Argue the position that AI PM is not a distinct discipline, then rebut it.

Castline is Tenorbridge's tool for cloning a voice actor's own voice, so a script they never read out loud still comes out sounding like them. Solmaz Karsten runs AI product there. Six weeks after Castline quietly saved a hardcover release date by cloning three missing chapters in narrator Ambrosine Thale's voice, Tenorbridge's VP of product stands up in a hiring review and argues that a separate "AI PM" title is a fad, no different from "mobile PM" or "API PM" before it, and should be cut before it becomes a permanent line on the org chart.

How to answer this, stage by stage

Nobody is grading whether you can sound confident about AI being important. They're grading whether you can build the other side's case honestly, then say the one specific reason it breaks, in a room where your VP of product is watching.

Let's learn

Castline reads a stack of a voice actor's old recordings and writes brand new lines in their voice, for a script they never read out loud.

Five steps. The fourth one, one score deciding whether a line ships, is the step the whole argument turns on.

Before Castline, when a narrator couldn't come back to record a late chapter, a publisher hired a sound-alike actor and re-recorded the missing pages from scratch. That took about three and a half weeks and cost around $21,000 a title, once you count studio time, the new actor's fee, and mixing. With Castline, the same missing pages come back in about two days, for about $1,200 a title.

For most of a year, the only gate on a cloned line was that one similarity score. It ran high, 96 percent on average, and nobody had a reason to look past it, because nothing had gone wrong yet.

Here's the turn. A stiff pause or a slightly off word in an ordinary line was never the real problem, because a listener's ear slides right past it without landing anywhere. The real problem showed up the one time the missing pages were not ordinary lines at all, they were the emotional center of the whole book.

This is the whole reason the steelman fails. One side of this picture has a ticket queue. The other one only has a listener's ear.

The whole answer, in one picture. A confident average is not the same thing as a confident answer about the one scene that actually mattered.

What I would leave alone: ordinary, low-stakes narration, scene-setting, a line of dialogue with no weight on it, never needed a second check. The similarity score alone is still fine there, and adding a listening panel to every line would just slow the whole product down for no reason.

The lesson: matching a voice and matching a performance are two different jobs. A great score on the first one tells you nothing about the second, and a metric that's confident about the wrong thing is more dangerous than one that admits it's unsure.

Now here is the same thing as a story

The short version above is what you actually say in the room. Read this one when you want to feel exactly what a 96 percent score was hiding, and how a habit thinned out until nobody was listening anymore.

Quentina Sarrow has reviewed narration for Tenorbridge for four years, and she has one habit nobody had to teach her: before any cloned scene ships, she listens to it, not just checks the score. She can hear the difference between a voice that sounds like someone and a voice that sounds like someone reading their own eulogy out loud by mistake, and for most titles, that difference never showed up.

Castline arrived properly two years ago, and for a long stretch it was the easiest part of her week. A cloned batch would land Monday morning, mostly short filler lines, "he nodded," "the door creaked shut," and she'd listen to a handful, hear nothing wrong, and sign off by ten.

So the full listen thinned. First she stopped listening to every line in a batch and started sampling a third. Then, as the similarity scores kept coming back in the mid-90s, she started trusting the number on anything that scored above 94 and only listening to the ones below it. Nobody decided this on purpose. It just kept being fine, the way habits do when nothing punishes them.

Ninety six percent never dropped. It was never the number that was going to warn anyone.

The Thornwake epilogue landed on a Thursday. Ambrosine Thale, who'd voiced all four books, had a family emergency the same week the publisher needed three new chapters recorded for the special edition. Castline cloned her voice from the earlier books. The whole batch scored 96 percent. Quentina sampled two lines from the funeral scene, both scored 97, and signed off.

Two weeks later, a colleague forwarded her a review screenshot with one line: "isn't this the flat-sounding bit people keep quoting." Quentina hadn't seen it. She pulled up the actual audio for the funeral scene and listened, properly, for the first time since it shipped.

It wasn't wrong, exactly. Every word was there, the timbre was unmistakably Ambrosine. But the pacing sat flat where the text called for a held breath, and the pitch never dropped the way it does when someone is actually grieving on the page. It sounded like her, reading, not like her, feeling it.

She didn't stop there. She pulled the last quarter's full batch, forty titles, and listened to every high-drama scene in it, not just the flagged ones. Two full days. Twelve titles had the same problem, all clustered in scenes tagged grief, revelation, or loss, all scoring in the mid-90s on similarity, none of them caught by anything anyone had built.

The decision Solmaz would take back traces to a scoping call fourteen months earlier, when Castline was mostly filling in short connective lines. Someone had asked whether emotionally loaded scenes needed their own check. The answer, reasonable at the time, was that there wasn't enough of that kind of work yet to justify slowing everything down, one similarity score was simpler, and they'd revisit it if it ever became a real share of the batch. Nobody circled back, because nothing forced the question until a third of one quarter's emotional scenes were quietly wrong and a publisher was on the phone.

Run the same batch again with the gate rebuilt the way it works now: any line inside a scene tagged high-emotional-stakes routes to a three-person human listen, regardless of the similarity score. The funeral scene scores 71 on that check, well under the pass bar, and routes to a human pass automatically, in the same overnight run that produces the draft. Eleven of the twelve problem titles get caught before a reader ever hears them. The twelfth needs a second pass, adding two extra days, not two extra weeks.

What Quentina would tell herself, back when the sampling habit first thinned: skipping the full listen wasn't careless. It was the sensible call when nearly every batch was low-stakes filler. Nobody ever agreed to revisit it once the cloned lines started landing on the pages readers actually cried on.

That's the story sitting behind the hiring review. Solmaz walks in already knowing the Thornwake numbers cold, and when Rowlandson Vail stands up and argues that "AI PM" is a fad no different from "mobile PM," Solmaz doesn't open by disagreeing. Solmaz opens by building Rowlandson's case first, out loud, better than Rowlandson built it, then says the one sentence that ends the meeting.

PICK: the four moves that ended the hiring-review argument

Not a way to win a debate on confidence. PICK is what forces a real commitment about which kind of wrong actually breaks trust, and what would have to change for the other side to be right.

The impact step isn't a rate. It's these two people, carrying two different weights.

One mistake announces itself on a spreadsheet. The other one hides inside a number that looks fine.

Three things worth stating directly, since this is where the real judgment sits. The alternative Solmaz's team considered, and rejected, was simply raising the similarity threshold higher, from 96 to 99 percent, instead of adding a separate human-graded check. It lost, because the funeral scene's problem was never about how closely it matched Ambrosine's voice, a tighter acoustic threshold measures the same wrong axis more strictly. The AI-specific failure worth naming by name is a confident proxy: a metric that measures the wrong thing while looking completely sure of itself, acoustic identity standing in for performance quality when the two only sometimes agree. The guardrail is a scene-type gate, not a smarter number: any line inside a scene tagged high-emotional-stakes requires a human-graded pass before it ships, no matter what the similarity score says. And the trade-off is real and accepted on purpose: routing those scenes to a listening panel adds two to three extra days and about $300 a flagged scene, and Tenorbridge accepts that cost for the slice of narration where getting it wrong can't be quietly fixed after a reader has already heard it.

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

Same four letters, a veterinary clinic instead of a recording booth, and this time the fragile line isn't a grief scene. It's whether a dog's limp gets called urgent.

Examline, built by Furrowdesk, listens during a vet's exam room visit and drafts the clinical note afterward, symptoms, findings, plan, so the vet doesn't have to type while they're still examining the animal. Ulrika Feste leads product there, and the same argument came up in Furrowdesk's own hiring committee, on a different kind of fragile sentence.

Same shape of fix as Tenorbridge's, built as a rule instead of a one-time save.

Jolyon Rasch, a board advisor from a traditional health-records background, made the same steelman Rowlandson made: electronic health records have used structured documentation for decades, this is just healthcare PM with a transcription feature bolted on. Ulrika's rebuttal turned on the same mechanism. Examline's eval set, 95 confirmed note-versus-recording mismatches, showed a 3.4 percent miss rate on ambiguous symptom language, phrases that sound clinically plausible either way.

The near miss that made the case real: Examline drafted a dog's limp as "intermittent, weight-bearing." The vet had actually said, half under her breath while examining the leg, something closer to "non-weight-bearing, sudden onset," a difference that decides whether the case gets referred urgently. A new vet tech, three weeks into the job, caught the mismatch by re-listening to the room audio, not because any process told her to, and flagged it before the note got filed.

Mapped straight onto PICK: the position is the same, AI PM is distinct because non-determinism changes what a passing note means, not because clinical software is hard. The impact splits the same way, a vet who over-trusts a wrong note risks a missed urgent referral, a clinic that panics and reviews every note by hand loses the entire reason to buy Examline. The cost asymmetry lands on the same shape: over-hiring a separate AI team is cheap and visible, missing a red-flag phrase is hidden until a case goes wrong. And the kill criteria transfer directly: if Examline's transcription variance on ambiguous phrasing ever hit near zero, the case for a separate check dissolves.

Swap the trigger and it still runs.
Speed: an interviewer caps you at ninety seconds. Skip straight to it: distinct discipline, because non-determinism changes what correct means, not because the tech is harder.
Cost: no budget for a full listening panel or a full vet review queue. Ship a lighter version, three-person quick listen or a single second-vet glance, only on the scenes or phrases already flagged as high-stakes, and grow it once volume justifies more.
The model got better, for real: say Castline's regeneration variance actually drops to 3. Keep the gate anyway, because close to deterministic is not the same as deterministic, and the one time it isn't is still the one that reaches a listener.

Where people run it wrong.
They answer with "AI work is just harder," instead of naming the specific reason it's a different kind of hard.
They build the steelman as a strawman, make the other side sound dumb on purpose, so the rebuttal feels easier than it actually is.
They fix a subjective-quality miss by raising a similarity or confidence threshold, instead of building a check for the axis that actually failed.

How to use it live. Before answering, ask yourself one question: could you write a fixed, repeatable test for the mistake you're describing? If yes, you're probably not describing the AI-specific case, you're describing an ordinary bug wearing an AI company's badge.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0