Prumo by Luciana Reynaud ↗
LLM FinOps · spreadsheet kit

Every AI invoice, attributed to the person who spent it.

The LLM Cost Attribution Kit splits your Claude, OpenAI, Copilot and Cursor invoices per user and per team, reconciles the totals to the cent, and reports cost per unit of work — in one validated spreadsheet. No platform. No scripts. No API keys.

Buy the kit — $79
Instant download · xlsx + 8-page playbook
Excel · LibreOffice · Google Sheets
invoiced $8,460.00 attributed $8,460.00 difference $0.00 33 checks · status OK
// The problem

Your AI invoices arrive as monoliths.

One number from Anthropic. One from OpenAI. One from GitHub. One from Cursor. Finance asks the only question finance ever asks — who is spending this and what do we get back? — and nobody can answer, because vendors bill at the account level, every billing model is different, and two of them changed their pricing mechanics this year alone.

The result is familiar: AI budget grows quarter over quarter, nobody owns it, and the first cost review threatens every tool indiscriminately — including the ones quietly paying for themselves.

// What's inside

A spreadsheet that refuses to look trustworthy until it is.

The workbook

8 tabs, fully formula-driven, preallocated for 200 users × 20 teams × 8 cost components. Seat, usage, and hybrid billing. A 33-check data quality battery gates every number behind an OK status.

The playbook

8 pages: the three attribution models, the allocation-basis rule that keeps your split economically valid, a vendor billing guide current to August 2026, a 30-day rollout plan, and a copy-paste email to Finance.

Markdown copy

The full playbook in plain text — for your internal wiki, your editor, or the part of your team that reads documentation in vim.

// How it works

One formula. Three billing models. Zero mystery.

user's allocation basis ÷ component's total basis × component invoice
  1. 01Classify each invoice line as a cost component: seat-based (allocated by roster), usage-based (allocated by billed cost or uniformly priced credits — never raw tokens across mixed model prices), or hybrid — two components, the model GitHub Copilot moved to in June 2026.
  2. 02Fill three input tabs — components, teams, users — from data you already have: invoices, rosters, and the vendor's own usage export. The Checks tab flags anything inconsistent before it can mislead anyone.
  3. 03Read the results: per-user attribution, team rollup, and CPUW per team — reconciled against invoice totals to the cent, screenshot-ready for your first monthly report.
// The metric

What is CPUW?

CPUW (Cost Per Unit of Work) is a metric that expresses AI spend as the cost of one completed unit of business work — a resolved ticket, a merged pull request, a delivered report — rather than raw tokens or invoice totals. It connects LLM billing data to business output, so AI cost becomes comparable, budgetable, and attributable per team.

CPUW = team AI cost ÷ units of work delivered in the same period

Why it works: tokens are not comparable across teams or months; units of work are. "$0.83 per resolved ticket" survives a budget meeting. "41 million tokens" does not. CPUW is comparable within a team over time and, cautiously, across teams that declare the exact same unit — which is why the kit computes it per team only, and shows N/A rather than zero when a team has cost but no declared output.

What CPUW is not: an ROI claim. It states what a unit costs in AI spend; it does not prove AI caused the output. That honesty is deliberate — credibility, once spent, doesn't come back.

How CPUW relates to "cost per task": benchmarks measure the cost for a model to complete a test task — useful for choosing models. CPUW is the operational sibling: the cost of a completed unit of business work, with retries, fallbacks, orchestration and human review included, measured from your own invoices. Benchmarks compare models; CPUW governs spend.

// Questions

Asked before buying.

Does it work in Google Sheets?
Yes — import verified: all formulas, dropdowns, and conditional formatting survive. One documented caveat: Sheets drops formula-cell protection; the README explains how to re-protect ranges in two clicks.
My vendor isn't on the list.
Component names are free text and headers propagate automatically. Any vendor fits one of the three billing models; the playbook shows how to classify it with a single question: does this invoice line change when usage changes?
More than 200 users?
That's the designed capacity. Beyond it you've outgrown spreadsheets — and the method (components, allocation basis, checks, reconciliation) transfers unchanged to a pipeline. The method is the part you're buying.
Is this financial reporting?
No. The kit produces management estimates from declared premises — invoices, rosters, unit counts — clearly labeled as such throughout. Validate premises with their owners before official use.
Support and updates?
Self-service product, sold as-is, with honest scope stated in the documentation. Defect reports are welcome and fixed; buyer-specific modeling support is out of scope by design — that's what keeps it $79.
// Get it

One monthly cycle pays for it.

If you administer seats on two or more AI tools, the first reconciled report answers questions your leadership has been asking for months.

Buy the kit — $79
Instant download on Gumroad
Fictional example data included — see the format before you fill it