The Problem
Open your billing dashboard for last week. Aggregate AI spend is $4,800 against ~$60K of MRR on the AI feature — about 8% of revenue. Sustainable. You move on.
Now sort the same week by customer. The top user sent 4,127 messages to your AI agent. At your blended cost of $0.18 per message, that single account cost you $743 in compute. They pay $19/month. You are losing $724 on them this week, and they're a self-described “power user” who is probably tweeting about how much they love you.
Scroll further. The top 12 accounts are all underwater. Together they're consuming 38% of your token budget while contributing 0.4% of revenue. The aggregate spend dashboard, the one you check every morning, has been hiding this for months because the median user is profitable and the bulk users are roughly break-even. The tail is the bleed.
A flat subscription on top of variable compute is a margin time bomb. Aggregate dashboards are designed not to show you when it goes off.
This is the structural problem with customer-facing AI features priced on a flat subscription: token cost scales linearly with use, revenue does not, and the gap shows up on a per-customer basis long before it shows up in the company total. By the time the company total moves, you have thousands of users who learned that the feature is unlimited and a Substack post going around telling new signups how to use you as a free personal assistant.
The Core Insight
Token cost is a customer-level metric, not a system-level metric. The aggregate dashboard is a lagging indicator. The customer-tail dashboard is the leading one.
The metric that matters is cost_per_customer / revenue_per_customer on a rolling 30-day window, per customer, with a histogram. You don't ship customer-facing AI without that histogram. If you only know the mean, you don't know your gross margin — you know an average that hides whether half your accounts are losing money.
This reframes the entire conversation. “Is our AI feature profitable?” is the wrong question. The right questions are: what fraction of our customers are underwater? at what cost-to-revenue ratio? what is our policy when a customer crosses the margin floor? Those questions have answers; the aggregate one only has averages.
// the metric that matters
For every customer with at least one inference in the window:
- cost_30d — compute cost attributed to this customer over the last 30 days
- revenue_30d — the share of subscription revenue allocated to this feature (e.g. plan_price × feature_weight)
- margin_ratio = (revenue_30d − cost_30d) / revenue_30d — negative means you are losing money on this customer
Plot the histogram. The shape of the left tail is your business.