What rarely lands is on the other side: measuring the ROI from AI. Today, AI spends on the cost line of the P&L—infrastructure, inference, licensing—and it is tallied and reviewed like any other technology bill.
Goldman Sachs reports AI-related spending is tracking to exceed $800 billion in 2026, with token consumption likely to reach 120 quadrillion tokens a month by 2030, 24 times today’s volume. Today, only 23% of C-suite leaders report widespread and sustained business value from AI across their organization. Tokens are the units of data a model reads and writes, and everything you ask AI to do burns through them. They began as a line on an invoice, but now they are a genuine factor of production, an input to be tracked and managed like any other. The urgency is hard to overstate. So the real question for executives watching invoices climb is not how much AI costs, but how to make every dollar return more. Businesses have to apply the same rigor to AI tokenomics as they do other technological resources.
Closing this gap is the defining challenge of enterprise AI today, and most companies haven’t mastered the tokenomics discipline: tying what AI consumes to the value it gives back.
Most companies treat AI as a software expense, but it behaves like an economic system. Software costs are predictable: you buy seats, pay license fees, and the bill moves only when headcount or contracts change. AI throws that predictability out.
In nearly every deployment we’ve examined, including our own, fewer than 10% of users and workflows drive the majority of spending. For example, a financial services firm might find that 8% of its workflows, like automated loan pre-screening, drive nearly 60% of monthly token spend, while the contact center AI actually retaining customers barely registers on the invoice. Consider a PowerPoint presentation that once required four finance analysts and 20 hours of work. With AI, one person can produce it in two hours at a token cost of roughly $200.
Cost climbs with what people ask AI to do, how deeply the model thinks and how much work flows through AI. And increasingly, the heaviest users aren’t people, but AI agents and automated workflows running around the clock. Cost also concentrates sharply. Two enduring economic principles sum up the challenge: Jevons paradox suggests that cheaper tokens will not necessarily mean smaller AI bills. Just as more efficient steam engines expanded Britain’s coal consumption rather than cutting it, cheaper inference pushes companies to run more AI, not less. Work once too expensive to justify suddenly makes financial sense. Telling employees to use less AI is not the answer; it caps the productivity gains AI exists to unlock. The smarter move is to route new demand to the cheapest model that does the job well, so the same budget buys more useful work. Meanwhile, the Pareto frontier represents the best tradeoff between cost and value. Most companies overspend where the work doesn’t justify it and underinvest where advantage is genuinely at stake. Telling employees to use less AI is the wrong instinct as that caps the productivity gains it exists to deliver. The better move is to match each demand to the right model and manage consumption at the source, so the same budget buys more useful work. A handful of moves can turn AI scale into measurable return. First, the CFO and CIO should together own a single enterprise view of AI cost, usage and return. Without a unified view, the CFO cuts the wrong line. Second, embed controls into AI deployment from the start, including deciding who can use certain models and for which workstreams. By governing from day one, organizations can reduce invoice surprises. Third, draw a line in the sand that dictates which work is best for which models. Let lighter models handle tasks that break into well-defined steps and reserve the heaviest models for higher value or greater risk work, such as complex analysis or multi-step reasoning across long agent chains. Most enterprise AI usage is still search and retrieval, while a smaller group of power users demands advanced models for more complex work. Sending both workflows through the same model is how cost piles up. Fourth, measure the return, not just the cost. The value here is the thousands of dollars of capacity freed up and redirected to higher-value work. And finally, repeat. Tokenomics is not a one-off exercise but a continuous process that must evolve as new models, tools and pricing emerge. It’s closer to treasury management than an IT project. The economics of AI are being set right now. Companies are dividing into two groups: those that treat AI as a technology cost with a bill that lands each month and those that manage it as a source of measurable return. The first will spend their energy fighting compounding bills. The second will make AI’s value compound faster than its cost, turning scale into durable competitive advantage. Companies that treat tokenomics as a discipline will spend the next decade ahead of the ones that wait. By Manish Sharma, chief strategy and services officer, Accenture and Lan Guan, Chief AI and Data Officer, Accenture Get the latest technology news and updates. Download the TOI App.