Use case

What does AI cost you to serve one customer?

If you sell an AI-backed product, this is the number that decides whether the business works. AI scales with usage, which makes it cost of goods sold, not a technology line item.

How do you calculate AI cost per customer?

By attributing model spend to the workflows that serve customers, then to the accounts consuming those workflows. Where requests carry a customer identifier the figure is measured directly. Where they do not, Camaze apportions using a usage driver you already have, such as seats or API calls, and labels it as apportioned rather than measured. The result is AI cost per account, per segment and per plan, reported against the revenue from each.

The problem

You can grow revenue into negative margin and not notice for a year.

Traditional software has near-zero marginal cost, which is why unit economics discipline atrophied in SaaS.

More detail

AI reverses that. Every customer action that touches a model costs real money, and heavy users cost multiples of light ones on the same contract.

Without per-customer attribution, that is invisible. The AI line grows, revenue grows, and the two look correlated. The fact that your five largest accounts are consuming eleven times the AI of an average account on the same pricing does not appear anywhere until gross margin moves.

  • AI reported as a technology cost rather than as cost of goods sold
  • No view of cost per account, so no view of marginal contribution
  • Unlimited-usage pricing set before anyone knew the unit cost
  • Enterprise accounts negotiated on seats while consuming on tokens
  • Feature-level costs unknown, so packaging decisions are guesses
AI cost to serve By segment, monthly
SegmentRevenue per account AI cost per accountAI as share of revenue
Enterprise, tier 1$18,400$2,180 11.8%
Enterprise, tier 2$9,200$1,640 17.8%
Mid-market$3,100$291 9.4%
Growth$980$142 14.5%
Self-serve$79$31 39.2%

Illustrative product view. Figures are examples.

What we do about it

Put AI in gross margin, where it belongs.

The mechanics are attribution down two levels: model spend to customer-serving workflows, then those workflows to the accounts consuming them.

  • AI cost per account, per segment and per pricing plan
  • Reported against revenue, as a share and as a margin component
  • Customer-serving workloads separated from internal and non-production use
  • Cost per feature, for deciding what is included and what is metered
  • Accounts whose AI cost is growing faster than their contract value, flagged
Unit economics

Cost to serve, by segment.

The segment view is usually where the first uncomfortable fact appears.

More detail

Self-serve customers frequently carry the worst AI-to-revenue ratio, because pricing was set when marginal cost was assumed to be near zero and usage is unconstrained.

Enterprise accounts tend to look healthier in aggregate and hide the widest variance, which is why the account-level view matters as much as the segment one.

  • AI cost per account and per segment, next to revenue
  • AI as a share of revenue and as a component of gross margin
  • Distribution within a segment, not only the average
  • Trend per account, so drift is visible before renewal
AI cost to serve By segment, monthly
SegmentRevenue per account AI cost per accountAI as share of revenue
Enterprise, tier 1$18,400$2,180 11.8%
Enterprise, tier 2$9,200$1,640 17.8%
Mid-market$3,100$291 9.4%
Growth$980$142 14.5%
Self-serve$79$31 39.2%

Illustrative product view. Figures are examples.

Outliers

The accounts eating their own margin.

Within any segment, AI consumption varies enormously.

More detail

A handful of accounts routinely account for a disproportionate share, usually because of one integration or one automated workflow calling your product in a loop.

Finding them early gives you options that disappear later: a rate limit, a packaging change, a conversation at renewal, or a genuine product fix. Finding them at renewal gives you one option, and it is uncomfortable.

  • Accounts ranked by AI cost against contract value
  • Consumption growth per account, trended
  • Alerts when an account crosses a cost-to-serve threshold
  • The specific workflow driving an outlier account
Agent and workflow economics Last 30 days
WorkflowRuns Calls per runCost per run Spent on retries
Ticket resolution agent 18,400 3.2 $0.41 34%
Contract review workflow 2,100 11.8 $2.90 71%
Lead enrichment agent 44,900 1.4 $0.06 12%
Code review agent 6,300 6.1 $0.88 48%
Weekly report generator 310 22.4 $4.15 9%

Illustrative product view. Figures are examples.

Pricing

Set prices on evidence.

Cost per feature is what makes packaging decisions defensible.

More detail

Which AI capabilities belong in the base tier, which are premium, which need metering, and what an unlimited-usage promise actually exposes you to at the top of the distribution.

Scenarios let you model a pricing or packaging change against actual consumption before committing to it, rather than discovering the effect a quarter later.

  • Cost per feature, across the accounts using it
  • Exposure modeling for unlimited-usage tiers, at the tail
  • Packaging scenarios run against real consumption
  • Margin effect of a pricing change, per segment
Scenario planning Monthly run rate
Five quarter AI spend forecast under three scenarios On the current trajectory monthly spend reaches about 447 thousand dollars. Rolling the feature out to everyone reaches about 662 thousand. Following the optimization plan holds it at about 238 thousand. 0 $175k $350k $525k $700k $447k $662k $238k NowQ1Q2Q3Q4Q5
Current trajectoryFeature rolled out to everyoneWith the optimization plan

Illustrative product view. Figures are examples.

Find out what a customer costs.

Bring your segments and we will show you AI cost to serve against revenue, on your own data.

In practice

Getting to a per-customer number

The precision available depends on what your requests already carry.

1

Measured

Requests carry a customer or account identifier. Cost per customer is calculated directly from usage. This is the most precise, and adding the identifier is usually a small engineering change.

2

Apportioned

No customer identifier, but customer-serving workflows are identified. Cost is apportioned using a driver you already have, such as seats, sessions or API calls, and labeled as apportioned rather than measured.

3

Segment level

Where neither is available, cost is reported at segment and plan level from workflow attribution. Less precise, still enough to show that self-serve is carrying a materially worse ratio than enterprise.

FAQ

Questions people ask

We do not tag requests with a customer identifier. Can you still do this?

Yes, at lower precision. Camaze attributes to customer-serving workflows and then apportions to accounts using a usage driver you already have. The figure is labeled as apportioned so nobody mistakes it for measured. Adding an identifier upgrades it, and it is typically a small change worth making once you have seen the apportioned version.

Should AI cost go in cost of goods sold?

For AI that serves your product to customers, that is the usual treatment, because it is a variable cost that scales with usage and it affects gross margin directly. AI used internally for productivity is operating expense. Camaze separates the two so the classification is evidenced. Your auditors will have a view on specifics, and the supporting detail is available.

How do we handle shared AI costs that are not per customer?

Keep them out of the per-customer figure and report them separately as platform cost. Mixing fixed platform overhead into a marginal cost calculation produces a number that is wrong in both directions depending on volume.

Can this feed our existing unit economics reporting?

Yes. Per-account and per-segment AI cost exports to Excel, Google Sheets, Snowflake or over the API, so it can join whatever margin reporting you already run rather than living separately.

Know your cost to serve.

Book 30 minutes. We will show you AI cost per account and per segment against revenue.

5 minute setup. Read-only. No engineering time needed.