Finance leaders have been handed two instructions that pull against each other. Prove the return on AI spending, and sign for it on terms where nobody can say in advance what the bill will be. The second instruction is the one setting the terms of the first, and the vendors closest to the problem have started saying so out loud.
Forty-six percent of the 200 North American CFOs in Deloitte's 2Q 2026 CFO Signals survey named cost uncertainty or a lack of transparency as their biggest internal concern about their own organization's AI use. They run organizations above a billion dollars in revenue, and it was the leading answer, ahead of every governance and capability worry on the list. Deloitte connects it directly to how the software is sold, noting that many AI providers now charge by consumption rather than a flat rate, and that when usage fluctuates the bill becomes hard to predict. In the same survey, 93% said their organization already uses AI extensively or modestly across multiple functions, which is what makes the pricing question urgent rather than theoretical. Adoption isn't waiting for the commercial model to settle.
Underneath that sits a measurement gap the finance function is being asked to close on a deadline. Research from CloudZero surveying 260 senior finance leaders, 135 of them CFOs, found that 87% must tie AI spend to business outcomes within the year while 22% can do it today, and that 66% of boards now condition further AI funding on proof of return. CloudZero sells AI cost-allocation software, so the survey measures a problem the company monetises, and the figures deserve to be read with that in view. The internal structure of the data is harder to wave away, though. Teams with same-day visibility into AI spend were roughly twice as likely to hold an aggressive investment posture as teams without it, and where budgets did overshoot, that same visibility tracked with materially fewer serious consequences. The association is consistent, though a survey of this shape can't say which way it runs.
The Pricing Model Changes Hands
Vendors have noticed, and one of them is making the argument from inside a finance organization. Patrick Villanova, CFO of BlackLine, told CFO Dive that token-based pricing leaves "the sky's the limit in terms of how much you can charge", which he described as very challenging for buyers who need predictability. His company's answer is to sell an outcome, a stated number of reconciliations automated, while BlackLine continues to buy the tokens itself and carries the consumption risk on its own books. He was candid about what that demands internally, saying the business has to be very cautious that whatever price it charges justifies the rate of token consumption underneath.
For buyers, the practical advice from people doing the buying is about sequence more than bargaining power. In a separate interview with CFO Brew, Villanova reduced it to an instruction that costs nothing to adopt, always define the outcome first, and never start with pricing. Matthias Steinberg, CFO of MindBridge, described the same shift from the other side of the table, noting that consumption and performance pricing make life harder for finance because the predictability that seat licences offered has gone. His answers are operational. Tell staff at all-hands that these are expensive tools, stop defaulting to the frontier model when most applications don't need the most cutting-edge one, ring-fence a deliberate experimentation budget, and build a routing layer that sits between what a user wants to do and the choice of which model actually runs. He expects every company to have one before long. None of that requires a new contract, and all of it shrinks the number that has to be defended later.
Whose Line Item This Becomes
The gap between deploying agents and governing their cost is where the finance seat is being drawn in whether it asks to be or not. OpenAI and PwC, announcing an expanded collaboration to build agents inside OpenAI's own finance organization, put token spend on the list of things a CFO will need to see. Their framing is that as agentic workflows scale, finance leaders need visibility into AI usage, token consumption and projected spend so adoption can be governed the way other enterprise operating costs are. The work starts with a procurement agent and runs through planning, forecasting, reporting, treasury, tax and the close, with OpenAI describing its own finance team as customer zero. Bloomberg Tax reported the same arrangement independently, and PwC's US advisory leader Tyson Cornell framed it in the firm's own release as finance moving from process efficiency toward decision-centric operations. The partnership matters less than the line item it hands finance. Deloitte's respondents put AI governance ownership with the CISO in a third of organizations and with the CFO in 19%, and consumption billing pulls steadily on that second number.
Counting the Attempts That Fail
What outcome pricing does not do is answer the question underneath it. Moving from tokens to outcomes shifts consumption risk from the buyer to the vendor, which is a genuine improvement in predictability, and it relocates the argument to what counts as an outcome. A useful way to see the size of that argument comes from a worked calculation by Pratik Rupareliya, co-founder of the AI development firm Intuz, published in Towards Data Science, which proposes measuring cost per successful outcome by loading the cost of failed attempts onto the successes. On its anonymized example, an agent at $3.40 per attempt with a 71% resolution rate costs $4.79 per resolved ticket against a $4.20 human baseline, before accounting for what Rupareliya calls double payment on escalation, where the buyer pays the agent to try and then pays a person to finish. The specific figures come from a single unnamed deployment and should not be carried into anyone else's business case. The arithmetic travels perfectly well on its own.
That is the clause worth negotiating hardest, because almost nobody has published one. HighRadius put outcome pricing in writing in February, offering office-of-the-CFO software with no implementation fee, no subscription until go-live and a share of realized savings, governed by a mutually agreed success criteria document with baseline and target metrics. What isn't yet in the public record, from that vendor or any other, is an independent customer account of how a success baseline was audited, how a disputed gain-share was settled, or who absorbed the cost of the attempts that failed. BlackLine's reconciliation-count unit is described so far in the vendor's own words. The contracts exist; the case law of them does not.
Which puts the achievable work back inside the finance function, where it can start this quarter without waiting for the market to standardize. The clause nobody has published yet is the one worth drafting first. Who pays for the attempt that fails, and for the person who finishes the job afterwards, is a question both sides will answer generously while they still want the deal and defensively once the first invoice lands. A finance team that gets that answer in writing beforehand has bought something more durable than a discount.
