The unit ladder

Ask a vendor what AI costs and you'll get an answer in tokens. Ask a board and the question comes back in customers and margin. Between those two lies every AI cost problem you will ever have — and a surprising number of them are just the same mistake: answering a question from one altitude with data from another.

It helps to draw the altitudes as a ladder:

token → request → session → call / feature → customer → P&L

Vendors bill at the bottom. Businesses decide at the top. Unit economics is the climb.

What each rung can — and cannot — tell you

Tokens. The atom. Every provider reports them; every price sheet is denominated in them. A rung-1 view answers "what is our blended cost per million tokens?" — and nothing else. No token has a customer.

Requests. One API call: a model, an input size, an output size, a price. This is the first rung where attribution is possible — a request can carry tags (customer, feature, team, environment) if you put them there. It's also the last rung most tooling reaches. Requests answer "which model and which endpoint cost what?"

Sessions. Requests joined by shared context — a conversation, an agent run, a working session. This is where the long-call trap lives (chapter 3): cost per session is not the sum of independent requests, because context accumulates. Sessions answer "what does one use of the product cost?" — the first rung with any business meaning.

Calls and features. Sessions rolled up to the thing you actually ship: a support call, a rendered video, a document review, a completed agent task. This is the rung your product is priced in, whether you admit it or not. It answers "what does the unit we sell cost to deliver?"

Customers. Everything above, summed per paying account and set against what they pay. The rung answers the question that decides pricing: "which customers make us money, and which quietly don't?" If you sell AI, this is your load-bearing rung.

P&L. Customers rolled up to the company: gross margin if you sell AI, spend versus budget by team if you buy it. The rung finance defends in a board meeting — and the only rung a CFO natively sees, which is precisely the problem: it arrives weeks late and explains nothing below it.

The climb is a series of joins

Nothing moves up the ladder for free. Each rung is a join someone has to build:

• Request → session needs a session identifier carried on every request.
• Session → call/feature needs the session mapped to the product event it served.
• Anything → customer needs a customer tag at request time, a key-per-customer scheme, or an allocation rule (chapter 4).
• Customer → P&L needs revenue joined in — billing data next to cost data, the join almost nobody has.

Miss a join and every rung above it goes dark. This is why most companies can quote their token spend to four decimal places and cannot say what their largest customer costs them: the data exists at rung 1 and the question lives at rung 5, and nobody built the stairs.

Wrong-rung errors

Two failure patterns account for most of the confident nonsense in AI cost conversations.

Answering up-ladder questions with down-ladder data. "Which customers lose us money?" answered with a blended cost per token is arithmetic theater — an average smeared across accounts whose usage differs by 100×. The number is precise, defensible, and useless. Same for "what should we charge?": a rung-1 answer to a rung-5 question prices your whales at your median.

Budgeting up-ladder with down-ladder controls. The reverse error. A company sets a P&L-level target — "AI spend under $80k this month" — with tooling that only observes tokens after the fact. The target has no enforcement rung to live on, so it's a hope, not a budget. Controls work at the rung where requests happen (chapter 6); targets set five rungs above their instrumentation are how month-end surprises are manufactured.

Choose your unit

The practical move that organizes everything else in this handbook: pick the rung you operate at, and name your unit.

• If you sell AI, your unit is the thing your pricing is written in — the call, the minute of product, the seat, the document, the render. Your work is joining cost up to that unit and to the customer above it.
• If you buy AI, your unit is the thing your budget is written in — the team, the project, the person. Your work is joining spend down from the invoice to those owners.

One test tells you whether the choice matters urgently: does user behavior move your AI cost? If usage is fixed and deterministic — one model call per record, same size every time — rung-1 math roughly holds and simple arithmetic will serve. If usage is open-ended — conversations, agents, anything a user can lean on — the rungs diverge, the averages lie, and the climb is where your margin went.

The rest of this handbook is the climb: what makes sessions expensive (chapter 3), how to attribute the stubborn costs (chapter 4), how to price at your unit (chapter 5), and how to make a budget hold at the rung where spend actually happens (chapter 6).