The Margin Floor Kit (free)
The complete starter kit for the Margin Floor Reset: turn a small AI software company’s invoices and usage export into a pricebook with a floor under every plan, and get paid $750 for your first reset and $1,250 for every one after. Free, and it stays free.
WHAT YOU ARE BUILDING (THE 60-SECOND OVERVIEW)
Tokens got a market price on August 19, 2026: Stripe agreed to buy OpenRouter, the gateway that routes AI requests across 400-plus models, for more than $7 billion, and DHH showed on Lex Fridman (August 26) that one coding job cost $23 on one model and about $550 on another for comparable output. Thousands of small AI software products sell a flat monthly plan on top of that swing. One heavy customer can run many times the median, a longer reasoning mode raises the bill with no change to the price sheet, and the vendor keeps every dollar of the variance. You are building the smallest service that fixes the SELL side: a seven-day reset that measures what one completed customer action costs, sets the minimum contribution each plan must keep (the margin floor), converts the leftover into included use plus a paid expansion pack, and hands the founder a migration plan and a billing handoff. You build it once from a spreadsheet, three documents, and a prompt pack, then deliver it for $750 the first time and $1,250 after. No code. You never touch the founder’s billing system or their customers.
START HERE: THE 5-MINUTE FIRST STEP
Pick any AI tool you pay for. Open its pricing page and find the plan you are on. Now try to fill in one line: “This plan includes ___ actions before it loses money at a cost of $___ per action.” You will not be able to, and neither can most of the founders who wrote those pages. Write the sentence you just could not finish; it is the opening line of your outreach note. Do it for a company on Product Hunt’s AI launch page and you have a prospect.
WHAT’S INSIDE
Who to pitch, who to skip (the Buyer Qualifier)
What to ask for and how to keep it safe (the Minimum Data Request and the Privacy Floor)
What one completed customer action really costs (the Feature-to-Cost Map)
How much model budget each plan can carry (the Margin Floor Calculator)
What the heaviest user does to the plan (the Heavy-User Stress Test)
Included use, expansion pack, exception policy (the Pricebook Builder)
How to change a price without losing the room (the Migration Decision and the Notice Templates)
Who flips which switch after you leave (the Handoff Checklist)
The four prompts that do the tedious parts (the Prompt Pack)
The 50-founder tracker, the outreach note, the teardown script, and the pricing (the Founder Sales Pack)
The worked example, filled end to end (the Policy Summaries Reset)
The 7-day delivery SOP and the works-every-time checklist
1. THE BUYER QUALIFIER
Score each candidate on six checks. Pitch at 5 of 6 or better.
# | Check | Pass looks like |
|---|---|---|
1 | Vertical AI SaaS, 3 to 30 people | A product for one industry (insurance, legal ops, property, clinics, logistics), not a general chatbot wrapper |
2 | Flat monthly plan | A per-seat or per-company price with “unlimited” or an unstated fair-use line |
3 | A usage-variable AI feature | Summaries, drafts, extractions, research runs, generated reports: anything that calls a model per customer action |
4 | 20 or more paying customers | Enough accounts to have light and heavy cohorts; a directory listing, review count, or customer logos usually tells you |
5 | 30 days of usable records | Vendor invoices (OpenAI, Anthropic, a router) and an action count from the app or its analytics |
6 | A reachable founder | A named founder on the site, a business email or contact form, a LinkedIn profile that answers |
Where to find candidates: Product Hunt’s AI launch pages (new products daily, the makers named on every launch page) and G2’s AI software categories (filter by market segment and industry, then check the vendor’s size on its own site or LinkedIn). Log every candidate in the tracker (section 10).
Vertical rule: pick ONE vertical and stay in it. The cost map you build for the first insurance tool is 80% of the map for the second. Specialists compound; generalists restart.
2. THE MINIMUM DATA REQUEST AND THE PRIVACY FLOOR
Ask for the least data that makes the math true. Send this list at intake:
Plan names, prices, and the number of paying customers on each plan.
The last 30 days of model-vendor invoices or usage dashboards (OpenAI, Anthropic, Google, a router such as OpenRouter or Vercel AI Gateway), totals by day or by model.
The count of completed customer actions in the same 30 days (from the app database, analytics, or a product export), by plan if possible.
Other variable costs per customer: hosting, storage, third-party APIs, payment fees.
Refunds, credits, or support tickets tied to limits or slowdowns in the period.
The current pricing page and the terms language about usage.
The privacy floor, written into your engagement note: you receive aggregates and redacted exports, never customer names, prompts, or documents. You do not paste any client data into an AI tool without the founder’s written authorization for that specific file. You delete the working files 30 days after handoff unless the founder asks you to keep them. You are a pricing analyst, not a data processor, and the note says so.
3. THE FEATURE-TO-COST MAP
One sheet tab, one row per AI feature. Columns:
| Feature | The customer outcome (one unit) | Model calls per outcome | Completion rate | Model cost per outcome | Other variable cost per outcome | Total cost per completed outcome |
|---------|----------------------------------|-------------------------|-----------------|------------------------|----------------------------------|----------------------------------|
Rules: the unit is the thing the customer would name (“one policy summary,” “one drafted response,” “one research run”), never a raw API call. Divide the period’s model spend for that feature by COMPLETED outcomes, so retries and failures land in the cost where they belong. If the vendor invoice does not split by feature, split it by the share of calls each feature made and say so in the notes column. If nothing splits, treat the product as one feature and move on; a rough true number beats a precise wrong one.
The invoice-classification prompt (section 9) does the sorting when the export is messy.
4. THE MARGIN FLOOR CALCULATOR
One tab per plan. The founder chooses the floor; you calculate everything else.
Row | Field | Example | Formula |
|---|---|---|---|
A | Plan price per month | $49 | input |
B | The floor: contribution the plan must keep | $34 | the founder’s choice, dollars or a percent of A |
C | Other variable costs per customer | $4 | from the cost map |
D | Model budget per customer | $11 | A minus B minus C |
E | Cost per completed outcome | $0.25 | from the cost map |
F | Safe included outcomes | 44 | D divided by E, rounded DOWN |
G | Included outcomes to publish | 40 | F rounded down to a clean number |
H | Expansion pack size | 20 outcomes | the founder’s choice |
I | Pack serving cost | $5 | H times E |
J | Pack price | $12 | at least I times 2, rounded to a clean number |
K | Pack contribution | $7 | J minus I |
Two rules the founder will try to break. First, the floor is chosen BEFORE you look at customer behavior, or the number drifts toward whatever keeps the heaviest user happy. Second, round down every time. A published allowance the plan cannot actually carry is the problem you were hired to remove.
If the founder cannot name a floor, offer three: the plan’s current average contribution, the number that keeps the plan profitable at the 90th-percentile user, and the number the founder would defend to an investor. Let them pick. Log which one they picked and why.
5. THE HEAVY-USER STRESS TEST
Three scenarios per plan, one row each: normal (the median customer’s monthly outcomes), heavy (the 90th percentile, or 5x the median if the export is thin), and failure (heavy plus a 30% retry rate, which is what a model swap or a bad week looks like).
| Scenario | Outcomes per month | Model cost | Other variable | Contribution today (unlimited) | Contribution under the new pricebook |
|----------|-------------------|------------|----------------|-------------------------------|--------------------------------------|
The test exists to produce one sentence for the founder: “Your heaviest customer costs you $X a month and pays you $Y.” Say it plainly in the delivery call and let the number do the persuading. Never round it up for effect.
6. THE PRICEBOOK BUILDER
The output document, three tiers, one page. For each tier: the price, the included outcomes (row G), the expansion pack (rows H to K), the exception policy, and a one-line promise in plain language. Rules:
Keep the entry tier’s allowance generous enough that a normal customer never sees the meter. The meter is for the top decile, not for everyone.
One expansion pack per product, same size and price across tiers, so support never has to explain a matrix.
The exception policy names what happens past the hard cap: a pause with a one-click pack purchase, never a silent overage bill. Surprise invoices are how a price change becomes a churn event.
Say what is NOT metered. If the customer’s day-to-day use of the product is unmetered and only the AI actions carry an allowance, the page says so in the first line.
Write the promise like a person: “40 policy summaries a month, then $12 for 20 more. Nothing else in the product is counted.”
7. THE MIGRATION DECISION AND THE NOTICE TEMPLATES
Three options, one decision, recorded in a one-page document with the reason:
Grandfather. Existing customers keep the old terms for a fixed window (90 days is the default) and move at renewal. Default for any product under 200 customers.
Phased notice. Everyone moves on a date at least 30 days out, with the allowance published now.
Immediate. New customers only, effective today; existing customers untouched until a later decision. Use when the founder is not ready to touch the base.
The notice, for the founder to send under their own name:
Subject: A change to how [Product] counts AI work, starting [date]
Hi [first name],
From [date], every [Plan] seat includes [40] [policy summaries] a month, and you can add [20] more any time for $[12]. Everything else in [Product] stays uncounted, and your price does not change.
Why: each summary costs us real model time, and a small number of very heavy months were being carried by everyone else. This keeps the plan fair and keeps us here for you.
Your account is grandfathered on today's terms until [date + 90 days]. If you regularly run past [40], reply to this email and we will find the right fit before anything changes.
[Founder name]Objection responses, kept to one line each: “We were promised unlimited” (the grandfather window is the promise kept, and the allowance was set from real usage so a normal month never meets it); “We will switch tools” (offer a call and a 60-day allowance bump, never a discount); “Why now” (the honest answer: model costs now vary by month, and a floor protects the product’s continuity).
The notice goes out only after counsel confirms the notice period against the founder’s own terms of service and the auto-renewal laws of the states where customers sit.
8. THE HANDOFF CHECKLIST
You do not implement. You hand the founder a list with an owner and a date for every line:
[ ] Billing: create the expansion pack as a product in Stripe, Chargebee, or the billing tool; set the grandfather flag on existing subscriptions. Owner: founder or developer.
[ ] Meter: the app counts completed outcomes per account and displays the remaining allowance. Owner: developer.
[ ] Pricing page: the new tiers and the plain-language promise. Owner: founder.
[ ] Support: three macros (allowance question, pack purchase, grandfather status). Owner: support lead.
[ ] Legal: counsel reviews the terms language and the notice. Owner: founder.
[ ] Finance: the accountant reviews how packs are recognized. Owner: founder.
[ ] Send: the notice goes out on the decided date. Owner: founder.
[ ] Review: 30 days after go-live, a 20-minute call on churn, pack sales, and support volume. Owner: you (this is the $300 quarterly re-run’s first touch).
9. THE PROMPT PACK
Four prompts. Run each on redacted data only, with the founder’s written authorization.
Prompt 1, invoice line classification:
You are sorting AI vendor invoice lines for a pricing analysis. Input: a table of invoice or usage-dashboard lines with date, model, tokens in, tokens out, and cost. Task: group the lines by model and by day, total the cost, and flag any day whose cost is more than 2x the median day. Output: a table (model, days active, total cost, share of total) and a list of flagged days. Do not estimate anything that is not in the table; if a field is missing, say which.Prompt 2, pricing-page critique:
You are reviewing a software pricing page for usage risk, not for design. Input: the page text. Task: list every plan, its price, and every phrase that promises or implies unlimited or unmetered AI use. For each, state in one line what a heavy user could do under that phrase. Output: a table (plan, price, phrase, exposure). No recommendations, no rewriting.Prompt 3, edge-case generation:
You are stress-testing a usage allowance for a software plan. Input: the product's AI feature, the included outcomes per month, the expansion pack size and price, and the exception policy. Task: list 10 customer situations that could produce a complaint, a surprise charge, or an unfair result under these rules, one line each, most likely first. Output: the numbered list only.Prompt 4, notice compression:
You are editing a customer notice about a pricing change. Input: the draft. Task: cut it to under 150 words, keep every number exactly as written, keep the grandfather date, remove any sentence that apologizes or explains model economics in technical terms, and keep the tone plain and direct. Output: the edited notice only. Do not use em-dashes.10. THE FOUNDER SALES PACK
The 50-founder tracker, one sheet: company, vertical, product, plan price, the unlimited phrase quoted from the page, founder name, contact path, qualifier score (of 6), date contacted, reply, next step, outcome.
The outreach note (email; your real name, a real reply-to, a postal address in the signature, and a one-line “reply stop and I will not write again”):
Subject: The maximum model cost your [Plan] plan can absorb
Hi [first name],
Your [Plan] page promises [quoted phrase]. One question: do you know the maximum monthly model cost that plan can absorb before it misses your margin floor?
I run a seven-day reset that answers it from your last 30 days of invoices and hands back a pricebook with a floor under every plan: included use, one expansion pack, a grandfather plan, and a notice your customers will not hate. No code, and I never touch your billing system.
Here is a 90-second teardown of a public pricing page, with illustrative costs, so you can see the method: [link].
First reset is $750 for a founder who shares the data and lets me publish an anonymized before-and-after. Standard is $1,250 after that. Worth a 20-minute call?
[Your name]The teardown script (90 seconds, screen recording): 0-15s, the public pricing page with the unlimited phrase highlighted; 15-45s, the calculator with clearly labeled illustrative costs, the floor, and the included number appearing; 45-75s, the stress test row for one heavy user; 75-90s, the one-page pricebook and the sentence “this is the method; your numbers replace mine in seven days.”
Pricing and kill rules: $750 founding (once), $1,250 standard, $300 quarterly re-run. Never a free audit, never a savings share, never a per-customer percentage. If 50 qualified notes produce zero calls, change the vertical before changing the price. If a founder will not share invoices, they are not a buyer yet; leave the teardown and move on. Publish the before-and-after as one client’s numbers with the founder’s written consent, labeled as one result, never as typical.
THE WORKED EXAMPLE, END TO END
The Policy Summaries Reset. Fictional, for the math.
The buyer. A six-person company sells a $49-a-month tool to independent insurance agencies. The main AI feature summarizes policy documents into a one-page brief. The plan says “unlimited summaries.” 140 paying seats. The founder shares 30 days of OpenAI and OpenRouter invoices, the app’s completed-summary count, and the pricing page.
The cost map. 30 days of model spend on summaries: $1,470. Completed summaries: 5,880. Cost per completed summary: $0.25. Other variable costs: $4 a seat (hosting, document storage, payment fees).
The floor. The founder picks $34 of contribution per seat on the $49 plan, about 69% of price, and holds that share on every tier. Model budget: $49 minus $34 minus $4 = $11. Safe included summaries: 44. Published: 40.
The pack. 20 summaries, $5 to serve, $12 to buy, $7 of contribution per pack.
The stress test. Median seat: 30 summaries, $7.50 of model cost, contribution $37.50 today and unchanged after. Heavy seat: 200 summaries, $50 of model cost, contribution today $49 minus $50 minus $4 = a $5 loss. After the reset: $49 plus eight packs at $12 = $145, minus $50 minus $4 = $91 of contribution. Failure scenario (200 summaries plus 30% retries): $65 of model cost; today a $20 loss; after the reset, $76.
The pricebook. Starter $29 (15 summaries: a $20 floor, $4 of other costs, $5 of model budget), Pro $49 (40 summaries), Team $49 a seat with a five-seat minimum (200 summaries pooled; the pool is the team benefit, not a discount). One pack, $12 for 20, on every tier. Past the cap: a pause and a one-click pack, never a surprise charge. Nothing else in the product is counted.
The migration. Grandfather, 90 days. The notice goes out under the founder’s name on day 7 with the calendar date filled in. Eleven seats were over 40 summaries last month; the founder emails those eleven personally before the general notice.
The handoff. Eight lines, each with an owner and a date. The 30-day review is booked before you leave the call.
The money. The founder paid $750 (founding). Your costs: about $50. Twelve hours in, including the four hours of outreach that found this founder: $700 net, $58 an hour. The second reset in the same vertical reuses the cost map’s structure, the notice, and the teardown: it sells at $1,250 and takes about the same twelve hours, $1,200 net, $100 an hour.
THE 7-DAY DELIVERY SOP
Day 1: intake. 30-minute call. Confirm the qualifier, send the data request, agree the floor choice deadline, sign the engagement note (scope, privacy floor, the “not legal, accounting, or tax advice” line).
Day 2: data in. Run Prompt 1 on the redacted export. Build the cost map. Flag the days that spiked and ask what happened.
Day 3: the floor. Present the three floor options. The founder picks. Fill the calculator per plan.
Day 4: stress test and pricebook. Run the three scenarios. Draft the one-page pricebook. Run Prompt 3 for edge cases and fix the ones that would produce a complaint.
Day 5: migration. Decide grandfather, phased, or immediate. Draft the notice, run Prompt 4, prepare the objection lines.
Day 6: handoff. Build the checklist with owners and dates. Send the pack for the founder’s counsel and accountant to review.
Day 7: delivery call. 45 minutes. Walk the pricebook, the heavy-user sentence, the notice, the checklist. Book the 30-day review. Ask for the anonymized before-and-after permission in writing if it was part of the founding price.
THE WORKS-EVERY-TIME CHECKLIST
[ ] The unit is a customer outcome, never an API call.
[ ] Cost per outcome divides spend by COMPLETED outcomes, retries included.
[ ] The floor was chosen before anyone looked at customer behavior.
[ ] Every allowance was rounded down; every pack price is at least 2x its serving cost.
[ ] The heavy-user sentence is written in plain dollars and was said out loud.
[ ] The pricebook says what is NOT counted in its first line.
[ ] The exception policy is a pause plus a one-click pack, never a silent overage.
[ ] The migration decision is written down with its reason and a calendar date.
[ ] The notice is under 150 words, keeps every number, and names the grandfather date.
[ ] Every handoff line has an owner and a date; counsel and the accountant are on it.
[ ] No client data touched an AI tool without written authorization for that file.
[ ] The engagement note says you are a pricing analyst, not a lawyer, accountant, or data processor.
LEGAL NOTES
Outreach: business email under CAN-SPAM needs a truthful sender, a non-deceptive subject, a physical postal address, and a working opt-out honored within 10 business days; no automated texts or robocalls without prior consent (TCPA). Outside the US, a cold email to a founder in the EU or UK falls under GDPR and PECR: send only where a business-to-business legitimate-interest basis holds, and honor any objection at once. Advice boundary: you calculate and package; you do not draft enforceable billing terms, give accounting or tax treatment, or offer investment advice; the founder’s counsel and accountant review, and your engagement note says so. Data: work from aggregates and redacted exports; keep a written record of every authorization to process a file with an AI tool; delete working files on the agreed date. Customer notices: the founder sends them under their own name, and any promise in a notice (grandfather window, allowance, pack price) must match the billing system before it is sent. Price changes for existing subscribers may be governed by the founder’s own terms of service and by state auto-renewal notice laws; the counsel review line on the handoff checklist exists for exactly that.
THINKING THIS IS A LOT OF WORK?
It is. Read back what you just agreed to do.
Pick a vertical and study its pricing pages. Score fifty founders. Send twenty notes and ask a stranger for $750. Then, for the one who says yes: collect invoices, classify them, build a cost map, get a floor chosen, fill a calculator per plan, run a stress test, write a three-tier pricebook, generate the edge cases and fix them, decide a migration, draft a notice, prepare objection lines, and build a handoff checklist with eight owners. Seven days, one founder, and every artifact built from a blank sheet.
That is a solid week of evenings if it goes well. Two if it does not.
Or you open the Vault, where seven deeper builds are already finished and this one is being built with the cohort.
The Pricebook Lab is this issue’s deeper build, and it ships to the member shelf as it is built with the founding cohort: the multi-plan cohort workbook with the calculator and the stress test pre-wired, the sensitivity test across three model suppliers and two reasoning modes, the retry-cost scenarios, the price-change decision tree, the customer interview scripts for the eleven heavy seats, and the quarterly review procedure that turns one reset into a standing $300 client. Founding members vote on which piece deepens first.
And the shelf it lands on is not empty. Seven finished deeper builds are live in the Vault right now: the Spread Desk, the Pack Line, the Operator’s Edition, the Rep Desk, the Call Pack Studio, the Proof Room, and the Cleanroom, with the Sprint Foundry in progress beside them. Each one is the done-for-you version of a past week’s play, filled in and tested, so you start at the step that pays instead of the step that builds.
Weigh the price against the library, not this one kit: 49 plays and counting, a new play banking most weeks, seven finished builds already on the shelf, and the price never moves after you join.
$129 for the year, never goes up, for the first 25 members. After the 25th seat, the standard price starts climbing toward $399.
Out-yield the average.
Overyield is educational, not financial, legal, or business advice.