I will name five technical terms. Explain each to me as if I were your CEO.

Cravenhurst Systems builds Steward, an AI executive assistant that reads a busy executive's email and calendar and actually moves things: reschedules, replies, holds a slot, without waiting for a human to click anything. Sabeen Doncastle is the seventh AI PM candidate Konstantin Stavrakis, the founder-CEO, has personally interviewed today. He puts his phone face down on the table. "I will name five technical terms. Explain each to me as if I were your CEO."

How to answer this, stage by stage

Nobody is grading whether Sabeen knows five definitions. They're grading whether she can land five of them, cold, one at a time, in front of a person who can end the conversation the moment one goes on too long.

Let's learn

What happens the first time an AI assistant has to explain itself to the one person in the building who can cancel the contract?

Steward is Cravenhurst Systems' AI executive assistant. It reads a busy executive's email and calendar, and it moves meetings, sends replies, and holds a slot, on its own, without waiting for someone to click a button first. Before Steward, Sabeen Doncastle spent three years at Yarrowick Systems, on Warden, an AI scheduling assistant built for law firm partners managing court deadlines and depositions.

Three people, three private answers to the exact same question, none of them written down anywhere.

For Warden's first stretch, four people talked to law firm clients about how it worked, and all four had helped build it. Whatever any of them said off the cuff about its limits was close enough to true, because they weren't really improvising. They were reporting.

Then the team that talked to clients grew to eleven, as Yarrowick hired sales engineers who hadn't built anything underneath the product. A cold audit that spring asked all eleven the same question a partner might ask: does it ever lose track of something. Only three gave an answer that was actually complete and correct.

Four things every one of the five lines does, in every telling, no exceptions.

The reschedule note that broke it sat in message twenty three of a long email thread with Fenrick and Manderley, a law firm client. Warden's window at the time held roughly the last eighteen messages of any thread. By the time the conflict actually mattered, the note that moved a deposition was five messages past the edge of what Warden could still see.

The third box is the one nobody was watching. That's the whole mechanism in one picture.

Nothing about the mismatch looked broken from the outside. Warden kept answering with a confident date every single time. It was just, past message eighteen, quietly answering from memory instead of from the thread.

What it cost at its worst: the firm's insurer flagged three hundred and forty thousand dollars of malpractice exposure over the missed filing deadline. Fifteen weeks after the sales engineer's overconfident answer, the firm did not renew, and Yarrowick lost a contract worth a hundred and eighty thousand dollars a year.

What I would leave alone: the engineering channel where two of Steward's own model engineers argue about window size in raw detail. Let that stay exactly as technical as it wants. Nobody in that channel is deciding whether to sign a contract based on it.

The lesson: "use your own judgment" isn't really an answer. It's a bet that everyone's judgment stays as sharp as the four people already in the room, and that bet pays off right up until the team outgrows the room.

Now here is the same thing as a story

The short version above is what you actually say in the room. Read this one for the fifteen weeks that taught Sabeen why it had to be five written sentences, not five good instincts.

By four in the afternoon, Cravenhurst Systems' one conference room has already run six interviews for the AI PM seat, and Sabeen Doncastle is the seventh. Three years ago, at Yarrowick Systems, she was one of four people who understood Warden's retrieval layer well enough to explain any part of it correctly, on the spot, to a nervous new client.

For eight months, that was enough. Warden was small, its client list was small, and every person who ever answered a hard question about it had watched the thing get built, line by line. When a partner asked something sharp mid-demo, whoever answered wasn't guessing. They were describing something they had actually made.

Then Yarrowick's client list grew, and so did the team who talked to clients about it, from four to eleven. Most of the new hires were sales engineers, sharp people who had never touched the retrieval code. The habit thinned out in three beats. First, their explanations just got longer, more careful, still roughly right. Then a few started rounding the nuance off for speed, because a tidy answer closed a demo faster than a careful one. By the third beat, a sales engineer six weeks into the job, cold-asked by a partner at Fenrick and Manderley whether Warden ever lost track of something, said the fully wrong line: "once it's read your calendar, it doesn't forget."

Fifteen weeks between one overconfident sentence and the day the account left.

It was a small, ordinary sentence, said to sound reassuring. Nobody in the room thought it was a lie. The problem was what the firm did with it.

Every Friday, Fenrick and Manderley's paralegal team had spent about two hours manually checking Warden's suggested schedule against the firm's own master case calendar, the old, slow way, the way they'd always done it before Warden existed. Reassured that Warden "doesn't forget," they quietly stopped. Nobody announced it. It just stopped happening, the way a habit stops when you've been told, by someone who sounded like they knew, that it isn't needed anymore.

Four weeks later, a deposition got rescheduled by email, buried by message twenty three of a long thread. Warden's window held the last eighteen messages. The note was already five past the edge. Nine weeks in, Warden confirmed the original, wrong date to a paralegal who no longer double checked it against anything. The deposition collided with a filing deadline nobody caught until it was almost too late to fix.

Fenrick and Manderley never had a number in their heads for how much to trust Warden. They had a habit: check it every week, or don't. One overconfident sentence, said once, in passing, flipped it to don't.

The decision Sabeen would take back sits in a small meeting eighteen months earlier, back when Yarrowick was still four people. Someone raised the idea of writing one standard answer for the hardest question clients asked. The room talked themselves out of it. A script felt wrong for a team that small, when everyone already knew Warden cold, and writing one down felt like admitting they might someday forget how to explain their own product. Nobody thought to ask what happens on the day someone new has to answer it instead.

Run the same afternoon again, months later, with the five lines written and rehearsed. Sabeen is mid-interview with Konstantin Stavrakis at Cravenhurst Systems, and he asks the exact hard version of the question, cold: "does it ever lose track of something." She doesn't improvise. She gives the context window line, the real number in it, the same words she'd say to any CEO who asked. Ninety seconds later, he's already said the next term.

One design hands every new hire their own judgment and hopes it stays calibrated forever. The other hands them five sentences that stay calibrated no matter who's the one saying them.

What I'd tell myself, sitting in that four-person meeting: "use your own judgment" isn't a policy. It's a bet that the team never grows past the people already in the room, and it's a bet that always eventually loses.

SPARK, run against a live clock

Not a script to sound rehearsed. SPARK is what forces you to design five sentences a whole company can lean on, instead of trusting each new hire to land on something close enough, under pressure, on their own.

Both edges lose. One loses the room today. The other loses a real decision made on a wrong picture, later.

All five terms sit high on both axes, on purpose. That's the whole design in one picture.

The recap, one line per letter: situation is a CEO asking five terms cold, payoff is that he never has to ask what something means twice, anchor is the five rehearsed sentences themselves, each tied to a real number or a real rival, risk is either extreme costing Cravenhurst on a different day, and keep out draws the line at the mechanism, every single time, not at the honesty.

One alternative Sabeen rejected on the way to this design: keep every answer deliberately vague, never name what actually breaks, to sound safe no matter what. She turned it down because a vague answer doesn't protect a CEO from a bad decision, it just delays the moment he finds out the hard way, and that's exactly the mistake that cost Yarrowick a hundred and eighty thousand dollars a year. She also rejected memorizing one long, fully technical glossary and reading from it: a rep under real pressure compresses six sentences into whichever extreme feels safest that day, which recreates the exact problem the five short lines exist to fix, just with more words in it.

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

Same five letters, a city's waste trucks instead of a law firm's docket, and this time the wrong guess isn't a filing deadline, it's a street that never gets collected.

Winterset Civic Systems builds RouteMind, an AI tool that plans a city's daily garbage truck routes. Trescott Osiel is interviewing for AI PM there. Merribell Beckwaite, the city's Public Works Director, hits the same wall eight months into the pilot: mid-demo, she names two terms cold and wants a straight answer to each before she signs off on a full rollout.

Same shape of decision, a different room, a genuinely different pair of terms this time.

Mapped onto SPARK: situation is Merribell asking cold, mid-demo, with a rollout decision riding on the answer. Payoff is that she never has to ask what something means twice. Anchor is two example lines, tuned to this domain: "Confidence threshold is how sure RouteMind has to be before it changes a route without asking a dispatcher first. We set that bar high on residential streets, low on open highway, because a wrong guess on a school street costs us more than a slightly longer route." And: "Model drift is when RouteMind's routes slowly stop matching how the city actually grows, new streets, closed lots, a shifted school schedule. We check it every month against real pickup logs, not just when a complaint comes in." Risk is the same shape as Cravenhurst's: too technical, and Merribell stops trusting the pitch; too soft, and a city council member gets blindsided by a missed street nobody flagged as a real risk. Keep out draws the same line: the routing model's actual optimization math stays in a written doc for the city's own traffic engineers, never in the live demo.

Swap the trigger and it still runs.
Speed: an interviewer caps each answer at one sentence. Drop the second sentence's "expect" clause and keep only the noun plus the one real catch.
Cost: no time to write and rehearse all five in advance, mid-interview, cold. Lead with the one term you know cold from a real incident, and say plainly the rest need a beat to compress right, rather than fake confidence you don't have yet.
The model got better, for real: say Steward's underlying model doubles its accuracy overnight. All five lines still hold, because "better" doesn't erase the catch. Hallucination and context window limits still exist in a stronger model, just at a lower rate, so the sentence stays true, only the number attached to it would eventually change.

Where people run it wrong.
They memorize a written glossary and read it flat, which lands like a legal disclaimer, not an answer from a person who's actually used the product.
They let each explanation's length grow with how nervous they are, not with what the CEO actually needs to hear.
They never say what's deliberately left out, so a sharp CEO assumes something's being hidden rather than sees a real choice about what he needed to decide today.

How to use it live. Before naming a single term, ask yourself one question: what's the one sentence this CEO will use to describe it back to his own board. Build your two sentences to survive being repeated exactly that way, word for word, by someone who wasn't in the room.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0