From "we should look into this" to a budget that holds — one month, four weeks, no heroics.
Everything before this chapter was method. This one is a calendar. It assumes nothing exotic — a spreadsheet, access to your vendor dashboards and invoices, and a few hours a week from whoever owns the question. Each week ends with something that exists and didn't before.
You cannot attribute what you haven't found, and most organizations are spending with more vendors than anyone can name from memory. So week one is a hunt, and its output is one spreadsheet.
One row per AI vendor or product, and be expansive about what counts: model providers, obviously — but also per-seat subscriptions (coding assistants, chat seats, embedded copilots), voice and speech services, transcription, telephony and SMS lines behind conversational products, GPU or hosting spend for anything self-hosted, and the AI add-ons quietly attached to tools you already pay for.
Three columns per row: what it bills (metered usage, per seat, flat), roughly what it cost last month, and — the load-bearing one — what it exposes: a usage API, request tags or per-customer keys, or invoice-only. That third column is next week's to-do list, pre-sorted.
Expect surprises. The forgotten lines from chapter 4 — telephony, the second and third models in a pipeline, seats nobody remembers provisioning — tend to surface here, and finding them is the point.
Two moves this week, one decision and one mechanical.
The decision: pick your unit, per chapter 2. If you sell AI, it's the thing your pricing is written in — the call, the render, the document, the seat. If you buy AI, it's the thing your budget is written in — the team, the project. Write it down and use it in every conversation from here on; half the value of a unit is that everyone means the same thing by "cost."
The mechanical move: for every vendor whose week-1 row says it supports tags or keys, turn them on. Request tags carrying customer or team; separate keys per customer, team, or environment where tags aren't supported; workspaces per team on subscription products that offer them. This is the "easy case" from chapter 4 — free attribution at the source — and this week is when it stops being left on the table. New spend starts arriving pre-attributed immediately; the earlier you switch it on, the more of the month you can actually see.
Now the spend that can't be tagged: pull last month's invoice for every invoice-only row from week 1 and bring it into the same ledger as the tagged spend, in the same unit.
Allocate each one with a simple, stated rule — subscriptions divided across active seats by team, shared infrastructure split usage-proportionally by your unit, telephony assigned to the product line that owns the phone number. Chapter 4 has the mechanics; the discipline this week is to write each rule down next to its number and mark every allocated figure estimated.
Resist the urge to perfect. Coverage before precision: by Friday, every row from week 1 should have an owner and a number, even if some numbers are rough. A complete rough ledger beats a precise partial one, because the rough ledger can answer questions this month.
Three closing moves.
First reconciliation. Sum your ledger per vendor and compare against what each vendor actually charged. Gaps are normal on a first pass — chase the big ones, shrug at the small ones, and note what the chase teaches you about your allocation rules. This is the monthly habit that keeps the whole system honest, starting now.
First blocking budget. Pick the scariest dimension from your ledger — the runaway-capable key, the team whose spend doubled, the frontier model everyone defaults to — and put a real, blocking limit on it, generous enough that nobody should hit it (chapter 6). The point of the first block is existence: from this week on, at least one class of surprise arrives as a denied request instead of an invoice.
The Monday number. Start a weekly one-email ritual: your unit, its cost, its trend, and anything anomalous — five lines, same time every week, to whoever owns the outcome. Unit economics dies in dashboards nobody opens; it lives in a number people expect on Monday.
A complete vendor inventory. A named unit. New spend tagged at the source, stubborn spend allocated by written rules, all of it in one ledger, reconciled against invoices once. One budget that can actually say no, and a weekly number that keeps the whole thing alive.
That's the method — and everything in this handbook can be run by hand, exactly as described.
It's also, we should say plainly, the work we've automated. UnitSense does this for a living: attribution across providers, subscriptions, and invoices; per-unit and per-customer cost; budgets enforced where requests happen. If you'd rather have day 30 without spending the month, we'll show you what it looks like on your own spend — unitsense.ai, or book a call at unitsense.ai/book.
Either way: pick your unit this week. The window is open, and it's shorter than the last one.
Start over: All chapters →