Describe the incident where an agent took a wrong action, and your remediation.

GUARD starts here. Not what the agent got wrong, but which of these two people could have stopped it.

How to answer this, stage by stage

Nobody is grading whether you can describe a purchasing agent that misfired. They are grading whether you can name the two people a wrong action actually touches, and put the fix where the harm lands, not where the bug lives. Seven moves get you there.

Let's learn

What happens the first time a purchasing agent gets to act on a number nobody taught it to doubt?

Say a distributor builds an agent that watches stock levels across thousands of parts, works out what to reorder, and sends that order straight to the vendor who makes it. No person has to type a purchase order by hand.

Before a tool like this, three buyers split the job across forty thousand parts and nine hundred vendors, about six hours a day each, reading each recommendation before it went anywhere. They caught a bad number the old way: a person who orders the same bolt every week glances at "twelve thousand" next to a part that always says "twenty-four hundred" and just knows something's wrong.

With the agent, most of that six hours disappeared. Orders inside a normal range go out in seconds. A buyer only opens the ones the system flags. Three months before this incident, Cargast raised how big an order could be before a person had to look at it, from five thousand dollars to fifty thousand, specifically so buyers would have even less to check.

Here is the turn. The agent is right almost all the time, comfortably above ninety-nine cases in a hundred. That was never the problem. The problem is what happens to the one case in a hundred once nobody stands between it and the vendor who has to build it. And the case that hurts is never the one Stockhand flags as unsure. It's the one where a stray number sits next to real numbers and nothing in the agent doubts any single input on its own, so a five-month-old test row gets treated exactly like real demand. Checking that number against the vendor's own pattern is the only thing that catches a case like this; the agent's overall accuracy tells you nothing about it.

At its worst, this costs far more than the order itself. A vendor who can't return a custom run, and can't get back overtime pay or spent material, ends up carrying a loss a much bigger company created and never noticed.

The choice I would take back. Raising the auto-approve ceiling to fifty thousand dollars, on its own, without ever checking a new order against that specific vendor's own history first.

What I would leave alone. Some parts really do swing wildly and it's normal, a seasonal item, a promotional part nobody could plan a year ahead for. A hold on every big swing everywhere would just slow down orders that were never wrong. The hold has to be sized to each vendor's own pattern, not applied as one flat rule across all nine hundred of them.

The lesson. An agent that can act, not just suggest, has to earn the right to skip the person standing on the other end of that action. Ninety-nine cases in a hundred doesn't earn it. The hundredth case is the whole job.

Now here is the same thing as a story

The shape above is the whole answer. What follows is what it actually felt like from Threadwell's side of the phone, on the Tuesday steel got cut for a part nobody needed twelve thousand of.

Gideon Baird has run procurement operations at Cargast Industrial Supply for four years. He knows which of his nine hundred vendors will call back if a spec looks off, and which ones will just build exactly what the order says, fast, no questions, because that's the service they're known for. Threadwell is the second kind. Kojo Antwi built that shop's name on doing precisely what a purchase order says, no follow-up call needed.

Stockhand went live for reorders under five thousand dollars that spring. For most of the year that was fine. Buyers stopped watching the small stuff and spent the saved hours somewhere that mattered.

Then, three months before the incident, Cargast raised the ceiling to fifty thousand dollars. Bigger orders started moving with nobody reading them first, and for a while, nothing went wrong. The agent really was good.

Somewhere in Cargast's forecasting tables sat a row nobody remembered: ten thousand units, logged five months earlier by a sales rep building a pitch deck, tagged TEST, never deleted. On a Wednesday morning, Stockhand's forecast for one hex bolt SKU pulled that row in as if it were real demand, stacked it on top of the SKU's normal signal, and recommended a reorder of twelve thousand units against Threadwell, five times the twenty-four hundred that SKU had ordered every three weeks for two years. The order value came out to eighteen thousand four hundred dollars. Comfortably under fifty thousand. Nobody at Cargast saw it before it reached Threadwell's supplier portal at 9:14 that morning.

Kojo saw it an hour later. Twelve thousand looked wrong to him immediately, the same way twenty-four hundred always looked right. He called Cargast's order line before he touched a single bar of steel.

The person who answered pulled up the order. It said confirmed. That was the only thing on their screen. They told him it was fine, go ahead.

So he did. Two weekend shifts, nine thousand four hundred dollars of raw steel rod, twelve thousand units cut, threaded, and shipped by the following Friday.

Kojo's call landed in the middle of this path, not at either end of it. There was nothing built there to catch him.

Eleven days later, someone on Cargast's accounts payable team flagged an $18,400 invoice from a fourteen-person vendor as unusually large and pulled the order history to check it. Twenty-four hundred, every time, for two years, then this. That's when anyone at Cargast looked at the stale test row at all.

The old decision, told as a memory of a meeting. Three months earlier, the team that raised the ceiling talked about it as a trust question: Stockhand had been right often enough that a five-thousand-dollar cap felt like it was there out of habit, not need. Nobody in that meeting was wrong about what the data showed them. They just measured trust in dollars, a number that meant nothing to a shop the size of Threadwell.

The replay, run the same way forward with a hold in place. The same recommendation drafts at 9:14, but the hold checks it against Threadwell's own last eight cycles instead of Cargast's dollar ceiling, sees twelve thousand against a trailing average of twenty-four hundred, and it never reaches the portal. It sits in Gideon's queue instead. He opens it by 9:35, follows the number back to the stale test row, deletes it, and the real reorder, twenty-four hundred units, goes out that afternoon. Threadwell never touches the steel.

The thing I'd tell my past self, from that meeting about the ceiling: the number that mattered was never how much we trusted Stockhand. It was how much of what it gets wrong, someone else has to eat.

The five checks GUARD makes you run, against this exact order

This is a risk question with a real subject in it, someone outside Cargast's own walls who never signed up to be part of Stockhand's decision, so GUARD fits: name both people, find where the harm lands, and ask who never gets a lever.

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

Roundel Home Health runs Aidewell, a scheduling agent that reassigns home visits the moment an aide cancels, and texts the family the new time. No person reviews a reschedule before it sends.

G. Two people again. Priel Nakamura, the scheduling coordinator, can see every aide's calendar and can override any reschedule from her desk. The patient on the other end of the visit never sees Aidewell at all, only a text.
U. When an aide cancelled a 2pm medication visit and Aidewell rebooked it for the next morning instead of finding same-day coverage, the harm landed entirely on the patient: one missed evening dose. Roundel's own schedule looked fine on paper, full, on time, nothing flagged.
A. The patient had no way to push back. The text goes one way, to a phone she barely checks. Her daughter found out the next morning, when she called to ask how the visit went.
R. Any reschedule that touches a medication-linked visit needs a two-way confirm from a family contact before it locks in, not a one-way text. A routine visit with nothing medical tied to it can keep moving the way it does now.
D. Flag every medication-linked visit Aidewell reschedules in the last day and check it against a same-day-coverage rule before the day starts, not after a dose gets missed.

Same missing lever, a different shape. Aidewell's gap isn't a dollar ceiling, it's a text with nowhere to reply.

Swap the trigger and it still runs.
Speed: an interviewer caps the answer at ninety seconds. Skip straight to naming the hold and who it protects.
Cost: engineering says a confirm hold slows down twenty percent of orders. Don't drop the hold to zero. Narrow it to smaller vendors or to actions a vendor can't undo, so the slowdown lands where the risk actually is, not everywhere.
The agent got better, for real: Stockhand's overall accuracy climbs to 99.9 percent. That still doesn't retire the hold, because the one case in a thousand it's wrong about is exactly the case a dollar ceiling was never built to catch.

Where people run it wrong.
They tune the model instead of asking who's standing on the other end of the action it takes.
They measure trust in a number that makes sense to the operator, dollars, overall accuracy, and never in a number that makes sense to the subject, steel, a missed dose.
They treat a person calling in to double check as the safety net, instead of building one that doesn't depend on someone happening to notice.

How to use it live. Say the reframe before any story: "the point of naming who can't push back is that the agent's accuracy was never the problem here. The problem is what happens to the fraction it gets wrong once nobody stands between the action and the person who has to live with it." That buys you room to name the actual hold, instead of promising to "improve monitoring."

Flashcards (click a card to flip it)

Check yourself Score: 0 / 0