Explain the pricing implications of a power-law usage distribution.

Compstack prices every account the same: $149 a month, unlimited comp pulls, whether you're an agent checking one listing or a buying firm scanning four hundred addresses a week. Silas Kirwan set that price off Compstack's own beta, sixty agents, before a single wholesale account ever touched the product. A year later, the monthly finance dashboard said gross margin was sitting near 93 percent. It was right, on average. It was also hiding twenty-four accounts running the plan at a straight loss, and nobody had gone looking until a question nobody could answer out loud forced someone to look.

How to answer this, stage by stage

Nobody is grading whether you can name a cost per pull. They're grading whether you know a flat price and a power-law usage curve are on a collision course from day one, and whether you can say, with real numbers, where that collision actually happens.

Let's learn

Compstack is the tool real estate agents and small investment buyers use to pull a comp set, similar recent sales nearby, plus a photo-based condition read and a valuation range, for one property, built by Oakmarrow Systems.

Before Compstack, building a comp set someone could actually defend meant an agent or analyst cross-referencing county tax records, recent sale prices, and permit history by hand. For someone who already knew the block, that ran close to 40 minutes a property. For someone looking at a neighborhood cold, closer to 70.

With Compstack, the same pull, record, comps, a read on the photos, a written valuation range, takes about 12 seconds. One flat price, $149 a month, unlimited pulls, whether an account runs nine of these a month or nine hundred.

Same product, same button. Two accounts can press it in completely different shapes.

The turn: more pulls was never the problem, Compstack wants accounts running it constantly. What a flat price does once usage stops looking like the sixty-agent beta it was priced against, that's the problem. A small slice of accounts, buying firms scanning whole neighborhoods instead of checking one listing, started running hundreds of pulls a month. The blended number never caught it, because thousands of light accounts were diluting it every single month.

Ninety-nine percent of accounts, and the sliver that quietly outweighs almost all of them.

At its worst: twenty-four accounts were running Compstack at a straight loss, close to $8,868 a month combined, while the company's own dashboard called the plan 93 percent healthy, right as Oakmarrow was pulling those same unit economics together for an investor update.

A flat price a wholesale account has already built a workflow around is a promise. A metered line above the point almost nobody reaches is just a setting.

What I would leave alone: the flat $149 price for the other 2,376 accounts, and the unlimited feel of the plan even for wholesale accounts below the line. They're a real segment worth keeping. Metering only above a real usage line fixes the number without pushing them out the door.

The lesson: a blended margin can hide a real loss the exact same way a blended accuracy number hides a failing segment. The fix in both cases is the same. Stop reading the average, and go look at the shape of the distribution sitting underneath it.

Now here is the same thing as a story

Read the short version above when you're in the room. Read this one when you want to feel why a margin number that cleared every single review could still have been wrong the entire time, twenty-four accounts at a time.

Silas Kirwan spent four years running pricing for two other subscription tools before Oakmarrow Systems hired him to launch Compstack's, and he had a rule he never broke: never sign off on a number you haven't traced back to a real account. For Compstack's first year, that rule meant a fifteen-minute ritual at the start of every month, pulling total AI spend, dividing by total paying accounts, and checking the result against the price.

The good months were genuinely good. Compstack launched with sixty agents in a private beta, and every single one of them used it close to the same way: check a listing before an open house, check a comp before writing an offer, a handful of pulls a week. The blended cost per account came out to a few dollars. Against $149, the margin was enormous, and it held, month after month, for the better part of a year.

It thinned in three quiet steps, and none of them looked like a mistake. Step one: word spread among a few small investment shops that Compstack could pull a defensible comp set in seconds, and a handful signed up through the same public form every agent used, no separate plan, no separate review. Step two: those accounts started running the tool differently, not five pulls a week but dozens a day, scanning whole blocks instead of checking one listing. Step three: because there were only a few dozen of them against thousands of agents, the monthly blended number barely moved. Ninety-three percent, ninety-three percent, ninety-three percent.

The trigger wasn't a dashboard alert. It was Amadea Loewe's first week. Newly hired onto the data team, she was asked to break the margin number out by account size for a routine board slide, something nobody had bothered to do in over a year because the top-line number had never once looked wrong. She came back with a question instead of a slide: why does the mean say $9.78 a month and the median say $3.06?

No single bad month. A number that held for a year, until a new hire's routine question split it open.

Silas pulled the raw logs himself that afternoon. Sorted by pulls per month, the account list didn't look like a bell curve, it looked like a cliff: two thousand three hundred and some accounts running single digits to double digits a month, then a short, steep tail of twenty-four accounts running hundreds. One of them, a regional buying firm, had logged 1,140 pulls the month before, more than the next forty accounts combined.

Twenty-four accounts, running the plan exactly as advertised, unlimited, were costing Oakmarrow about $8,868 a month more than they paid, and the company had been thirteen months from noticing.

The decision that opened the door traced back to a pricing meeting eighteen months earlier, three weeks before the public launch. Someone asked whether the beta's usage pattern would hold once real investment buyers started using it. The honest answer at the time was that none had signed up yet, there was nothing to model. "Unlimited" tested well with every beta agent who was asked, and a single flat price was one line in a pricing doc instead of three. Nobody in that meeting was picking the wrong price. There wasn't yet a wrong price to pick.

Run that meeting again with one change: a soft usage line at 60 pulls a month, set from real percentile data instead of guessed at launch, with everything above it billed at 65 cents a pull instead of folded into the flat fee. Same twenty-four wholesale accounts, same 850 pulls on the average one. Instead of a $369.50 loss, that account now nets Oakmarrow close to $144 a month, comps included. The blended margin barely changes for the other 2,376 accounts, because they were never close to the line in the first place.

One design let a beta's usage pattern set the price for every account type that would ever sign up after it. The other let the actual shape of usage, whatever it turned out to be, set where the price stopped being flat.

What I'd tell myself, back in that first pricing meeting: "unlimited" isn't a decision about a number. It's a bet that everyone will use the product roughly the same amount. Compstack's first sixty agents made that bet look safe for a year. Nobody asked what happens the day it stops being true.

BOUND, run against a curve instead of a single number

Not a story wearing a framework's clothes. This is an estimation problem with a real skew hiding inside it, and BOUND is what turns "the model is cheap on average" into a number that actually holds up account by account.

Two of these four barely move account to account. Two of them, comps and photos, are set entirely by how the account chooses to use the tool.

The whole range sits well under a thousand dollars. Doing the same work by hand sits well above it.

Three things worth stating directly, since this is where the real judgment sits. The alternative Oakmarrow's team seriously discussed, a flat cap of 100 pulls on every account, lost because it protects the margin by refusing to serve the exact accounts a comps tool is most valuable to, and does nothing to fix the price for the ones still inside the cap. The AI-specific failure worth naming is a cost curve that isn't flat by nature: every pull's real cost depends on how many comps and photos that specific account's own workflow pulls in, so a blended per-account average will always understate the heaviest users and overstate the lightest ones, the same shape as an accuracy number that looks fine until you split it by segment. The guardrail is tracking cost by pull-volume cohort as a standing metric, not a blended portfolio number pulled together once for a board slide. And the trade-off is real: Oakmarrow is keeping the full photo-based condition read for every wholesale pull, the expensive part, instead of downgrading blind bulk accounts to a text-only record, because the photo read is the actual reason a buying firm trusts a number it's never walked through in person, and it's choosing to meter the price instead of the model, so quality never has to be the thing that pays for this.

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

Same five letters, an insurance claims photo estimator instead of a comps tool, and this time the lever isn't which comp count to trust, it's which kind of account files the claim.

Claimlight is Halberd Casualty's tool for estimating repair cost from the photos a claim comes in with, and drafting the adjuster's write-up. Halberd licenses it to independent adjuster portals at a flat $89 a seat a month, unlimited claims. Cathal Feathergill owns the unit economics on it.

A routine claim, six photos, a short write-up, runs about 19 cents in AI cost. A public adjuster firm, filing on behalf of dozens of policyholders at once and submitting fifteen to twenty photos a claim to document every angle of the damage, runs closer to 58 cents a claim. A typical in-house adjuster files about 8 claims a month, comfortably inside the $89 seat. One public adjuster firm on the platform was filing 190 claims a month by itself, about $110 in AI cost against $89 of revenue, on a seat structure that assumed one person, one modest caseload.

The account that actually drives Claimlight's cost was never the busiest in-house adjuster. It was the one filing for many policyholders at once, with the heaviest photo documentation.

Same rank, different lever: the account that needed the guarded slice here was never the highest-volume in-house adjuster, it was the one filing on behalf of many policyholders at once with the heaviest photo documentation, because that's what actually drives Claimlight's cost. Cathal's team held the $89 seat price, metered per claim past a real threshold instead, and kept the same photo-based repair estimate for every account, because a public adjuster's whole case for a payout rests on Claimlight having actually looked at every angle of the damage.

Swap the trigger and it still runs.
Speed: an interviewer caps you at ninety seconds. Skip straight to it: cost every pull by what it contains, find where real usage actually breaks a flat price, meter only past that line.
Cost: there's no engineering budget this quarter to build metered billing. Ship the cheapest version first, a monthly usage report and one warning email to the accounts already over the line, before the billing change ships.
The model got better, for real: say the vision model gets twice as accurate for the same price. That doesn't change the verdict. A cheaper, better model moves the threshold, it doesn't remove the fact that a flat price needs to know the shape of usage underneath it before it can be trusted.

Where people run it wrong.
They watch one blended margin number and never once break it out by how much an account actually uses.
They fix it by raising the price for everyone, which punishes the ninety percent who never caused the problem.
They ban or hard-cap heavy usage outright instead of pricing it, and lose a real segment to whoever prices it correctly first.

How to use it live. Ask the shape question before quoting a number: "is usage on this plan roughly the same for every account, or does a small slice of them drive most of it?" That's the question that decides whether a flat price is even the right shape, before any number gets named.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0