Some of your heaviest users cost you more than they pay.
Two reports, two weeks: margin per customer by usage decile, and the right size for your next vendor commit. Both with the arithmetic shown.
30 minutes. No deck, no obligation.
Built on a panel of 2,886 model SKUs over 35 months
Margin per customer, by usage decile
IllustrativeDollars per customer per month · every customer pays $200
Use
Margin
- 1st
- 2nd
- 3rd
- 4th
- 5th
- 6th
- 7th
- 8th
- 9th
- 10th
- +$196
- +$194
- +$191
- +$187
- +$181
- +$172
- +$156
- +$124
- −$40
- −$380
3,000 customers on one price, ranked lightest first. The heaviest tenth drives 57% of the tokens and pays the same $200 as the lightest.
Illustrative figures — a worked example, not customer data.
Five numbers, and where each came from.
On flat-priced AI products, the top 10% of users drive more than 70% of tokens (Kyle Poyar, Growth Unhinged).
AI-product gross margins run 45–53% (ICONIQ, N=305), so the heaviest accounts decide where in that band you land.
75% of AI companies changed their pricing within a year (Growth Unhinged, N=230+).
“Seven minimum spend commitments… wasted a couple million” — a Series C finance head, on the record via Ramp’s blog, April 2026.
Anthropic and OpenAI credits lapse at about 12 months, and no tool tracks the pools (vendor terms, verified 2026-08-14).
What you get in week two
Five screens, built from your vendor exports and your contract terms.
Every figure is a worked example, not customer data, and the arithmetic ties.
01The decile table
01 · The decile table
IllustrativeEight deciles pay for themselves. Two do not.
Every customer ranked by usage, with what each one costs to serve and what is left over.
Decile | Cost to serve | Margin | Margin % |
|---|---|---|---|
| 1st | $4 | +$196 | 98% |
| 2nd | $6 | +$194 | 97% |
| 3rd | $9 | +$191 | 96% |
| 4th | $13 | +$187 | 94% |
| 5th | $19 | +$181 | 91% |
| 6th | $28 | +$172 | 86% |
| 7th | $44 | +$156 | 78% |
| 8th | $76 | +$124 | 62% |
| 9th | $240 | −$40 | −20% |
| 10th | $580 | −$380 | −190% |
Blended gross margin 49%. Served at a loss: the 9th and 10th deciles, 600 of 3,000 accounts.
Illustrative figures — a worked example, not customer data.
02The repricing simulation
02 · The repricing simulation
IllustrativeMove the loss-making deciles to a usage tier at cost plus 25%. Blended margin 49% to 61%.
One move, priced off the same table, with the accounts it touches counted.
It touches 600 of 3,000 accounts. 41 of them renew inside 90 days.
Illustrative figures — a worked example, not customer data.
03The commit position
03 · The commit position
IllustrativeEvery commitment, credit pool and seat plan, with the date it expires.
One page for what you have already bought, ordered by how soon it stops being yours.
- Vendor A — annual commitexpires 31 Mar 2027
- $2.40M
- Vendor B — prepaid creditsexpires 14 Nov 2026
- $400k
- Vendor C — seat plan, 120 seatsexpires 1 Jan 2027
- $288k
At $45k a month with 3 months left, $265k of the credit pool lapses unused.
Illustrative figures — a worked example, not customer data.
04The two exposures
04 · The two exposures
IllustrativeStranded $0.30M if the year lands low. Runaway $0.40M if it lands high.
Two ways the same commitment costs money, priced side by side before anyone signs it.
Both are priced off the same forecast, so the choice between them is a number rather than a preference.
Illustrative figures — a worked example, not customer data.
05The signed number
05 · The signed number
IllustrativeCommit $3.20M at the annual tier. Leave $0.50M on demand.
The deliverable, with the assumptions it rests on set out where you can argue with them.
Committed
$3.20MOn demand
$0.50M- Accountsyour FY27 plan
- 3,000 → 3,540
- Tokens per accountyour last six months
- flat
- Model priceunerr price panel
- unchanged
- Repricingyour decision
- not applied
$3.20M + $0.50M = $3.70M, the midpoint of the forecast band above.
Illustrative figures — a worked example, not customer data.
Why nobody hands you this today
Cost-per-customer tools exist. None of them joins your revenue system, so none can tell you margin per customer. And nobody derives commit advice from your observed usage — every commitment optimiser on the market covers cloud instances only.
The reason is structural. A bill shows price and work delivered multiplied into one number, never the two apart. Deciding what to buy takes a measure of the work.
The levers work without the meter. The business doesn’t.
How it works
You send
Vendor usage exports, gateway logs and your contract terms
Hours of your team’s time, not a project.
We work
Two weeks, fixed scope
You get
Two reports with money attached, and every assumption editable.
30 minutes. No deck, no obligation.
Fixed scope, two weeks, two reports.
$5,000–$15,000
unerr, engagement pricing
Scope
- Fixed. Two reports, not an open-ended retainer.
Timeline
- Two weeks, start to both reports.
What we need
- Vendor usage exports, gateway logs and your contract terms.
Questions we get before the call
Answered here so they don’t slow down the call.
We already see our token spend — Anthropic’s console, Ramp, our gateway.
Good — clean history makes this faster. Those show spend: single-vendor, list-price, or cost-only. None of them joins your revenue system, so none can tell you margin per customer, and none sizes a commit against your own observed usage. That is a different question.
Our engineers already optimise.
They should — the levers work. What the levers cannot do alone is target a budget, attribute the saving, and defend it to whoever signs the next commit. Optimisation lowers a bill. It does not tell you which customers you serve at a loss, or how much to buy next year. The levers work without the meter; the business doesn’t.
What do you need from us?
Vendor usage exports at model granularity, your gateway logs, and your contract terms. That is hours of your team’s time, not a project. We install nothing and run nothing against your production systems. If your billing history is already clean, the first report lands sooner.
What if your number is wrong?
It is a range with a named assumption register, not a point estimate. Every driver carries the assumption it rests on and where that assumption came from. Disagree with a driver and the model updates — that is the product, not a caveat.
Book 30 minutes.
Two reports, two weeks. The arithmetic is yours to keep.
Loading calendar…