Dr. Suresh Rajappa, Global Data / Tax Leader at KPMG LLP.
I believe the next serious conversation about enterprise AI will not be about which model is smartest. It will be about which organization can turn every token into measurable business value. I like to call this AI tokenomics: the economics of prompts, responses, retrieval calls, agent steps, model choices, latency, infrastructure, governance and human adoption. In traditional software, leaders bought seats and negotiated licenses. In AI, every useful interaction creates a variable cost. The strategic question is no longer simply, “Can we use AI?” It is, “Can we afford the intelligence pattern we are designing?”
The Economic Times recently captured this issue well by showing how token spending can surprise even disciplined enterprises. A regulated industry company reportedly processed 400 billion tokens in one month, producing a $78,000 bill, what would amount to roughly $1 million annually. The lesson here is not that AI should be slowed down. It’s that unoptimized workload design can quietly become a new tax on transformation.
I see this as a defining issue for chief data and AI officers, CDAIOs, because they sit at the intersection of ambition, architecture, governance and value realization.
How Good Tokenomics Can Help Your Business
CDAIOs need to treat tokens as a management signal, not just an engineering detail. Imagine a bank using AI agents to support relationship managers. The same customer insight could be generated three ways: by sending every request to a frontier model, by routing routine analysis to a smaller model or by using retrieval, caching and templates before asking any model to reason. The customer experience may look identical, but the economics may be completely different. A CDAIO who measures cost per insight, cost per resolved case and cost per revenue opportunity can move AI from experimentation into scalable operating discipline.
This is where agent economics becomes critical. A single agentic workflow can create hundreds of hidden model calls, while still appearing to the user like one simple request. An agent that researches a prospect, reads CRM notes, summarizes emails, drafts outreach, checks compliance language and logs the result may be valuable. It may also be wasteful if every subtask uses the most expensive model. The best CDAIOs need to design agent teams the way supply chain leaders design networks: by focusing on the right capability, the right task, the right cost and the right control point. Token routing will become as important as data routing.
Deloitte has described tokens as “variable input costs,” which is a helpful analogy. The point is not to minimize every token, just as manufacturers do not minimize every raw material. The point is to convert inputs into margin, speed, quality and growth. AT&T’s reported experience, cited in Deloitte’s report, is instructive: It scaled AI across employees and agents, then redesigned orchestration to cut costs dramatically even as token volumes increased. That is the future I expect: more tokens, not fewer, but with better economics, better measurement and clearer accountability for outcomes.
How Poor Tokenomics Can Burn You
The burn risk is equally real, though. CDAIOs can get hurt in five predictable ways. First, they may celebrate adoption without measuring value. Token consumption is not transformation; it is activity. Second, they may allow pilots to become production systems without cost architecture. Third, they may let vendors hide metered usage inside broad contracts. Fourth, they may mistake frontier models for universal solutions. Fifth, they may underinvest in data quality, forcing AI systems to ask longer questions, produce longer answers and repeat work. In that environment, the AI program may look successful in demos but painful in the P&L.
Consider a hypothetical retailer deploying a customer service agent. On one hand, the CDAIO can use clean knowledge bases, policy aware retrieval, small models for classification, larger models for complex exceptions and dashboards that show cost per deflected call. Customer satisfaction may rise, call volumes may fall, and the CFO may see a credible return. On the other hand, the same retailer may launch a glamorous chatbot on a powerful mode. But it may handle repetitive questions inefficiently, lack escalation design and discover after scale that token spend rises. The technology is similar, but the economics are not.
AI tokenomics needs to be a board-level conversation. CDAIOs should not own it alone. Finance departments should help define value thresholds. Technology leaders should manage architecture and observability. Risk assessors should define guardrails. Business units should be accountable for outcomes, not just usage. Procurement teams should negotiate transparency into tokens, rate cards, latency commitments, data residency and model switching rights.
Companies need to build an intelligence operating model where cost, quality, speed, safety and adoption are managed together. That operating model should also include chargeback rules, model approval tiers, red-teaming budgets and clear retirement triggers for workflows that never reach value targets. Without those mechanisms, AI portfolios can accumulate like unused software.
What To Watch For
When vetting your operating model, watch for a few key things. First, watch model routing. The question should shift from “Which model are we using?” to “Which model should handle this task at this moment?” Second, watch small language models and open-source deployment. Many enterprise tasks do not need the most powerful model available. Third, watch inference optimization, caching, compression and orchestration platforms. They will become the quiet infrastructure behind scalable AI. Fourth, watch pricing models. Seat-based software economics will increasingly collide with usage-based AI economics. Fifth, watch governance metrics. CDAIOs will need dashboards that connect tokens to revenue, productivity, risk reduction and customer experience.
My view is simple. AI tokenomics is not a cost-cutting exercise; it is the discipline of making intelligence economically sustainable. CDAIOs who master it can earn the right to scale AI responsibly. Those who ignore it may discover that the fastest path to AI adoption is also the fastest path to margin leakage. Many organizations do not yet have an AI strategy. They have enthusiasm, vendors, pilots and a rising bill. Tokenomics is how that enthusiasm can become an operating model.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
