Tokenomics: CFOs have a new challenge for CIOs—make AI spend financially legible. Accenture data suggests new disciplines needed

Four in five dollars of AI token spend lack a quantified link to business outcomes. Or in other words, less than one dollar in every five can be shown to have a verifiable financial outcome. For CIOs faced with demand from the C-suite to get with the AI program, that creates a twin dilemma - they can’t predict consumption, resulting in a token bill for the CFO to pay that they can’t forecast, and they can’t demonstrate what that spend creates in terms of value. That’s the top line conclusion from new research from Accenture - The CIO’s Guide to AI Tokenomics - which polled 750 senior global executives across 17 countries. The enterprises in the survey spent about $2.5 billion on AI tokens last year in total, but the findings suggest that this could climb as high as $3.6 billion over the next two years, with respondents predicting token consumption to grow 78% even as per token pricing declines by 19% over the same period. The CFO sees every dollar that’s been spent but in most cases has no explanation for ROI on that spend. There are three questions that need to be answered - what is driving AI costs, how fast will those costs grow, and what provable business value is AI creating? The study notes: > AI creates value where Finance cannot see it The cost side of AI is cleanly decomposable: tokens are 25% or less of total AI spend per our survey - identifiable, traceable and attributable to specific workloads and teams. The value side compounds across a whole workflow: a productivity gain created by an AI-assisted workflow reflects the model, the data it retrieved, the infrastructure running it, the engineer who designed the workflow and the employee who used it. No amount of careful measurement will isolate the share of the realized value that belongs to tokens alone. In hard cash terms, the data is unedifying for the fiscally minded: > Today, only one dollar in five of token spend can be translated into a financial return. With full chargeback and precise measurement, provable value climbs to around 32 cents on every dollar spent. On every $10 million of spend, that shift makes approximately $1.15 million of value provable for the first time. Consuming needs Accenture identifies that token consumption originates across three broad categories - employee and developer tools, which account for around 30% of spend; internal process automation and enterprise software, roughly 29%; and customer-facing applications, analytics and agentic workflows, at around 39%. (Coming up as another driver of cost growth are agentic workflows which multiply inference calls at every step and as such fuel token consumption very fast. These currently only account for 11% of spend, but in future…) To complicate matters, everyone buys into the idea that there is ROI from AI investment; they just can’t prove it financially. So the CFO has a challenge that the over-worked CIO must add to their to-do list , namely creating the instrumentation and operating discipline that makes both cost and value financially legible, as well as reducing the AI spend bill. That’s going to mean getting the corporate head around a whole new way of thinking and counting: > Companies still manage AI the way they managed enterprise software a decade ago: They wait for the bill. That breaks down fast now that spending has shifted from fixed software licensing to usage-based AI, and consumption is climbing across employees, agents and workflows. The study adds: > Budget models built before agentic tools became mainstream cannot keep up….One in three organizations has already exhausted its annual token budget before the year ended. The token bill will not stabilize on its own, and approaching AI spend like this is no longer sustainable. Blind spots Accenture breaks out six blind spots when it comes to AI budgeting: 1. Unknowable task costs - as tokens have no intuitive human equivalent; cost emerges only after tasks run. 2. Using frontier models by default - people choose the most capable model even when lower-cost models would be sufficient. 3. Agentic multipliers -the likes of chained calls, expanding context and repeated processing only serve to multiply consumption. 4. Budgets built for a pre-AI model of enterprise computing -those that pre-date mainstream agentic tools and usage growth and, as noted above, are essentially unfit for purpose. 5. Fragmented pricing - tapping into multiple providers, models and pricing structures prevents a consolidated view. Less is more… 6. Visibility without accountability, ‘showback’ trumps playback. But limited chargeback weakens cost ownership. Discipline needed To counter all this, Accenture suggests five disciplines for the CIO to get his or her head around: 1: Make every workload observable before it scales. Deploy an AI gateway or observability layer that captures, for every AI interaction: the workload name, the owning team, the model used, the token cost and the output. 2. Make teams pay for what they use. Bill consuming teams for their token costs. Start with engineering and the top three or four consuming functions. Showback without financial consequence changes awareness but only changes behavior when coupled with other incentives. In our research, chargeback is one of the strongest predictors of cost optimization and value attribution, but fewer than one in 12 organizations use it today 3. Send each task to the model it actually needs. Identify the approximately 10% of workloads or process steps that genuinely need frontier capability—the complex reasoning, planning, multi-source synthesis and high-stakes generation—and route everything else to capable mid-tier or open-weight models. Enforce this as a routing policy at the gateway, not as guidance left to each developer’s judgment. The price spread is significant: a frontier model can cost 15 times or more as much as a capable lower-tier model. 4. Demand a value case and a management standard before any workload reaches production.Define three things before you commit capital to any new AI workload: what the process costs today, in money or time; what specific financial outcome AI will improve; and how that outcome will be measured in dollars. Then hold the gate. Approve the business case only when observability tooling is in place to measure consumption at the workload level, optimization is on a credible path toward break-even, and the expected value can be expressed as a quantified financial outcome, whether revenue, cost avoidance or productivity in dollars. This does not need to be complex. It needs to exist before the spend, not while trying to justify it afterward 5. Build AI skills across the organization. Most routing mistakes happen because people lack clear guidance, not good judgment. Developers and employees choose the most capable model because they can’t see the cost difference—and often don’t have a simple rule for when a less expensive model is enough. Training helps, but only to a point: 65% of companies have introduced some form of user training, and the routing gap remains. Awareness doesn't change decisions unless it's re-inforced at the moment of choice. My take > The discipline is the differentiator. It’s not too late to get on top of this problem, argues Accenture, but it does demand a new way of thinking and doing: > CIOs who act now will have an answer. Tokenomics is becoming a standard discipline for running enterprise AI. Build it now and the CIO can give the CFO confidence in the economics, give the CEO a return they can quantify and give the enterprise permission to scale AI on evidence rather than hope. A big ask perhaps, but one that is impossible to avoid addressing.