ConceptIntermediateQuality, Cost & Token Economics / Cost modeling and unit economics / #8
Explain the gross margin problem for AI products with flat-rate pricing.
GUARD · pricing & margin risk
Satchel's board saw one healthy number every quarter. Nobody split it by how much a student actually used the app, until a proposed fix to save the margin nearly landed on the exact students who could least afford to lose it.
The direct answer
Do not price or budget a flat-rate AI product off its blended average cost per user. Track cost by usage percentile and by subscriber cohort, because a flat price hides the fact that your heaviest, most loyal users are usually the most expensive ones to serve. Protect the margin with usage-based model routing, sending only the highest-volume turns to a cheaper model, not with a blunt usage cap, because a cap punishes the students who need unlimited access most and gives them no way to see it coming or push back.
Do this, in order
Track cost per subscriber by usage percentile, not one blended average.Why: Satchel's board saw eighty two percent margin for a year while its top one percent of users were already running at a loss.
Track gross margin by signup cohort as each cohort ages, not just company-wide.Why: the erosion never showed up as a bad quarter. It showed up as the same slow bend in every cohort's curve, quarter after quarter.
Fix the cost with model-tier routing above a usage threshold, not a flat cap on turns.Why: a cap breaks the one promise, truly unlimited, that made the product spread in the first place.
Check who actually sits in the heavy-usage decile before shipping any fix aimed at it.Why: hardship-tier students, who rely on Satchel most, were nearly four times as likely to hit the proposed cap as everyone else.
Give any usage-based decision a visible number and a way to ask about it.Why: no subscriber could see their own turn count, so nobody capped could have known it was coming or said anything back.
Leave the bottom ninety percent of users alone entirely.Why: their turns already cost pennies. Building routing logic for them protects a margin that was never at risk.
How to answer this, stage by stage
Nobody is grading whether you know the phrase "unit economics." They are grading whether you would trust one healthy-looking average, or go find the segment it's hiding.
1
Scope it to one product before answering in the abstract
Say it like this
"Let's ground this in one product. Satchel is a homework help app from Quillback Learning. Students pay nineteen ninety nine a month, flat, for unlimited step by step help, typed questions or a photo of the problem. Calantha Kolstad owns pricing and product for it."
Why this works
A general question about margin turns into a lecture fast. One real product turns it into a number problem with a person attached.
2
Say your structure out loud before touching a single number
Say it like this
"I'd use GUARD here, because this is a risk question wearing a pricing question's clothes. Who holds the pricing lever and who doesn't, where the cost lands unevenly, who can't see it coming or push back, the actual product fix, and how you'd catch it early next time."
Why this works
Two seconds naming the plan tells the interviewer you have a method, not a guess with a percent sign on it.
3
Name what actually breaks about flat-rate pricing here
Say it like this
"Flat-rate pricing works when the marginal cost of one more use is close to zero, which is true for most software. It is not true here. Every question Satchel answers costs real money, a model call, sometimes a photo read. So the more a student uses the exact thing they're paying a flat price for, the less that student's subscription is worth to the company. Usage and margin move in opposite directions, and a flat price never shows you that."
Why this works
This is the reframe. A checklist answer jumps straight to "raise the price." A real one explains why flat-rate breaks for AI specifically.
4
Give the one decision
Say it like this
"So: never budget off the blended number. Split cost by usage percentile and by cohort. And when the top of that curve needs a fix, route their extra turns to a cheaper model instead of capping anyone's turns, because a cap breaks the promise and lands hardest on the students who need it most."
Why this works
This matches the direct answer word for word, which is what an interviewer is listening for.
5
Prove it with the near miss
Say it like this
"Here's what almost shipped. Blended margin had dropped from eighty eight percent to seventy one percent over a year, still looked fine on a board slide. Finance proposed a flat cap, a hundred and fifty turns a month, to stop the bleed fast. Two days before it shipped, Calantha pulled the list of who'd actually hit that number first. Nearly four times as many hardship-tier students, the ones on need-based pricing, were about to get capped as everyone else, in the middle of finals week, with zero warning."
Why this works
Four sentences, a real number, and the harm lands on someone specific. That's the compressed version of the story below.
6
Say what you'd measure, and what you'd leave alone
Say it like this
"I'd track cost per active subscriber by percentile every week, and gross margin by signup cohort as it ages, alerting when the top ten percent's cost crosses forty percent of the subscription price. I'd leave the bottom ninety percent completely alone. Their turns already cost pennies. There's no risk there to manage."
Why this works
Shows judgment, not blanket caution. Watching everything equally is the same as watching nothing closely.
7
Close on the decision, not the arithmetic
Say it like this
"So: a flat price hides where the real cost lives. Split it by percentile, fix it with routing, not a cap, and never let a margin fix ship without checking who it actually lands on first."
Why this works
Ending on the rule, not the last number crunched, is what makes this sound like judgment instead of a spreadsheet read aloud.
Let's learn
What happens to a subscription's margin when the people who use it most are exactly the people it costs the most to serve?
Satchel is a homework help app from Quillback Learning. A student types a question or uploads a photo of a worksheet, and Satchel walks through the answer step by step, math, chemistry, physics, essay feedback. Sixty thousand students pay nineteen dollars and ninety nine cents a month, flat, for unlimited use.
Before Satchel, a family paying for a private tutor spent about fifty dollars an hour, maybe two hours a week during a hard semester. Nineteen ninety nine a month for unlimited help looked like a landslide deal, and it was one of the reasons the app spread fast among students who could never have afforded a tutor at all.
Knowledge spark: why flat-rate breaks for an AI product
Most software costs almost nothing extra per use once it's built. A flat price works fine there. An AI answer is different: each one is a real model call, and it costs real money every single time. So flat-rate pricing quietly bets that most people won't use the "unlimited" part very much. When they do, the bet starts to lose.
Company-wide, the number on the board looked great: gross margin at eighty two percent. That number is not wrong. It is just an average, and averages are exactly where this kind of problem hides.
Gross margin, blended vs. split by how much a student actually uses Satchel
HealthyThinningLosing money
The board only ever saw the green bar. The top one percent of subscribers, about 600 students, cost more to serve than they pay, every single month.
The average subscriber runs about forty two tutoring turns a month, costing Satchel around three dollars and eleven cents once every cost is counted, model calls, image reads, support, infrastructure. Against nineteen ninety nine, that is healthy.
The top ten percent, about six thousand students, run closer to two hundred forty turns a month, many of them photo uploads of full worksheets, and cost about seven dollars and fifteen cents. Still profitable, but thinning fast. The top one percent, about six hundred students, run closer to nine hundred turns in an active month, mostly during exam weeks, and cost about twenty eight dollars and sixteen cents. Satchel loses over eight dollars a month on every one of them.
Satchel did not get less profitable because the model got worse. It got less profitable because the students who loved it most started using it the way it was actually advertised.
That top ten percent of students, by volume, account for fifty nine percent of Satchel's entire model bill, even though they are one tenth of the subscriber base. The top one percent alone account for twenty two percent of it. This is the turn nobody saw coming: in most software, your heaviest users are your best users, because usage costs the company almost nothing. In an AI product with a flat price, your heaviest users can be the ones quietly sinking the number everyone else is watching.
The choice that mattered
Satchel launched using the same capable, photo-reading model for every single turn, from any student, at any volume. That was the right call at launch: five hundred early users, light and uniform usage, and the whole point was proving the answers were good enough to trust. It stopped being the right call once usage stopped being uniform and started concentrating hard in one decile.
At its worst, this doesn't end in a dramatic collapse. It ends in a quiet, sensible-sounding fix that lands on the wrong people: a flat usage cap, applied evenly, that happens to hit the students who could least afford to lose unlimited access the hardest, with no warning and no way to say anything about it.
What I would leave alone: the bottom ninety percent of Satchel's subscribers need none of this. Their turns already cost a few cents a month combined. Building routing logic, alerts, or monitoring around light users spends engineering time defending a margin that was never actually at risk.
The lesson: a blended margin can be completely honest and still be the wrong number to run a pricing decision on. Eighty two percent was real. It was never the number that told anyone the top one percent was already underwater.
Now here is the same thing as a story
Read the long version below for how a healthy board slide and a genuine near miss sat two days apart, not just to be told that they did.
Calantha Kolstad had priced two products before Satchel, both flat-rate, both fine, because both were ordinary software where one more use cost the company almost nothing. Calantha brought that same instinct to Quillback Learning without thinking twice about it, because it had never once been wrong before.
For the first year, it wasn't wrong here either. Satchel launched with five hundred students, mostly light, occasional use, a question before bed, a photo of one tricky problem. Margin sat above ninety percent. The board loved the number, Calantha loved the number, and there was no reason yet to ask what was sitting underneath it.
Word started spreading in a specific place: students cramming for AP exams and finals, the exact group for whom "actually unlimited" wasn't a marketing phrase, it was the whole reason to switch from a competitor that charged per question. By the start of year two, usage had stopped being uniform. A small slice of students were using Satchel the way it had always promised they could.
Quarter by quarter, the blended margin bent down. Eighty eight percent. Eighty five. Seventy nine. Seventy one. Nobody panicked, because nobody number ever moved more than a few points at once, and seventy one percent still sounded like a company doing fine.
Then finance ran the numbers ahead of the year-end board meeting and didn't love the trend line. Someone proposed the fast fix: a hundred and fifty turns a month per student, framed internally as a "fair use policy," with an automatic message once a student crossed it. It would move the number. It could ship in a sprint. Calantha signed off on it on a Tuesday.
Two days before it went live, Calantha was reviewing the actual support message that would go out to capped students, mostly to make sure the tone was kind, and pulled the list of who would receive it in month one to spot-check a few names.
The list wasn't random. It was nearly four times as likely to contain a hardship-tier student as anyone else on the platform.
Satchel offers a need-based tier at half price for verified low-income families. Those students make up twelve percent of subscribers overall. On the list of students about to hit the cap in its first month, during finals week, they were thirty four percent. A student on that tier isn't using Satchel for fun. For a lot of them, it's the only tutor they have.
Calantha didn't have a name and a face for one specific student that week, the way a support ticket sometimes hands you. What surfaced instead was a pattern: no subscriber, hardship tier or not, had ever been shown their own turn count. Nobody could see a cap coming. Nobody had a way to ask why they'd been capped, or to say "this is the week I actually need it." The policy would have shipped as a silent wall, hitting hardest exactly where the app mattered most, and nobody on the receiving end would have had any lever at all.
The decision that opened the door went back to launch, to serving every turn with the same capable model regardless of who was asking or how often. That made complete sense at five hundred light users. Nobody ever sat down and decided it should still be true at sixty thousand, with a top decile running two hundred and forty turns a month. It just kept being true because nobody had reason to look at it again.
Calantha killed the cap two days before launch and asked engineering for a different fix instead: route any turn past a set volume threshold to a smaller, cheaper model for routine, practice-level questions, while keeping the full-strength model for anything genuinely complex or photo-heavy, no matter how many turns a student had already used that month. No cap. No message telling anyone they'd used too much of something they were promised was unlimited. Just a quieter model doing the easy half of a very active month.
Run that quarter again with the routing fix live instead of the cap: the top one percent's average cost falls from twenty eight dollars and sixteen cents to about eighteen dollars, still costly, but no longer a guaranteed loss on most of them, and nobody on a hardship plan gets a message telling them they've hit a wall during finals week.
What I'd tell myself, back at launch: we asked whether the model was good enough. We never asked what it would cost the day the students who trusted it most started using it exactly the way we told them they could.
GUARD, the five checks behind a margin nobody was watching by segment
This isn't a pricing question wearing numbers. It's a risk question, and GUARD is what stops one comfortable average from hiding who the fix would actually land on.
GGroups. Who holds the lever, and who doesn't?
Finance and product hold the lever, they can change the price, change the model, or change the rules whenever the numbers ask them to. Subscribers sit underneath one flat price with no lever at all, especially the heaviest users on Satchel's need-based hardship tier, who rely on it the most and can see none of this coming.
Naming both sides before touching a number is what keeps this from turning into a spreadsheet exercise.
UUnequal. Where does the cost actually land?
Usage follows a steep curve, not a bell curve. The top ten percent of subscribers drive fifty nine percent of the entire model bill. The top one percent drive twenty two percent of it alone, and inside that heaviest decile, hardship-tier students are overrepresented three to one against their overall share of the subscriber base.
The unevenness isn't random noise. It's concentrated on the exact students the product exists to help most.
Quarterly blended margin, one year of quiet erosion nobody flagged
Actual, four quartersProjected if unchecked
No single quarter looked alarming, three to six points at a time. Unchecked, the same slope crosses zero blended margin in about six more quarters, roughly a year and a half out.
AAbility to contest. Who never gets to push back?
No subscriber, hardship tier or not, can see their own turn count against any limit, because Satchel never published one. When the hundred and fifty turn cap was proposed, no student had any way to know it was coming, appeal being capped mid-finals, or explain that this was the one week they actually needed it. The people who'd feel the fix had zero visibility into it, and the people with visibility weren't the ones who'd feel it.
This is the hardest step, and the one most pricing answers skip entirely. A fix nobody affected can see coming isn't a policy, it's a wall.
Finance held the lever. A hardship-tier student doing homework at 11pm held nothing, and was about to get the same outcome anyway.
RReduce. The specific product decision.
Route turns past a defined monthly threshold to a smaller, cheaper model for routine, practice-level questions, while keeping the full model for anything complex or photo-heavy no matter how many turns a student has used. No cap on turns, no rationing, "unlimited" stays true in the one sense students actually care about: nobody is ever told to stop asking.
A real design decision, not a policy document. The bar isn't zero cost on heavy users, it's a threshold, checked against an eval set so routed answers don't quietly get worse.
DDetect. How you'd know before the board meeting.
Track cost per active subscriber by percentile, P50, P90, P99, every week, not one blended number a quarter. Track gross margin by signup cohort as it ages, watching for the same slow bend showing up cohort after cohort. Alert when a cohort's P90 cost crosses forty percent of the subscription price, catching the drift a couple of quarters before it ever reaches a board slide.
The blended number is exactly the thing that let this run for a year. Detection has to live below it, not inside it.
The proposed cap had three working steps and a fourth that was never built at all.
Three things worth stating directly, since this is where the real judgment sits. The alternative finance tried first, and Calantha rejected two days before it shipped, was the flat hundred and fifty turn cap, applied the same to every subscriber. It lost for two reasons: it broke the one promise, truly unlimited, that made Satchel spread faster than any per-question competitor, and it landed nearly four times as hard on hardship-tier students with no way for any of them to see it coming or contest it. The AI-specific failure worth naming by name is usage-cost inversion: in most software your heaviest, most loyal users are your best users, because usage costs the company almost nothing extra. In a flat-rate AI product, your heaviest users can be the ones quietly losing you money, and a flat price hides exactly that inversion from anyone only watching the blended average. The guardrail is percentile-based cost monitoring paired with threshold-based model routing, checked against an eval set so a routed answer doesn't quietly fall below the same accuracy bar. And the bar here was never zero cost on any single heavy user, no flat-rate product can promise that. It's a percentile bar: keep the P90 segment's cost under forty percent of the subscription price on a trailing basis, checked weekly, not a demand that every single active month turns a profit on every single student.
And if you want to be sure it really works, try it somewhere else
Same five letters, small law firms instead of students, and this time the group that can't push back isn't underprivileged, it's just outnumbered by its own paperwork.
Redline is a contract review tool from Griswall, a legal AI company. Small firms pay a flat monthly fee for unlimited contract uploads, flagging risky clauses before anyone signs. Emberlynn Achterman runs pricing on it.
The build-up: Griswall's blended margin looked steady for over a year, until a firm doing high-volume merger and acquisition work started uploading two hundred page agreements daily instead of the short leases and vendor contracts most firms upload. That one firm's monthly cost ran higher than four typical firms combined, on the exact same flat plan.
The decision Emberlynn would take back
Pricing Redline as one flat tier for "unlimited contracts," with no distinction between a two page NDA and a two hundred page merger agreement, because at launch almost every uploaded document was short.
G, groups. Griswall's pricing team holds the lever. Small firms on the flat plan hold none, especially solo practitioners who took on one big, document-heavy case and had no way to know it would flip their account from profitable to a loss. U, unequal. Document length and complexity, not firm size, drive the real cost. A solo lawyer handling one merger can cost more than a five-person firm doing routine leases all month. A, ability to contest. A firm has no visibility into its own document-cost total and no warning before a plan change lands on them mid-case. R, reduce. Route documents past a page and complexity threshold to a review queue billed separately, instead of raising the flat price for every firm regardless of what they actually upload. D, detect. Track cost per firm by document complexity weekly, not firm count or revenue tier, so one merger-heavy account doesn't hide inside an average built from mostly short leases.
Swap the trigger and it still runs.
Speed: an interviewer caps you at ninety seconds. Skip straight to it: never price a flat AI product off the blended average, split cost by usage percentile, and fix the tail with routing, not a cap that lands on whoever uses it most.
Cost: there's no budget this quarter for both a routing rebuild and a full monitoring dashboard. Monitoring wins first. You cannot safely fix a cost curve you cannot see broken out by segment yet.
The model got better, for real: say Satchel's underlying model gets cheaper per call across the board. That helps the blended number, but it doesn't undo the concentration, the top one percent still uses the model far more than everyone else, just at a lower price per call.
Where people run it wrong.
They watch one blended margin number and call it healthy, without ever splitting it by how much any one group actually uses the product.
They fix a thinning margin with a price hike or a usage cap applied to everyone equally, without checking who that flat rule actually lands on hardest.
They build "unlimited" once at launch and never revisit which model serves which turn, even after usage stops looking anything like it did on day one.
How to use it live. Say the real question out loud before answering: "is this margin number an average across everyone, or does it hold up for the people who use this the most?" That buys a beat to think, and it's almost always where the real answer is hiding.
Flashcards (tap any card to flip it)
1 · THE FRAMEWORK
What framework fits a question about a flat price quietly losing money, and why?
Tap to flip
ANSWER
GUARD, for risk. The real question is who holds the pricing lever, who doesn't, and who a margin fix would actually land on, exactly what GUARD is built to find.
2 · THE PERSON
Who is this answer about?
Tap to flip
ANSWER
Calantha Kolstad, who owns pricing and product for Satchel at Quillback Learning, and had priced two ordinary flat-rate products before this one without issue.
3 · THE OLD DEFAULT
What default, set at launch, quietly stopped making sense as Satchel grew?
Tap to flip
ANSWER
Serving every turn, from any student at any volume, with the same capable model. Fine at five hundred light users. Expensive once a top decile ran two hundred forty turns a month.
4 · THE GAP
What did the eighty two percent blended margin hide?
Tap to flip
ANSWER
The top ten percent of subscribers ran sixty four percent margin, and the top one percent ran negative forty one percent, losing money every month, hidden inside one healthy average.
5 · THE OLD DECISION
What decision did Calantha reject two days before it shipped, and why did it almost happen anyway?
Tap to flip
ANSWER
A flat cap of a hundred fifty turns a month for every subscriber. It almost shipped because it was fast and would move the margin number before a board meeting.
6 · THE NUMBER
Fill in the blank: on the list of students who would hit the proposed cap first, hardship-tier students made up thirty four percent, against ___ percent of the overall subscriber base.
Tap to flip
ANSWER
Twelve percent. Nearly four times their overall share, meaning the cap would have hit the app's most dependent students first, during finals week.
7 · THE REPLAY
Same quarter, new fix, what changes?
Tap to flip
ANSWER
Turns past a set volume route to a cheaper model for routine questions, the full model stays for anything complex. The top one percent's average cost falls from about twenty eight dollars to about eighteen, and no student ever sees a cap message.
8 · CROSS-PRODUCT TRANSFER
Section 4 answers this same question again for a different product. Which product, and what plays the role of "usage volume" there?
Tap to flip
ANSWER
Redline, a contract review tool from Griswall. Document length and complexity play the role turn count plays for Satchel, a single merger agreement can cost more than dozens of short leases.
Check yourself Score: 0 / 0
Fill in the blank
1. Satchel's company-wide blended gross margin read ___ percent, while the top one percent of subscribers by usage actually ran a margin of ___ percent.
Show hint
Look at the three bars in the first chart in Section 1.
Show answer
82 percent, then negative 41 percent. The blended number was honest. It just described almost nobody at the actual top of the usage curve.
True or false
2. True or false: because Satchel's blended margin only dropped a few points each quarter, three to six points at a time, it was reasonable for nobody to treat the trend as urgent.
True
False
Show hint
Look at the projected line in the quarterly margin chart in Section 3, and where it crosses zero.
Show answer
False. Each quarter's drop looked small on its own, but the same slope, left unchecked, crossed zero blended margin in about six more quarters. A slow, steady bend is still a real trend, and it's exactly the shape that a quarter-by-quarter view makes easy to wave off.
Multiple choice
3. Why did routing heavy usage to a cheaper model win over a flat cap on turns?
A. Routing is technically simpler to build than a usage cap.
B. A cap breaks the "unlimited" promise driving Satchel's growth, and it landed nearly four times as hard on hardship-tier students who rely on it most, with no way for them to see it coming.
C. Routing makes every answer higher quality than the original model did.
D. A usage cap is illegal for subscription products to enforce.
Show hint
Look at the near miss in Section 2, and who was on the list of students about to be capped first.
Show answer
B. Routing keeps the core promise intact and controls cost only on the highest-volume, mostly routine turns. A cap is a blunt dial that would have hit the neediest heaviest users hardest, invisibly.
Short answer, name the rejected alternative
4. What did finance propose first to fix Satchel's thinning margin, and why did Calantha kill it two days before launch?
Show hint
Look at the R step in the framework recap, where the rejected alternative is named directly.
Show answer
Model answer: A flat cap of a hundred fifty turns a month for every subscriber, framed as a "fair use policy." Calantha killed it after finding that hardship-tier students, who make up twelve percent of subscribers, were thirty four percent of who'd hit the cap first, with no way for any of them to see it coming or say anything about it.
Short answer, apply it yourself
5. Pick a flat-rate app you use, or one you've heard of, that promises "unlimited" something. Name one group of its users who probably costs it the most, and how you'd check.
Show hint
Think about who uses the "unlimited" part of the product the hardest, not the average user.
Show answer
Model answer: A flat-rate AI writing app might lose money on users who paste in entire long documents for rewriting every day, versus someone who polishes one short email a week. I'd check by pulling cost per active user split by how many words get processed a month, not by looking at one company-wide average.
Multiple choice
6. The top ten percent of Satchel's subscribers drive fifty nine percent of the total model bill. If that concentration grew to seventy five percent next year with subscriber count unchanged, what would you expect to happen to blended margin, all else equal?
A. Blended margin would improve, since more usage means more value delivered per subscriber.
B. Blended margin would keep falling, since a growing share of total cost concentrating in a small high-volume group pulls the average down further.
C. Blended margin would stay exactly the same, since total subscriber count didn't change.
D. It's impossible to say anything without knowing the exact subscription price.
Show hint
Think about what happens to the average when a fixed group's share of total cost keeps climbing, even if nobody new signs up.
Show answer
B. The same mechanism driving Q1 through Q4's decline: as the heaviest decile's share of total inference cost climbs, it drags the company-wide blended average down with it, even with the subscription price and subscriber count both held flat.
Before you close the answer
Why this works
Tests whether you'll trust a healthy blended number or go looking for the usage curve hiding inside it, and whether your fix for a cost problem accounts for who it actually lands on. Most candidates jump straight to "raise the price" or "cap usage" and stop there.
Follow-up traps
"Why not just raise the price for everyone instead of building routing logic?" Response: a flat price hike punishes the ninety percent of light users who were never the problem, and does nothing to fix the actual cost concentration sitting in the top one percent.
"Isn't routing to a cheaper model just a quieter version of the same cap, since quality might drop?" Response: no, because routing is gated by an eval set checking accuracy stays above bar, and it only applies to routine, practice-level turns, complex or photo-heavy questions still get the full model regardless of a student's total volume that month.
If pressed
The routing threshold isn't one fixed number for every student. Quillback set it relative to each student's own trailing volume, so a student having one unusually heavy week before a single exam doesn't get routed the same way a student who is heavy every week does, the threshold tracks a sustained pattern, not a single busy Tuesday.
From U2xAI Academy
From answering questions to owning outcomes.
A live workshop where you ship a working AI agent, defend a launch decision, and walk away with a portfolio recruiters can't wave off, not just more questions to study.
