Blog

How to calculate AI cost per customer

AI cost per customer is the delivery cost required to serve one customer over a period, including the AI work that customer consumes. Calculating it well is the first step toward AI customer cost analysis and gross margin tracking.

· Arbiter Margin

Start with a clear definition

Decide the period (day, week, or month) and the customer identifier you already use commercially. Cost per customer is only useful if it reconciles to the same accounts that generate revenue.

Gather delivery-cost inputs

Include the costs required to deliver AI features for that customer — not only tokens.

  • Model / LLM usage
  • Tool and third-party API calls
  • Retries and failed runs allocated to the customer
  • Evaluation or guardrail passes tied to their traffic
  • Shared infrastructure allocated with an explicit, documented rule

Example calculation (synthetic)

Example only — not a customer result. Suppose Customer A generates $1,000 of plan revenue in a month. Attributed delivery costs are $420 model, $90 tools, $40 retries, and $50 allocated infrastructure.

Delivery cost = 420 + 90 + 40 + 50 = $600. Gross profit = 1000 − 600 = $400. Gross margin = 400 ÷ 1000 = 40%.

A second customer on the same $1,000 plan with $920 delivery cost yields 8% margin. Same revenue, different AI cost per customer.

Formula summary

AI cost per customer = Σ attributed delivery costs for that customer in the period.

Gross profit = revenue − AI cost per customer (for the same period and customer).

Gross margin = gross profit ÷ revenue.

Go deeper than the customer total

If margins look thin, split cost by workflow or task. Customer averages can hide a single unprofitable agent path. See AI cost-to-serve for the multi-level view, and AI cost per completed task for the task-level arithmetic.

FAQ

Should I include sales or support labor?
Only if your finance policy treats those as delivery cost for the analysis. Keep the rule explicit so comparisons stay consistent.
What if infrastructure is shared?
Use a documented allocation method. Perfect precision is less important than consistency and transparency.