Your AI costs are only going higher. It is time you knew where.
The fastest growing line in the budget is the one with the least evidence behind it. Get the total, the attribution, the forecast and the savings, in a form you can defend.
What does a CFO get from an AI cost platform?
Four things a CFO cannot currently produce: a complete AI total across every provider, tool and self-hosted workload; that total attributed to departments, products and customers; a forecast built on usage drivers with a defensible range; and a quantified list of cost savings. Together they turn AI from an unexplained line into a managed cost with an owner, a plan and a variance you can decompose in a sentence.
You are being asked to approve a number nobody can explain.
AI is now material, it is growing faster than anything else in the budget, and it is the line with the weakest evidence behind it.
More detail
When the board asks what it costs and what it returns, the honest answer is usually an approximation of the first half and silence on the second.
That is an uncomfortable position, and it has a predictable outcome. A cost that cannot be explained, attributed or forecast gets capped, because capping it is the only responsible action available. The cap then lands on the AI work that was going to pay off, because nobody can distinguish it from the AI work that was not.
- A total assembled by hand from invoices, card statements and cloud bills
- No attribution, so no owner and no accountability
- Variance discovered at close, with no explanation available
- A forecast that is last month's figure with a growth percentage applied
- No return figure, so the line cannot be defended on its merits
Illustrative product view. Figures are examples.
Make it a managed cost.
Nothing here is exotic. It is the standard treatment every other significant cost in the business already receives, applied to a category that grew too quickly for anyone to apply it.
- One complete total, across every provider, tool and self-hosted workload
- Attributed to departments, products and customers, with showback and chargeback
- Budgets with owners, and notice on day 5 rather than day 31
- A driver-based forecast with a range, and variance decomposed into volume, unit cost and scope
- A quantified list of cost savings, tracked from identified to realized
Answer the question in one sentence.
The value of attribution is not the chart.
More detail
It is being able to say, without preparation, that AI cost 1.4 million last quarter, that 62% of it sits in two production workflows in Product and Support, that unit cost fell 18% after a model change, and that the overrun is one workflow that shipped in month two.
That sentence is what separates a cost that gets approved from a cost that gets capped.
- Spend by department, product and customer, from one source
- Variance split into volume, unit cost and scope
- A quarterly summary generated rather than assembled
- Drill from any figure down to the underlying usage
Illustrative product view. Figures are examples.
Roughly half of it, typically.
Most companies find they can get the same AI output for roughly half the cost.
More detail
The findings are unglamorous: a frontier model running a classification step, prompts carrying context the task does not use, caching available and switched off, agent loops with no ceiling, workloads still running behind removed features, seats nobody opens, and committed tiers your volume already qualifies for.
Each arrives with a dollar figure, a named owner and a status, and stays on the ledger until it is done or declined with a reason. Realized savings are verified against actual spend, not assumed.
- Optimizations ranked by annual savings
- Identified, in progress and realized, tracked separately
- Verification against actual spend after implementation
- No reduction in AI output required
| Optimization | Workflow | Status | Savings |
|---|---|---|---|
| Open-weight model covers the classification step | Support triage | Done | $15,600 |
| Cache the system prompt, 78% of input tokens | Document workflow | In progress | $8,700 |
| Put a ceiling on the agent retry loop | Internal agents | In progress | $7,400 |
| Move batch-eligible work off synchronous | Data enrichment | Planned | $5,100 |
| Committed-use tier now earned | All providers | Planned | $14,200 |
| Right-size a GPU endpoint at 11% utilization | Self-hosted models | Under review | $6,900 |
| Reclaim unopened seats on two AI tools | Engineering | Under review | $3,900 |
| Total | $61,800 a month |
Illustrative product view. Figures are examples.
The market repricing itself is a planning input.
Model pricing changes several times a year, in both directions. New models arrive that do the same work for less.
More detail
Caching and batch pricing appear on workloads that have been paying full rate for months. Volume crosses into committed tiers.
New models, new prices and new discounts land every few weeks. You hear about them the week they land, priced against your own usage, in time to plan the change instead of reading about it afterward. For finance that turns optimization from an annual cleanup into a standing item with a pipeline, which is the only way it survives contact with a budget cycle.
-
A cheaper model now covers your largest workload
Ticket triage runs 41M tokens a month on a frontier model. A newly released mid-tier model matches it on your own eval set.
$21,300a month -
Input pricing dropped on a model you already run
No change needed. The saving lands automatically. The forecast has been updated and plan variance is now favorable.
$4,900a month -
Prompt caching is now available on the document workflow
78% of that workflow's input tokens are an unchanged system prompt. Caching them is a configuration change, not a rewrite.
$8,700a month -
Volume now qualifies for committed-use pricing
Twelve months of usage supports a commitment at the next tier. Draft terms and a break-even are attached.
$14,200a month
Illustrative product view. Figures are examples.
Put AI where it belongs in the P&L.
For any company selling an AI-backed product, AI is a variable cost that scales with usage.
More detail
It behaves like cost of goods sold and it belongs in gross margin, not buried in a general technology line where it distorts both.
Camaze reports AI cost per account, per segment and per plan against the revenue from each, so pricing and packaging decisions are made on evidence and margin compression is visible before it appears in the quarter.
- AI cost per account and per segment, next to revenue
- AI as a component of gross margin and cost to serve
- Accounts whose AI cost is outgrowing their contract value
- Cost per feature, for deciding what is included and what is metered
| Segment | Revenue per account | AI cost per account | AI 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.
Walk your own numbers.
30 minutes, read-only, no engineering time. You will leave knowing what AI actually costs you and roughly how much of that you could save.
The first quarter
What a finance organization typically has in place after ninety days.
Month one: the number
Every source connected, history backfilled, spend attributed to departments and products. The total is almost always higher than the working assumption, and the gap is usually cloud-billed model access and self-hosted GPU time.
Month two: the controls
Budgets set with named owners. Alerts live and routed. The savings ledger opened and worked largest first. Weekly digests going to budget owners, which is the change that moves behavior.
Month three: the plan
A driver-based forecast replaces the extrapolation. Chargeback runs into the ERP. The quarterly board summary is generated rather than authored, and the AI line arrives with its assumptions attached.
Keep reading
For FP&A teams
AI as a driver you can model, with variance you can decompose and a forecast that holds.
Read moreBoard reporting
The questions a board asks about AI spend, and the numbers that answer them.
Read moreAI budget ownership
Giving the AI line an owner, a budget and a review cycle instead of a shared concern.
Read moreQuestions people ask
Is this a cost-cutting tool?
It is an optimization tool, and the distinction matters. Every finding removes spend that is not producing output: the wrong model for a task, context nothing reads, retries that fail, seats nobody opens. Volume is unchanged.
The purpose is to remove the reason AI budgets get frozen. When the number can be explained and half of it turns out to be avoidable at the same output, the conversation moves from defending the line to deciding where to reinvest.
How long before we see a number?
The total is available the day sources are connected, with history backfilled. Attribution takes a mapping session, usually an hour or two. An itemized savings list with dollar figures is typically ready inside the first week.
Should AI spend sit in cost of goods sold or in operating expense?
That depends on the workload, which is exactly why attribution matters. AI serving your product to customers behaves like cost of goods sold and affects gross margin. AI used internally for productivity is operating expense. Camaze separates the two so the classification is evidenced rather than assumed. Your auditors will have a view on the specifics, and the underlying detail is available to support it.
Do we need engineering to be involved?
Not for implementation. Connection is read-only credentials and takes about 30 minutes. Engineering becomes involved when a finding requires a code change, which is a small subset. Configuration and commercial findings do not need them at all.
We only use one AI provider. Does this apply?
Yes. One invoice still gives you a total with no explanation, no attribution, no early warning and no forecast. See single-provider visibility.