Cost Attribution: Tracking LLM Spend by Feature and User

"Our AI bill went up 40%" is a fact you can't act on. "The summariser's output tokens tripled after Tuesday's prompt change, and 80% of it is one enterprise customer" is a fix. The distance between those two sentences is LLM cost attribution - tagging spend so every dollar has a feature, a model, and a user attached.

Tag at the call site

Attribution starts where the request is made. Every call carries a small envelope of metadata that travels with the recorded usage:

record({
  project: "prod-app",     // which app/environment
  feature: "summariser",   // which product surface
  user: currentUser.id,    // who triggered it
  model: res.model,        // which tier
  inputTokens: res.usage.input_tokens,
  outputTokens: res.usage.output_tokens,
  costUsd: price(res.model, res.usage),
});

The dimensions to roll up

• By feature - almost always concentrated. One or two surfaces are usually the majority of spend.
• By user - long-tailed. A handful of accounts drive a disproportionate share; find them before they find your invoice.
• By model - reveals silent upgrades and mis-routed traffic.
• By time - a step-change lined up with a deploy points straight at the cause.

What attribution unlocks

1. 01Targeted optimization - tune the 20% of features that are 80% of the bill, not everything.
2. 02Fair per-user caps - set hard caps from real usage instead of guesses.
3. 03Meaningful alerts - baseline per feature so a spike in a small surface isn't hidden by total volume. See budget alerts.
4. 04Unit economics - map spend to revenue per customer and know which accounts are underwater.

Attribution is the foundation the rest of cost control stands on - it's the first thing to add when you start monitoring Claude costs, and it's what makes a Slack cost alert specific enough to act on.