Every finance team I have worked with over the last thirty years has a story about the moment a quiet line item on the ledger stopped being quiet. In the print houses I started out in it was computer time, metered on a shared workstation until personal machines got cheap and the line vanished. In the two-thousands it was outbound bandwidth. More recently it was storage, then egress. The pattern is always the same: a new input arrives, nobody owns it, it grows faster than anything else on the sheet, and at some point the CFO walks into a room and asks who approved this.
AI token spend is that line now. The industry is past the is it big enough to care about phase and into the phase where an enterprise either gives it a budget line of its own or watches it bleed out of a dozen cards and a dozen pilots that never get reconciled. The question a buyer has to answer this quarter is who, inside the organisation, owns the arithmetic once it arrives.
I went and read what analysts, finance leads and platform executives are saying about that on the record. The honest answer is that nobody who has to make the decision agrees on the shape of it either.
The eight voices this reads from
Trade-press noise on AI budgeting is thick and mostly unattributed. The piece below only uses named statements with a link to where the person said them. Eight of them, chosen for sitting on different sides of the same question.
- Shari Lava, Group Vice President of AI, IDC (IDC blog, 10 September 2026).
- Guan, Chief AI and Data Officer, Accenture, co-author of CIO’s Guide to AI Tokenomics, 10 September 2026 (full name on the linked report).
- Ajoy Menon, Digital Core Reinvention Partner Lead, Accenture, same report.
- Karim Atiyeh, Co-Founder and Co-CEO, Ramp (Ramp blog, 9 April 2026).
- Ara Kharazian, Lead Economist, Ramp (InformationWeek, 4 August 2026).
- Hemant Kapadia, Chief Financial Officer, Anaplan (same piece).
- Mohan Sankararaman, Executive Vice President and CIO, First Horizon (same piece).
- Yogesh Joshi, Senior Vice President of Global AI Platforms, TransUnion (same piece).
The bill is already real, and most of it is off the ledger
Start with how large the line has already grown. Karim Atiyeh at Ramp reports that the average monthly AI token spend across Ramp’s 70,000-plus customers has risen roughly thirteenfold since January 2025. That is not Ramp’s inference, it is their own invoicing data. The biggest spenders see their AI bill jump by fifty percent or more in roughly one month out of every four. Atiyeh’s line about what that feels like inside a finance team is worth quoting straight: “Your AI bill? You’re guessing.”
Accenture’s CIO’s Guide to AI Tokenomics, 10 September 2026 by Guan, Ajoy Menon, James Burrows and Surya Mukherjee, sizes the global number. Enterprises spent roughly 2.5 billion US dollars on AI tokens last year, and the authors project that aggregate token consumption grows by seventy-eight percent over the next twenty-four months. They also report the tell every CFO recognises: one in three organisations exhaust their token budgets before year-end, and only thirty-five percent can compute a cost per business outcome for even their largest AI use case.
IDC’s Shari Lava lays the governance claim on top of those numbers. CFOs already influence more than two-thirds of enterprise technology investment decisions, she writes, and nearly three in four organisations now flag excessive AI spending as a material risk to their broader technology plans.
“Every dollar of wasted AI agent spend is a governance failure.”
Shari Lava, Group Vice President of AI, IDC, 10 September 2026
Position one: Give the line its own owner, or accept that nobody owns it
The first position, held by the analysts and the platform vendors, is that the fix is structural. A budget line nobody owns gets spent by everybody. Lava’s IDC data makes this sharply: only one CEO in five has given AI its own budget line. The rest absorb the cost into existing software or cloud lines, which makes it invisible at the point a decision is made about how much to spend.
Accenture’s report is more specific about what owning the line looks like. Guan and her co-authors recommend a workload-level observability layer, which they call an AI gateway, that attaches owner, model and token cost to every interaction. On their numbers, fifty-three percent of enterprise token usage is already traceable to a user or team when the plumbing is there, and fifty-four percent of AI requests are currently misrouted to more expensive models than the task needs. Both are governance failures, in Lava’s sense, not technology failures.
Atiyeh’s view from the ledger side agrees. The examples in his April post line up into a pattern: one Ramp customer uncovered 120,000 US dollars a year of AI spend that showed up on no provider dashboard, only on employee cards. Fourteen ChatGPT Team subscriptions turned up on personal expense claims at a single company. Finance teams already know how to do this, he argues: they did it for cloud and for SaaS. They need the same visibility over the AI column.
Position two: Chargeback is the wrong lever, and too early
The second position, the sceptical one, is held by people who run finance or measure it for a living. Their point is not that the line should stay invisible. It is that reaching straight for chargeback in year one is a category error, because the arithmetic of chargeback assumes you can tie spend to outcome, and nobody has solved that half yet.
Ara Kharazian, Ramp’s Lead Economist, is explicit in InformationWeek. “It’s very difficult to measure AI’s impact on a business,” he told Stephanie Overby for the 4 August 2026 piece. The usage numbers everyone is reporting, he added, “say nothing about the quality of work being produced.” Kharazian is the economist for the company telling finance teams to track every token, which is why his caveat matters.
Hemant Kapadia, CFO at Anaplan, frames the same objection as a measurement trap. “You end up measuring a pilot instead of measuring a capability,” he told InformationWeek. AI chargeback gets built from whatever spend you can see first, which is pilots, and pilots do not tell a CFO what the fully deployed capability costs and returns.
Accenture’s report acknowledges this half-solved status by naming its intermediate stop, showback: making consumption visible without yet charging it back. Only seven percent of companies in their sample have gone as far as chargeback. The rest are either in showback or in the fog before it.
Position three: ROI per use case, or stop opening more of them
The third position comes from inside enterprise IT. The operational reply to the budgeting debate: you fix the measurement problem not by inventing a smarter cross-charge, but by refusing to let an AI use case into production until it has a declared purpose with a return attached. Mohan Sankararaman, CIO at First Horizon, put it to InformationWeek in one sentence: “Every AI use case should have a clear purpose and ensure that purpose has ROI.” He added that “doing more AI experiments doesn’t equate to more AI value.”
Yogesh Joshi at TransUnion carries the same discipline into the platform layer. His advice in the same piece: “start small and go deep,” and “AI can’t be a hammer looking for a nail.” Both lines describe a procurement posture, not a technical one, and that is why they belong in a budget conversation.
Where Sankararaman and Joshi land matters for what you put on the invoice. If a use case cannot defend its own ROI, chargeback for its tokens is an accounting cosmetic. The spend that deserves its own budget line is the spend that has already been justified somewhere else, and the function of the line is to keep that justification honest over time.
Position four: Finance is a capital-allocation seat, not a cost-containment one
Running underneath the three positions above is a quieter argument about which seat in the C-suite holds the AI budget. Kapadia at Anaplan gave the clearest version: he treats AI measurement less like a software purchase and more like a capital allocation decision. That framing moves AI out of the OpEx column, where the instinct is to trim it, into the capital-allocation conversation, where the instinct is to size it against the return it is expected to produce.
Three things follow. The CFO sets the hurdle rate, not the CIO. The CIO stays responsible for keeping the AI usable and governed, which is where Accenture’s gateway lives. And the lead team for a given use case, marketing or operations or support, holds the business case for the capability. Ownership is distributed by function, and the budget line exists to make the distribution auditable. Greg Cooley at AngelList, quoted in Atiyeh’s Ramp post, captured the current state: “AI is moving faster than the finance context around it.”
What a mid-market commerce team should put in a supplier contract next Monday
If you are an ecommerce director or a digital lead about to sign an agency or platform contract that involves AI, the above translates into four clauses you can insert without inventing anything. We have used versions of these inside our own agency billing since 2026-09-13, when we started pricing AI as its own line item instead of absorbing it into fixed fees.
Clause one, unit declaration. The contract names the billing unit used for each AI input: tokens at a provider’s published unit price for model spend, API calls at the provider’s price for external services like transcription or embeddings, human hours for human work. No internal “AI credits” only the supplier can reconcile. Our own weighted-tokens formula (input + 5 × output = weighted tokens, one unit of 100,000 weighted tokens roughly 10 to 15 minutes of active AI work) compresses model-tier spend into one legible number, and it is on our invoice-line-items post.
Clause two, audit right. The contract gives the buyer a right to inspect the raw provider invoices that underlie the AI line, through an auditor where commercially sensitive. When 120,000 US dollars a year of AI spend can hide on employee cards, the audit clause is the only mechanism that keeps the supplier honest without forcing the buyer to build a gateway of their own.
Clause three, measurement baseline. Before the AI feature goes live, the buyer records a baseline of the business metric it is supposed to move. For a WooCommerce store, that is usually a conversion rate at a step, a support deflection rate, or a time-to-first-draft for product descriptions. Six weeks in, the baseline is re-measured and the result lands in the quarterly review alongside the token line. This is the Sankararaman clause: no use case without an ROI, on paper, before the first token.
Clause four, routing policy. The contract states which model tier is used for which class of task, and commits the supplier to the cheaper tier when the task does not require a frontier model. Accenture’s finding that fifty-four percent of AI requests are currently misrouted to more expensive models than they need makes this material, not theoretical. Inside WooCommerce, this maps to product description generation (Haiku-tier) versus agentic customer support (Sonnet or Opus for the complex paths, Haiku for the bulk of greetings).
These four clauses do not solve the measurement problem Kharazian and Kapadia flagged. They do something smaller and more useful: they give the buyer a budget line they can read, audit and route, while the measurement side catches up. The audit clause alone is what makes the AI line credible to a procurement team.
What is still unresolved
The eight voices above agree on about half of this. They agree the line exists, that it is growing fast, that one in five organisations is already behaving as if the line is real and the rest are not, and that visibility has to come before chargeback. They disagree on what to do with the line once you have it. The Accenture and IDC view treats it as the start of a governance capability. Kharazian and Kapadia warn that measuring it too hard, too early, produces a pilot-grade number mistaken for a capability-grade one. Sankararaman and Joshi sidestep the whole fight by refusing to spend on a use case without an ROI argument attached. Atiyeh builds the tooling that makes either side’s answer legible to a CFO.
My own position, because the rest of this is quoted and I owe you mine: the line goes in now, the chargeback waits a year, the ROI check is enforced on every new use case from this quarter onward, and the supplier contract carries the four clauses above whether or not the internal finance context is ready. If you are evaluating a platform or an agency with AI deliverables in the statement of work, and your current contract does not let you see the token line separately from the hours, that is the conversation worth having first. We spend a lot of our week on exactly this kind of contract, and the agency page is where that one begins.
Last modified: November 9, 2026
