How do you find opportunities where AI changes the unit economics rather than just the experience?

Tallyrook is Vintermark's pricing engine. It watches what competitors charge across the web and hands an online seller a new price back, all day, without anyone opening a spreadsheet. Yrsa Aabakken owns its roadmap. Oxcombe Outdoor Supply, a fourteen thousand SKU camping and gear retailer, is one of its biggest clients. For most of a year, the number everyone watched on Tallyrook was how satisfied Oxcombe's sellers felt. It barely moved. Underneath that flat number, something much bigger already had.

How to answer this, stage by stage

Nobody is grading whether you found a bigger number. They are grading whether you would have gone looking for it at all.

Let's learn

Here's what happens when a feature makes everyone a little happier and the actual bill barely changes size.

Say a company builds Tallyrook, a tool that checks what competitors charge and hands a seller a new price back, without anyone opening a spreadsheet.

Four steps in one seller's pricing workflow. Only one of them was ever expensive.

Before Tallyrook, Oxcombe Outdoor Supply ran this whole workflow by hand. Three pricing analysts, a hundred eighty thousand dollars a year in loaded pay, checking competitor prices at roughly three dollars a check: six minutes of work at about thirty dollars an hour. That budget could only cover about eight hundred of the fourteen thousand SKUs on a real weekly cycle, the ones that sold enough to be worth an analyst's morning. The other thirteen thousand two hundred sat on prices nobody had rechecked in a quarter.

Same catalog. One version of it only ever got checked for one SKU in seventeen.

Now, an automated check costs about four cents: a scrape, a match against the right competitor listing, and a price recommendation, with no analyst involved unless the model flags a shaky match. Every one of the fourteen thousand SKUs gets checked, most of them daily. Oxcombe's whole pricing operation, one analyst left plus the automation bill, now runs about sixty one thousand dollars a year. Full coverage, at a third of the old cost.

What it costs at its worst: if nobody had run that math, Vintermark would have kept pitching Tallyrook on "sellers like it a bit more," a soft story that is easy to cut from a roadmap. The real fact, a ninety eight percent drop in the cost of running a whole catalog, would have sat undiscovered, and it was the actual reason a client would sign a longer contract.

What I would leave alone: the price suggestion screen itself, the part sellers actually rate, is fine exactly as it is. Nobody would notice if it never got another design pass, because it was never carrying the real value. Leave the screen alone. The win never lived there.

The lesson: a number can hold perfectly still and still be hiding something enormous underneath it. A satisfaction score measures how a screen feels. It was never built to catch a cost line falling ninety eight percent somewhere nobody's dashboard was pointed at. If a metric only watches the interface, it will miss the moment the real economics change, every time, and nobody will know until someone happens to run the math.

Now here is the same thing as a story

Use this version when you have the time. The short version above is what you would actually say out loud in the room. This one is for feeling why it matters.

Every Monday, Yrsa Aabakken pulls the same three numbers before she opens her inbox, out of habit more than need. She has run Tallyrook's roadmap at Vintermark for four years, since before the model could tell a hiking boot from a hydration pack, and she built the very first competitor matching pass herself, on a spreadsheet, checking forty SKUs by hand just to prove the idea would work at all.

Tallyrook launched properly eighteen months ago, and for the first stretch, the number Yrsa watched was the one everyone watches: how satisfied a seller felt with the price suggestions Tallyrook handed them. It opened at three point six out of five. Sellers said the screen was nice. Nobody said much more than that.

Oxcombe Outdoor Supply came on as a client that spring. Fourteen thousand SKUs, three pricing analysts, and a habit going back years of manually checking whatever competitor prices there was time for. Within two months of turning Tallyrook on, satisfaction crept up to four point one. It sat there. Week after week, four point one, sometimes four point zero, once briefly four point two before sliding back. Steady. Unremarkable. The kind of number a roadmap meeting glances at and moves past.

Nobody was hiding anything. The number just never moved enough to make anyone stop and ask what was actually happening underneath it.

Then, in October, Vintermark's VP of Growth, Bjarke Hemsley, brought that same flat four point one into a board prep deck, arguing the pricing engine had plateaued and the roadmap should shift toward something friendlier to use, starting with a one tap price approval button. His deck had eleven months of satisfaction data behind it. It looked responsible. It looked like evidence.

One of these was already true. Nobody had asked it the right question yet.

Yrsa almost let it go. The number really had sat still. Nothing in her own dashboard gave her a reason to ask a different question.

What stopped it was smaller than a meeting. A finance partner, prepping Oxcombe's renewal call that same week, asked Yrsa for one line: how much had Tallyrook actually saved this client, in dollars, this year. Not a satisfaction score. A dollar figure, for a slide.

Yrsa didn't have one ready. She built it in an afternoon, pulling Oxcombe's own pricing operations line from the shared finance sheet, before Tallyrook and after.

Before: three analysts, a hundred eighty thousand dollars a year, and real coverage on maybe eight hundred of the fourteen thousand SKUs, the ones that sold enough to be worth checking. The other thirteen thousand two hundred sat on prices nobody had rechecked in a quarter.

After: one analyst left, handling only the cases the model flagged as uncertain, plus the automation bill itself. Sixty one thousand dollars a year, total, covering every single SKU, most of them rechecked daily.

Yrsa remembers the exact meeting where satisfaction became the metric on Tallyrook's own dashboard. It made sense at the time. Tallyrook was young, mostly a nicer screen wrapped around a rough model, and satisfaction was the number every other product review already used, sitting right there on the same dashboard as five other features. Nobody sat down and asked whether a screen feeling number could ever catch a cost line collapsing behind it. It just came with the template.

She brought the real number to Bjarke before the board deck shipped. Cost per repricing check, three dollars down to four cents. Ninety eight percent, not a rounding error. Oxcombe's own renewal, which had been shaky, closed for three years within the month, on the strength of that one dollar figure, not the satisfaction score that had sat flat the whole time.

The one tap approval button still shipped, a few months later. It is a fine button. It just is not the reason anyone renewed.

What I'd tell myself, back when the dashboard template asked for a satisfaction number and nothing else: the screen was never lying. It just was never built to see the ledger underneath it. I would have gone looking for the dollar number a year earlier, not because the satisfaction score was wrong, but because I never should have trusted it to tell the whole story on its own.

LEAD, so a cost story doesn't hide behind a satisfaction score

Not a way to prove a feature is popular. LEAD forces you to name the real cost line before anyone gets to call something a unit economics win.

Four questions, multiplied together. That is the whole tool.

The cost line finished falling weeks before the meeting that finally looked at it.

Two features can sit at opposite corners of this chart and still show up in the same roadmap deck.

Three questions, asked before a single design screen gets drawn.

The recap, one line per letter: link "changes the economics" to a real cost per transaction, not a satisfaction score. The early signal is cost per unit, tracked weekly, and it moved weeks before anyone's dashboard caught it. Name the abuse plainly: a feature that trims a few seconds off an already cheap step, wearing the language of a real cost collapse, and a metric like acceptance rate that can be hit without the underlying cost ever moving. And the decision is what makes any of this real: price the workflow first, build against the priciest step, and route anything the model is unsure about to a person rather than trust a round confidence number.

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

Same four letters, a freight brokerage instead of a webstore, and this time the metric that almost fooled everyone was speed, not satisfaction.

Ingimar Kestholm runs pricing operations for Redcastle Freight Partners, a mid size freight brokerage. Loadline, their AI quoting tool, used to get judged the same way Tallyrook almost did.

Before Loadline, a broker manually checked four or five competing carrier rates to quote one shipping lane: about twelve minutes, at a loaded rate near forty five dollars an hour, roughly nine dollars a quote. Only the big repeat customers' lanes got that real check. Smaller, one off shipments got a flat rate guess, because nobody had time to check them properly, and that guess quietly bled margin on lanes nobody was watching.

Loadline's automated lane check costs about fifteen cents. Every quote now gets a real competitive check in seconds instead of a guess. The team's headline metric was quote turnaround time, which fell from two hours to about four minutes, and sales loved it. That number almost became the whole story.

The turnaround number was true. It just was not the number that mattered.

Mapped onto LEAD, the shape holds. The link is a real cost per quote and the margin sitting behind it, not how fast the quote appears on screen. The early signal is the same lesson Yrsa learned: a leading cost number can already be telling the real story while the metric everyone is watching stays flat, or in this case, looks great for the wrong reason. The abuse Ingimar watched for was the same too: an "instant quote PDF download" button got pitched as an economics win, but cost per quote never moved, since the quote was already computer generated; only the delivery screen changed. And the decision matched Yrsa's: price the workflow, chase the step that actually costs the most, and keep the win moving to the next expensive manual step instead of resting on it.

Swap the trigger and it still runs.
Speed: an interviewer caps you at ninety seconds. Skip straight to it: map the workflow's cost, chase the priciest step, check whether the transaction cost actually moved.
Cost: no budget for a dedicated finance review. Whoever owns the model spot checks a random sample of quotes or reprices by hand every week instead of waiting on a quarterly report. Slower to build, but it closes the same gap.
The model got better, for real: say the matching model's accuracy genuinely improves. Keep chasing cost per transaction anyway, because the point was never how right the model looked. It was always what the manual step used to cost, and whether that cost actually fell.

Where people run it wrong.
They let a feature borrow the language of the last real economics win because it feels similar, without checking the actual cost number.
They watch a satisfaction or turnaround metric and assume flat, or good, means nothing else is worth checking.
They chase the most visible manual step instead of the most expensive one, and those are not always the same step.

How to use it live. When an interviewer asks you to find an AI opportunity, buy yourself a second by asking out loud what the workflow actually costs today, per unit, before naming a single feature. That question does most of the answer already.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0