Explain to a non-technical executive why the last ten percent of quality costs more than the first ninety.

Scrivano listens to a recorded legal deposition and produces a speaker-tagged draft transcript for a court reporter to check and certify. Morwenna Pettigrew is the product manager who owns its transcription model. Hadleigh Coldicutt is Scrivano's chief revenue officer, and he has never once opened a training log. Threadgold and Sconce, a litigation firm, runs its whole deposition practice through Scrivano, and Ambrogio Featherwick, the firm's director of litigation support, is the one whose court reporters caught a line that almost went out under the wrong witness's name.

How to answer this, stage by stage

Nobody's grading whether you can say "word accuracy" with a straight face. They're grading whether you can turn that number into something a revenue-focused executive can actually act on.

Let's learn

Five boxes. The third one, flagging what the model is unsure about, is the whole reason the last five points are survivable at all.

Scrivano is a button a court reporter presses when a deposition wraps for the day. It listens to the recording, writes out every word, and tags who said it, so the draft is mostly done before anyone opens a laptop.

Before Scrivano, turning a four hour deposition into a rough draft took a court reporter about nine hours of listening and typing. With Scrivano running at its first real number, 80 percent word accuracy, that draft is ready in about fifteen minutes, though the reporter still spends close to five hours fixing what it got wrong. Ten weeks later, once the model reached 95 percent, that clean up time dropped to about forty minutes, mostly spent checking the handful of lines Scrivano itself flags as unsure.

Here's the turn. The five points Scrivano still misses at 95 percent are not spread evenly across an ordinary transcript. They cluster in the exact moments a legal record gets argued about hardest: two lawyers talking over each other during an objection, a witness testifying in a second language, a specialist using words a general model has barely seen before.

Six separate causes, not one blurry five percent. Each row needs its own fix, its own data, its own timeline.

What it costs at its worst: during a deposition of a structural engineer testifying as an expert witness, two attorneys spoke over each other during an objection. For about four seconds, Scrivano's draft briefly put a damaging line in the wrong person's mouth. The line scored low confidence, so it got flagged, and a senior court reporter caught it before certification, exactly the way the system is meant to work. Nothing went out wrong. It was close enough that the client called the same week Hadleigh was writing his renewal pitch.

What I would leave alone: a routine, single-witness commercial contract deposition, clear audio, plain vocabulary, sits on a flat part of this same curve. Pushing that kind of case from 95 to 98 percent is genuinely cheap, because almost everything left in it is one or two causes, not six. That's worth chasing hard. The expensive climb only shows up where the causes pile up.

The lesson: a confident number on a dashboard and an honest one look exactly the same, right up until somebody has to defend it out loud, on a call, to a client deciding whether to renew.

Now here is the same thing as a story

The short version above is what you'd actually say out loud in an interview. Read this one for the Thursday the promise almost went out wrong twice in the same week.

Scrivano's accuracy dashboard lives on a shared screen in the sales bullpen, and by nine most mornings, Hadleigh Coldicutt has already glanced at it twice. He's not an engineer, and he's never pretended to be one. What he's good at is closing a room, and for the past year the dashboard had been doing half his job for him.

Morwenna Pettigrew has run Scrivano's transcription model since before it had a single paying client. She reads an eval report the way a proofreader reads a page, catching a number drifting off before anyone else has noticed it moved.

January was launch month, 80 percent word accuracy, rough but real. By mid-March, ten weeks in, the line had climbed to 95, and Hadleigh had started opening every sales call with a screenshot of that climb. New firms signed partly because the line kept going up.

The habit thinned out in three small steps. In February, on a call with an existing client, he said "it keeps getting better every month," true, harmless, and he'd checked the rough shape of it with Morwenna first. In April, pitching a new firm, he said "give us a quarter and we'll be near-perfect," an extrapolation now, not a fact, and he didn't check with anyone before he said it, though it happened to still sound roughly right. By June, prepping the renewal deck for Threadgold and Sconce, he typed "99 percent by Q4" straight into a slide, a specific number and a specific date, and he never asked Morwenna a single question about it.

Then, on a Thursday, Ambrogio Featherwick called Scrivano's account team, not angry, just unsettled. One of his senior court reporters had caught a line during a routine review: four seconds of crosstalk during an objection, a damaging statement briefly attributed to the wrong person. The system had flagged it low confidence, exactly as designed, and nothing had gone out wrong. But Ambrogio couldn't stop thinking about the version of that afternoon where nobody happened to check that particular line.

Three things landing on the same Thursday. Only one of them was actually a coincidence.

Morwenna heard about the near miss at 1pm. Hadleigh's call with Ambrogio's firm, the renewal call, the one with "99 percent by Q4" already sitting in the deck, was set for 5pm the same day.

She spent the afternoon building the one chart she should have built months earlier: what it actually cost, in dollars and weeks, to move each point past 95. By 4:45pm she was in front of Hadleigh with it.

Both lines start in the same place. Only one of them describes what actually happened next.

He remembered the January meeting where the single blended dashboard got decided. "Nobody in that room was being lazy," Morwenna told him. "There was one client. There was one kind of case. A blended line was the honest line, then." It just stopped being honest the moment a second, harder client came along and nobody split the number back apart.

At 5pm, Hadleigh got on the call. He didn't say 99 percent by Q4. He told Ambrogio the real curve, the six causes, the near miss, and the actual cost of chasing the last points automatically. Then he offered something better than a guess: 99 percent on Threadgold and Sconce's routine single-witness matters within the quarter, since that segment was cheap to push, and a slower, staged target for the multi-expert cases, backed by the same human review that had just caught the crosstalk line.

The renewal closed anyway, two years, at 840,000 dollars a year. And the following week, the segmented dashboard Morwenna should have built in January finally got built, this time before anyone needed it to survive a phone call.

What she'd tell her January self: she built that dashboard for engineers to trust. She never once asked what happens the day somebody in sales reads a promise straight off of it.

BOUND, for pricing the three points Hadleigh thought were free

Not a way to make a bad number sound careful. BOUND turns "it's getting better" into a cost per point Hadleigh, or anyone else at Scrivano, could actually stand behind on a client call.

Four points on one line. The gap between best case and worst case is entirely about how many causes are actually left, not how hard anyone works.

The analogy that actually lands with someone who has never opened a training log: renovating an old building costs the same way quality does.

One alternative considered and rejected: hire enough human proofreaders to check one hundred percent of every transcript by hand, instead of continuing to fund model fixes. It lost, because it scales the wrong way. More depositions would mean more hours, forever, instead of a fix that keeps paying off once it's built. The AI-specific risk sitting under all of this is quiet: a wrongly attributed line inside four seconds of crosstalk reads exactly as clean and confident as a correct one, with nothing on the page to warn a tired reviewer. The guardrail is the per-segment confidence score, tied to a mandatory human check on anything flagged low-confidence or overlapping, before certification, the exact thing that caught the near miss. And the trade-off is real, not free: five months and 310,000 dollars to chase three more automated points, against holding the model at 98 percent and paying a person to catch the rare cases for good.

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

Same five letters, a crop-disease photo checker instead of a deposition tool, and this time the executive's real question isn't whether the curve bends, it's whether bending it is worth the money at all.

Loamwright checks a phone photo of a crop leaf and flags the ones likely to be diseased, so a farmer doesn't wait for a spreading blight to become obvious to the eye. Evrart Stonemere is Loamwright's product manager. Longbarrow Growers Collective, a farm co-op, runs it across its member farms, and Guntram Ashencroft, the co-op's field operations director, wants near-certain detection before trusting it across every farm in the collective.

Run BOUND on it. Break it down: after broad tuning on common blights and rust in good daylight, what's left splits into blurry or motion-blurred photos, a rare strain nobody's photographed much, backlit glare hiding a leaf's true color, and overlapping leaves hiding the actual spot. Own the numbers: eight weeks and 25,000 dollars took accuracy from 85 to 96 percent, 11 points. Sixteen weeks and 180,000 dollars took it from 96 to 98.5, only 2.5 points, over seven times the cost per point.

Different modality, same shape of problem. A photo model's last few points splinter into named causes exactly the way an audio model's do.

Swap the trigger and it still runs.
Speed: an interviewer caps you at ninety seconds. Skip straight to it: split what's left into its real, separate causes, own a real cost for the common part, and turn the unknowable rare part into a checked range, never a guessed one.
Cost: no budget to hold a promise past one month. Shrink the range honestly instead of hiding it, and say plainly what "good enough" has to mean by then.
The model got better, for real: say Loamwright's overall accuracy jumps to 99 percent right as the promise is being made. Still check by segment, since a strong headline number can hide a badly-performing slice just as easily as a weak one, maybe more easily, since nobody goes looking underneath good news.

Where people run it wrong.
They let "it's getting better" stand as the whole plan, instead of naming what's actually left.
They read one dashboard average and call the model even everywhere, instead of checking case by case, or farm by farm.
They treat a fast early climb as proof the next climb will move at the same pace.

How to use it live. Before answering, ask yourself one plain question out loud: "is what's left one shared problem, or several separate ones." Whichever it is, that answer decides whether the next stretch is cheap or expensive.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0