Cost tracking, attribution and budgets

0:00 Cost tracking and budgets

At the end of the month, the invoice arrives and it's forty percent higher than last month. Which feature caused it? Which customer? Which change? If you can't answer in five minutes, this lesson is for you. You'll learn how to track AI costs per call, attribute them to features and customers, set budgets that actually protect you, and talk about AI spend in business terms.

0:29 Why it matters

Why does this matter? AI costs behave differently from most cloud costs. They scale with usage and prompt design, they can jump overnight when a feature goes viral or an agent loops, and prices vary hugely between models. Think of a household electricity bill. The total tells you what you spent, but a smart meter per appliance tells you it was the old freezer in the garage. Your per call ledger is that smart meter.

1:02 Usage per provider

Every provider returns usage with each response. Claude reports input and output tokens, plus cache reads and cache writes. OpenAI's Responses API reports input and output tokens, cached tokens, and reasoning tokens. Gemini reports prompt tokens, output tokens, cached tokens and thinking tokens. Here's the catch: reasoning tokens are billed as output on providers that expose them, so a thoughtful model can bill far more tokens than the visible answer suggests.

1:33 The cost ledger

So build a ledger. One row per model call, with the timestamp, provider, model, feature, tenant, a pseudonymized user id, environment, every token type, whether it was a batch request, latency, outcome and estimated cost. Calculate cost from prices loaded from configuration with an effective date, never hard coded. And once a month, reconcile your estimates against the provider's usage and billing reports, so you know your numbers are right.

2:03 Simple example: three features

A simple example. A solo developer runs a writing assistant with three features: rewrite, summarize and translate. After adding a ledger, she sees translation is a third of her requests but two thirds of her cost, because users paste whole documents. She adds an input size limit with a friendly message and moves long documents to a cheaper model. Her bill drops, and nobody complains.

2:31 Hands-on: cost meter

The lesson's code gives you a provider agnostic cost meter. Small adapter functions convert each provider's usage object into one common shape: input, cached input, cache writes and output. A cost function applies prices from your config, with a batch multiplier where relevant. And a record function writes a row with feature and tenant tags. Check each provider's token accounting, like whether cached tokens are included in the input total, and adjust the adapters if needed.

3:04 Budgets that protect

Now budgets that protect you. Per request caps, like max tokens, input size limits and agent step caps. Per user and per tenant quotas enforced in your backend. Per feature budgets with alerts at fifty, eighty and a hundred percent. Provider side spend limits and separate projects. And anomaly alerts when cost per request or per tenant jumps compared with the recent average. One customer's runaway loop shouldn't become everyone's problem.

3:35 Per-tenant limits

Here's a practical detail that saves real money: per tenant limits. If you sell an AI feature to many customers, one customer's integration can accidentally call you in a loop overnight. Without limits, their mistake becomes your invoice. Give every tenant a daily token or cost allowance, enforced in your backend before calling the provider. When they hit it, return a clear message and alert your team. Most customers will never notice the limit; the one who hits it will thank you for catching their bug.

4:12 Business example: agency billing

Now a realistic business example, with illustrative numbers. An agency working across the UK and Pakistan runs AI workflows for twenty five clients. With the ledger tagged by client and feature, finance produces a monthly AI cost report per client, adds a margin, and invoices it. The report also revealed that one client's weekly report feature cost about ten times more than similar clients, because it attached a sixty page PDF every run. Caching and page selection fixed it.

4:46 Unit economics

Finally, speak the language of the business: unit economics. Translate tokens into cost per support conversation resolved, per lead enriched, per product description published, or per active user per month. Then compare that with the value: time saved, revenue, conversion. This is how you decide where a bigger model is worth it, where caching or batch deserve engineering time, and how to price AI features for your own customers.

5:16 Common mistakes

Common mistakes. Relying only on the provider invoice. Hard coded prices that silently go stale. Ignoring reasoning tokens and cache writes. And no per tenant limits, so one customer's loop becomes everyone's bill. Each of these is cheap to prevent today and painful to untangle after a surprising invoice arrives.

5:38 Showing costs

A common question: should you show AI costs to your customers or users? For internal teams, yes: showing each team its monthly AI spend changes behavior quickly. For external customers, it depends on your pricing model. Many products bundle AI into plan tiers with fair use limits, while agencies often pass through costs with a margin. Either way, your ledger gives you the numbers to decide pricing with confidence rather than guesswork.

6:09 Forecast monthly

One last habit: forecast. Take the last four weeks of ledger data per feature, work out cost per outcome and outcomes per week, and project forward with known changes, like a new client, a marketing campaign or a model switch. Add a buffer for growth and price changes, share it with finance, and compare forecast with actual every month. The gap tells you which assumption was wrong, and over time your forecasts become boringly accurate, which is exactly what finance wants.

6:44 Deeper: agency AI billing (illustrative)

Let's deepen the UK and Pakistan agency's billing example with illustrative numbers. Twenty five clients, three AI features each. The ledger showed one client's weekly report costing about ten times the median, caused by a sixty page PDF attached to every run. After caching the document and selecting only the relevant pages, that client's cost dropped in line with the others. Finance then introduced a simple pricing rule: AI usage passed through at cost plus a fixed margin, shown as a line on each monthly invoice. Clients appreciated the transparency, and the agency stopped absorbing AI costs it couldn't see before.

7:28 Watch me do it: provider-agnostic cost meter

Watch me do it. Let's walk through the cost meter. Prices load from a JSON file with an effective date, never hard coded. The usage data class has provider, model, input, cached input, cache writes and output. Three adapter functions convert each provider's usage object. From Claude reads input tokens, cache read and cache creation tokens and output tokens. From OpenAI subtracts cached tokens from input so they aren't double counted. From Gemini subtracts cached tokens from the prompt count and adds thinking tokens to output, because they're billed as output. The cost function multiplies each part by its price per million, with a batch multiplier when relevant. Record builds a row with the feature, tenant, batch flag, cost and timestamp, and prints it, standing in for your data warehouse. I record one call from each provider for the feature product copy and tenant noor, and three comparable rows appear.

8:33 Recap + try this now

Quick recap. Invoices show totals, so build your own per call ledger with feature and tenant tags, prices from config, and monthly reconciliation. Enforce budgets at every level, and express spend as cost per business outcome. Try this now: add a cost ledger to one application, set one budget alert, and calculate the cost per outcome for one feature, like cost per resolved support conversation.