Solutions

Measure AI cost-to-serve

AI cost-to-serve is the delivery cost of completing work for a customer. It is wider than a token invoice, and it is the input to AI customer profitability — not a substitute for billing or accounting.

Connect what you earn with what it costs to deliver AI — at the customer, task, and workflow level.

What belongs in AI cost-to-serve

A completed AI task often fans out across several cost lines. AI delivery costs are the sum of the lines your finance policy treats as the cost of serving that work.

LLM cost attribution by customer only becomes useful when those lines can be tied back to the account and the workflow that caused them.

  • Model calls and token consumption, including longer context and higher-tier models
  • Tool calls and external APIs
  • Retries after failed or low-quality runs
  • Evaluations and guardrail passes
  • Caching: cache hits can reduce model calls, while cache storage and refresh still cost something
  • OCR and document processing
  • Vector search and retrieval
  • Infrastructure allocated with an explicit, documented rule

AI cost per customer and AI cost per workflow

AI cost per customer is the delivery cost attributed to one account over a period. AI cost per workflow splits that total by the kind of work — a short answer, a tool-heavy support path, a research agent, a document pipeline.

Customer averages can look acceptable while one workflow is expensive. AI workflow cost analysis is how you see that split. AI cost per task sits one level lower: the delivery cost of a single unit of work, including the retries it took to finish.

How to calculate AI cost-to-serve

Pick a period and the same customer identifier you use for revenue. Sum the delivery-cost lines attributed to that customer. That sum is AI cost per customer for the period.

To see AI cost per workflow, group the same lines by workflow before you sum. Shared infrastructure stays out, or goes in, according to one written allocation rule — consistency matters more than false precision.

Illustrative only, not a customer result. A workflow attributed with $0.80 of model cost, $0.25 of tool cost, and $0.45 of retries and evaluations has $1.50 of delivery cost before any shared infrastructure you choose to allocate.

Why this is not an observability dashboard

Provider consoles and monitoring tools remain the right place to inspect latency, errors, and raw spend. Billing remains the commercial system of record.

Cost-to-serve is the commercial reading of that spend: what it cost to deliver the work a customer paid for. AI customer profitability is the comparison of that cost with revenue.

Continue with how AI agent pricing changes when delivery cost varies.

FAQ

Is AI cost-to-serve the same as cloud spend monitoring?
No. Monitoring explains infrastructure spend. Cost-to-serve attributes delivery cost to customers, tasks, and workflows so it can be compared with revenue.
Do I need to replace my billing system?
No. Arbiter Margin sits beside your billing stack. It does not issue invoices or replace your subscription system.
Should caching, OCR, and vector search be included?
Include them when they are required to complete the customer’s work. A cache hit is not automatically free if you still pay for storage, refresh, or retrieval.

From the blog