfortably above the line. It auto-posted in the same overnight batch as everything else, and nobody at Vesterholt ever saw it.

The whole story in one picture. A model can be genuinely, honestly sure of itself and still be wrong about the one thing nobody thought to test.

Three weeks later, a woman from Thrune's accounts receivable team called Vesterholt's AP line. Polite, a little tense. Their next shipment was scheduled to leave that week, and their books showed a thirty thousand four hundred fifty dollar invoice sitting open, marked paid for thirty dollars and forty five cents. Was there a problem with the account.

Ingerlise pulled ninety days of Thrune's invoices that afternoon and found it inside twenty minutes. Every digit correct. Every field above the line. The decimal point simply pointed the wrong way, and nothing downstream of the confidence score had ever asked whether thirty dollars for a truck of lumber made any sense.

The decision she would take back traces to a planning meeting a year earlier, when the 96 percent threshold first got built. Someone on the call had asked whether the gate should also check the total against what a vendor normally bills. With a launch date six weeks out and almost every vendor in the pilot billing in plain US dollars, the answer, reasonable at the time, was to start with digit confidence alone and add more checks later if international vendors ever became a real share of the volume. Nobody put a date on the calendar to go back and check.

Run it again with the plausibility check in place. Thrune's invoice reads $30.45 against a vendor whose last six invoices averaged twenty eight thousand dollars, a mismatch of roughly a thousand times. It never clears the gate. It routes to a person in the same overnight batch, gets confirmed in under two minutes the next morning, and posts correctly before Thrune's AP team has any reason to pick up a phone.

What I would tell myself, back in that planning meeting: I mistook the model being confident for the model being right. It was completely sure of every digit it read. It was never once asked whether the number those digits made was a number that made sense.

ORDER: ranking the four jobs by what breaks first

FLIPS would fit if this were about a person's trust wearing thin. Nothing here turns on a habit fading, it is a fixed set of four things competing for rank, which is ORDER's job.

One of these you can fix whenever you get to it. The other one already cost Vesterholt a vendor's trust.

The eval set is the only box on the left with nothing feeding into it. Everything else on this chain waits on it first.

Three things worth stating directly, since this is where the real judgment sits. The alternative Ingerlise's team considered, and rejected, was simply raising the auto-post threshold from 96 to 99.5 percent across the board, instead of building a separate plausibility check. It lost, because Thrune's invoice already scored above 98 percent on every field. A stricter blended bar would still have let it through, since the model was never unsure about anything, it was only wrong about which country's punctuation it was reading. The AI specific failure worth naming by name is exactly that: a confidence score measures how sure a model is about what it read, not whether what it read makes sense in context, and a distribution shift toward a new invoice format the eval set never covered can hide completely behind a high confidence number. The guardrail is the plausibility check itself, comparing the extracted total against that specific vendor's own recent invoice range, not a flat dollar cutoff for every vendor alike. And the trade-off is real and accepted on purpose: a wider plausibility check means more invoices, a few percent more, take ninety seconds of a person's time instead of twelve seconds of the model's. Vesterholt accepts that cost on purpose, because the alternative is a number that is a thousand times wrong sitting in the books for three weeks before anyone notices.

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

Same five letters, a veterinary emergency room instead of an accounts payable desk, and the fragile thing this time is not a decimal point, it is the wording a front desk tech happens to type.

Intakewell, built by Fenlight, reads a veterinary ER's intake notes and vitals and drafts a triage summary plus an urgency flag for the on-duty vet: see this one now, or it can wait. Renske Corbary owns Intakewell's model behavior at Fenlight. Oxlip Creek Emergency Vet has run it for eight months and is deciding whether to keep it past the trial.

Different desk, same shape of danger. The cases that matter most are exactly the ones the eval set is least likely to have seen worded that way.

Intakewell's eval set held 620 real intake cases at launch, built mostly from how a trained vet tech types a presenting complaint: "distended abdomen, non-productive retching" for a dog showing early signs of bloat, a condition that can turn fatal within hours. Six months in, a newly hired tech typed the same case in her own words: "dog restless, hard belly, drooling." Intakewell's confidence on the urgency call landed just under the cutoff, not because the case was mild, but because the wording matched nothing closely in the eval set, so it defaulted to the lower urgency bucket and filed the dog as routine, can wait.

Dr. Marence Colquhoun caught it walking the waiting room about twelve minutes later, well inside the window that matters for bloat, the same way Ingerlise's old habit once caught Vesterholt's format gaps: not from the flag, from actually looking.

Same rank, different lever, mapped straight onto ORDER: the outcome is the same shape, that the urgency flag Intakewell gives a vet matches what a trained vet would actually say looking at the same dog. The reversibility argument holds just as hard, retuning a confidence cutoff is an afternoon, rebuilding a vet's trust in the flag after one missed bloat case is not, especially once a clinic starts double-checking everything the tool says. The dependency is identical: no plausibility override for known critical symptom combinations means anything without a real eval set behind it. And the evidence check transfers directly, pull the last ten new hires' intake notes and see how many phrasings never appeared in the eval set at all, Oxlip Creek's answer was seven of ten.

Swap the trigger and it still runs.
Speed: an interviewer caps you at ninety seconds. Skip straight to it: eval set first, then the threshold, then drift, then fairness, and here's the one failure that proves the order.
Cost: no budget this quarter to widen the eval set. Ship the plausibility check as a flat rule for known critical symptom combinations first, and grow the eval set's wording coverage with every new phrasing a real tech types.
The model got better, for real: say Intakewell's confidence calibration improves across the board. The critical-combination override does not loosen. A better model still needs the same eval set coverage to know what "better" even means for a case it has never seen worded that way.

Where people run it wrong.
They treat the eval set as a one-time launch task instead of something that has to grow every time a new person types a case in their own words.
They raise the confidence threshold higher instead of asking whether the threshold was ever checking the right thing.
They add a plausibility check everywhere at once instead of starting with the small number of cases where being wrong actually costs hours, not an afternoon.

How to use it live. Before ranking anything, ask yourself out loud what breaks first if you only had time to build one of the four. That question alone buys real thinking time, and it is usually the exact distinction an interviewer is listening for.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0