Explain why high-volume, low-stakes, tolerant-of-error tasks are the best first targets.

Cantillion pulls licensing records out of a dozen-plus databases every night and works out who gets paid what, every stream, every sync, every airplay. Wynford Bagshaw runs rights operations there. Nkiruka Athelstan runs partnerships, and she just watched Delacroix-Webb Publishing, one of Cantillion's biggest clients, spend six weeks fighting over how to split one song's royalties between three co-writers. She wants Cantillion's first AI feature to fix that. Wynford has to decide if she's right.

How to answer this, stage by stage

Nobody is grading whether you can recite "high volume, low stakes, error tolerant." They're grading whether you can turn those three words into a number that would actually change a real roadmap meeting.

Let's learn

What happens when a team picks its first AI project by which one gets the loudest applause in a room, instead of by the numbers?

Say a company builds a tool that reads licensing records from a dozen databases and works out who gets paid, every time a song streams, airs on the radio, or gets used in an ad. One of the jobs that tool has to do constantly, quietly, in the background, is match a new catalog record against records it already knows, because two databases almost never spell a song title or a writer credit exactly the same way.

The whole feature in five boxes. The fourth one, the flagged full check, is the reason a wrong guess never turns into a wrong payment.

Before any AI touched this, an analyst compared each pair by hand, title against title, writer against writer, duration against duration, about 45 seconds a pair. With roughly 48,000 pairs landing every month, that was 600 hours of pure comparison work a month, spread across a small team.

With the model proposing each match and flagging how sure it is, the sure ones get a ten-second glance instead of a full compare, and the unsure ones still get the same full 45-second check as always, nothing skipped. That drops the monthly total to about 181 hours. A real save of 419 hours a month.

Here's the turn. The thing that got everyone excited in a roadmap meeting wasn't this. It was something messier: teaching the model to work out which of three songwriters gets which cut of a disputed song's royalties, all on its own. Say plainly what's true: that excitement is not the problem. Chasing it first, before there's any track record, is.

What it costs at its worst: the team spends a full quarter building an autonomous split-resolver nobody can actually check, because the "right" split is the exact thing being disputed, there is no ground truth to test it against. If it ships anyway and gets one call wrong, in a way that pays real money to the wrong person, that's worse than never building either feature. Now a client is disputing the AI's answer on top of disputing the original songwriting credit. Meanwhile the 48,000-pair-a-month backlog keeps eating hours nobody touched.

What I would leave alone: this isn't an argument for never building the disputed-split feature. Some rare, high-stakes problems really are worth solving first, when there's no safe workaround and someone is walking away today. This wasn't that. Cantillion could keep resolving disputes by hand for another two quarters without anyone getting hurt, so there was no honest reason to rush it.

The lesson: don't let the size of the reaction in the room stand in for the size of the number. The task nobody claps for in a pitch meeting can still be the one worth 400 hours a month.

Now here is the same thing as a story

The short version is above, for saying out loud. Read this one for the actual Thursday the two numbers finally sat in the same room.

Wynford Bagshaw can tell, from three lines of a licensing feed, whether two records describe the same song or two different ones that just happen to look alike. He's run rights operations at Cantillion for four years, and the catalog-matching backlog has been his job, unglamorous and constant, the whole time.

Nkiruka Athelstan is good at a different thing: a room. She runs partnerships, and she'd just come off a call with Delacroix-Webb Publishing, a mid-size independent publisher and one of Cantillion's oldest clients. Three songwriters on one Delacroix-Webb track had spent six weeks arguing over the royalty split, each one certain of a different number, no contract clean enough to settle it outright. Delacroix-Webb's founder had said, almost as a joke, "can't your AI just tell us who's right?"

Nkiruka did not think it was a joke. At the next roadmap review, she pitched it as Cantillion's first real AI investment: an autonomous dispute-resolution feature, something to show every publisher client. The room liked it. Rooms like a good story.

The two candidates as they actually looked, side by side, before anyone had run a single number.

Wynford was asked to scope it. He started, as he always did, by trying to write down what "right" would even mean. For the matching backlog, right meant matching Cantillion's own master catalog, a real answer already sitting in the database. For the disputed split, there was no master answer. The correct percentage was the exact thing three people were shouting about. He couldn't build an eval set for a question nobody agreed had a right answer.

So instead he ran the numbers he could actually check. Forty-eight thousand pairs a month, 45 seconds each by hand, 600 hours. With a confidence flag doing the sorting, 85 percent get an eight-second glance, 15 percent still get the full 45-second check. That's about 181 hours. A save of 419 hours, every single month, starting the month it ships.

Then the disputed splits: six cases a month, sixteen hours each of contracts, session logs, and old statements. Automate every one of them, perfectly, no mistakes, no walk-backs. That ceiling is 96 hours.

Four real candidates, plotted honestly. Only one of them sits in the corner where a first AI project actually wants to be.

He brought both numbers to the next review, on one slide, no framing, just the arithmetic. Nkiruka looked at it longer than anyone else in the room. She hadn't been wrong that the dispute mattered, Delacroix-Webb's frustration was real. She'd just been ranking by how loud a problem sounded, not by how much it actually moved.

The same four candidates, this time sorted by the actual decision each one earns.

"We're not saying no to the dispute feature," Wynford told her. "We're saying not yet, and not alone. It needs a way to check itself, and right now it doesn't have one."

Six weeks, start to first ship. The loud project didn't disappear. It just stopped going first.

Title matching shipped in week eight. By the end of its first full quarter, it had saved rights operations just over 1,250 hours, more than one analyst's entire year, and every royalty it touched had still passed through a human glance or a human check, nothing paid out unattended. The disputed-split feature is still on the roadmap, redesigned now as a tool that hands a specialist the contracts, the session logs, and a first-pass draft split, and lets a person make the actual call.

What Wynford would tell his past self: he'd let "how big does this feel in the room" quietly stand in for "how big is this, actually," for longer than he'd like to admit. The task nobody was excited about in that first meeting was the one that was already worth more than the one everyone wanted to talk about.

BOUND, for choosing which royalty problem earns Cantillion's first model

Not a way to make an exciting idea sound suspicious. BOUND turns "this feels like the right first project" into a number Wynford, or anyone else at Cantillion, could actually defend in the room.

Three properties, one gauge. Miss any one of them and the needle doesn't actually point at a safe first project.

Four separate reasons this specific task can absorb a wrong guess without anyone getting hurt.

One alternative considered and rejected: running both projects in parallel, on split teams, so nobody had to say no to Nkiruka's pitch. It lost, because a small rights-ops team split two ways ships neither project well, and title matching's 419 hours a month were sitting there, unclaimed, for every week that decision sat unmade. The AI-specific risk underneath all of this is quiet and easy to miss: a model can confidently propose a match between two records that only look alike, in the exact same tone as a correct match, with nothing on the screen to flag the difference. The guardrail is the confidence threshold itself, calibrated against real held-out pairs, plus the mandatory full check on anything it flags, so a wrong guess never reaches a payment unattended. And the trade-off is real, not free: even the confident 85 percent still gets a human glance, not a fully hands-off pass, trading a little speed and cost for a workflow that can catch its own mistakes.

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

Same five letters, a library system instead of a rights ledger, and this time the tempting rare-and-risky project is an insurance appraisal instead of a royalty split.

Spinecheck matches book records across the branches of the Fenlake Library Consortium, so a hold or a fine follows the right physical copy instead of stalling on a title that's spelled two different ways in two different systems. Undine Thorsby leads cataloguing there, and she ran into the same choice Wynford did, on a much smaller shelf.

Run BOUND on it. Break it down: matching near-duplicate book records happens constantly, across a busy consortium, and a wrong link just delays one hold by a day, easily caught and fixed. Appraising a rare book's insurance value after fire or flood damage happens rarely, maybe four times a year, and a wrong number changes what an insurer actually pays out on a claim, hard to reverse once it's filed. Own the numbers: Fenlake's branches generate about 14,000 candidate record-matches a month, at roughly 30 seconds of manual comparison each, 116.7 hours old way. A confidence-flagged match, glance-confirmed on the easy 90 percent, drops that to about 41 hours, a save of 75.7 hours a month. Four disputed appraisals a year, at about 20 hours of expert review each, is a ceiling of 80 hours a year, or well under 7 hours a month.

A different shelf, the exact same shape of choice: a lot of cheap misses, or a few expensive ones with no real way to check them yet.

Swap the trigger and it still runs.
Speed: an interviewer caps you at ninety seconds. Name the three properties, own one real number, and say plainly that a range beats a single flattering guess.
Cost: no time to run a full comparison before the pitch. Shrink the sample to the handful of cases that would actually decide the argument, and say honestly that it's a smaller sample.
The model got better, for real: a new release makes the confidence flag far more reliable. Recheck the flag rate anyway, because a better model can still shift which task deserves to go first next, not just how fast the current one runs.

Where people run it wrong.
They rank by how much a single case matters instead of by how often it happens.
They assume "we can automate it eventually" means "we should try to automate it first."
They treat a rare, high-stakes task as unsolvable forever, instead of just wrong to build first, before there's a track record.

How to use it live. Before answering, ask yourself one plain question out loud: "does this happen constantly, does a miss cost almost nothing, and can the workflow catch it?" If you can only say yes to two of the three, say which two, out loud, and let that be the honest hedge instead of a guess dressed up as confidence.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0