Most software costs land on one predictable line: a licence fee, once a month, the same figure every time. AI usage doesn't behave that way. Claude and other large language models charge by the token, and token counts shift with every prompt, every attached document, and every extra step an agent takes to finish a job. For a CFO used to flat SaaS pricing, that variability is usually the first objection raised in the budget meeting, and it's worth having a straight answer ready before you're asked, rather than improvising one on the spot.
Why “what if usage doubles?” is the first question
In a fixed-fee tool, doubling how much your team uses it changes nothing on the invoice. In usage-based AI billing, cost scales directly with what the model reads and writes. A Sydney accounting firm running Claude across forty staff during tax season will spend more in July than in November, because more documents move through the system, not because the vendor changed the price. That's a different budgeting problem to the one most finance teams have already solved, and it needs a different kind of model rather than a bigger number pasted into last year's spreadsheet. The good news is that the underlying usage is measurable in a way a headcount decision never quite is.
Input tokens: everything the model reads, including the prompt itself, any attached documents, and anything pulled in from a knowledge base or search.
Output tokens: everything it writes back. These are typically priced higher per token than input, so verbose responses cost more than terse ones.
Tool calls and agent loops: every round trip through a connected tool, database, or MCP server adds another read-and-write cycle, and a multi-step agent can make dozens of these in a single task.
Prompt caching: system instructions or reference documents that repeat across calls can be cached at a fraction of the standard input price, often the single biggest lever for cutting recurring costs.
Model tier: Haiku, Sonnet, and Opus price differently by design. Routing routine work to a cheaper tier and reserving the top tier for genuinely hard problems is a cost decision, not just a technical one.
Building a variable cost model your CFO will trust
Start with one workflow, not the whole business. Pick something with stable volume, ticket triage, invoice matching, meeting summarisation, and measure its actual token usage over a real week rather than guessing. Most API providers, including Anthropic's own Console, break usage down by day and by model, which is enough to build a believable per-workflow unit cost without commissioning a separate analytics project.
Once that first workflow has a number attached to it, the same method scales sideways. A firm running five or six agent-assisted workflows doesn't need five or six separate cost models, it needs the same per-workflow discipline applied five or six times, sitting on a shared dashboard so finance can see all of them without asking the operations team to run a manual report every month. That repeatability is what turns a one-off estimate into something a CFO will actually sign off on.
A support-ticket triage workflow handling 3,000 tickets a month might use around two million input tokens and four hundred thousand output tokens on Sonnet. At current per-token pricing that lands somewhere between $180 and $260 a month, not the $65,000 a year a Melbourne business might budget for a junior analyst doing the same first-pass sorting. Put both numbers on the same slide and the usage-based model stops looking risky and starts looking like the cheaper option it usually is.
A five-minute answer for the budget meeting
CFOs don't need a token-by-token breakdown. They need three things: a worst-case monthly figure if usage triples overnight, a per-workflow unit cost they can compare against the alternative, whether that's a new hire, an outsourced service, or doing nothing, and a spend alert set in the billing console so nobody finds out about a cost spike from the invoice. Set that alert at a threshold agreed on in advance, tied to a specific workflow rather than the whole account, so a genuine anomaly gets caught without an over-eager cap throttling a busy but legitimate month.
For regulated clients the conversation usually extends one step further: where the data sits while it's being processed, and whether that satisfies obligations under the Privacy Act or, for APRA-regulated entities, internal data governance policy. That's a separate conversation to the billing one, but it tends to surface in the same meeting, so it pays to walk in with both answers ready rather than promising to follow up later.
If you're building this model for your own team and want a second pair of eyes on the numbers, book a short call and we'll walk through your actual usage pattern together.
