By Adam Albrecht, Head of Domain Intelligence, James Thompson, Head of Engineering, & Chase Hammon, Advisor at Translucent

Deloitte published a survey on the gap between AI investment and AI value in healthcare. 73% of CFOs say they're expected to be heavily involved in enterprise AI decisions. Only 49% feel equipped to contribute effectively. Among organizations furthest along in AI adoption, only 18% can consistently connect AI spending to a measurable financial outcome. The pattern is consistent across the survey: investment is accelerating faster than the discipline required to govern it.

Deloitte's prescription for closing this gap includes a new function it calls AI tokenomics: treating token consumption with the same rigor CFOs have historically applied to licenses, headcount, and capital equipment. The logic is sound on its own terms. Track usage, model volatility, and forecast a cost base that scales with adoption.

For a specific category of AI deployment, this is the correct response to a real problem. For most health system finance functions, this model doesn’t make sense.

The Cost Structure Is a Function of the Purchase, Not of AI Itself

Token-based cost volatility is not an inherent property of artificial intelligence. It is a consequence of a specific commercial model: licensing direct access to a general-purpose model and building an application layer on top of it internally. Under this model, usage and cost are structurally coupled. As adoption grows, so does spend, at a rate that is difficult to forecast and, without active management, difficult to attribute to a specific outcome.

Deloitte cites an illustrative case: a large health system experienced token consumption growth of 8 to 10 percent per month, translating to more than $6 million in annualized, unplanned cost increases before finance had visibility into the underlying driver. This is a legitimate and growing risk for organizations that have adopted this purchasing model. It is not, however, an unavoidable cost of AI adoption in general.

The forecasting challenge compounds further because pricing itself is unstable. Model providers adjust token pricing on their own timelines, for reasons that are rarely disclosed and difficult to anticipate. New models are released every few weeks, each with a different cost structure and a different performance profile for a given task. A forecast built this month can be obsolete by the next model release. Organizations that have accepted token-metered pricing are left re-running cost scenarios on a recurring basis simply to stay current, a task that consumes finance capacity without adding financial insight.

Once an organization accepts token-metered pricing, the governance response Deloitte outlines follows logically: model routing to match task complexity to cost, caching and batching to reduce redundant computation, chargeback mechanisms to align incentives, and scenario-based forecasting to manage volatility. These are sound practices, but incur massive resourcing costs on the health system. Governance requires ongoing operating discipline, layered on top of finance functions that are already stretched thin.

An Alternative: Price the Outcome, Not the Input

Translucent was built on a different commercial premise. A single subscription fee. Unlimited users. Unlimited token consumption. No usage tiers, no metering, and no chargeback infrastructure for the client's finance team to design or maintain.

This model is viable because the token is not the product. The product is a defined set of financial outcomes for health system performance: margin improvement opportunities identified, payer contract leakage surfaced, service line economics clarified at a level of granularity most systems cannot produce internally. Token consumption is an input cost that Translucent manages on the client's behalf, and that cost is more variable than it appears. Different models carry different token costs, and the right model for a given task changes as the underlying technology evolves. Deciding which model to use for a given task, whether that model is being used efficiently, when to adopt a newer model, and how to test one model against another for a specific workflow is a continuous, technical discipline. Translucent absorbs that discipline as part of the platform. The client purchases an outcome whose value is a substantial multiple of the compute required to produce it, without needing to evaluate the model decisions that produced it.

Three structural conditions in health system finance make the economics sustainable:

The function is systematically underserved. Few health systems can hire or retain the volume of skilled financial analysts that tasks such as payer contract modeling or multi-entity service line analysis require at scale. Demand for this analysis consistently exceeds the capacity to perform it. Where AI closes this capacity gap directly, the value generated per unit of AI output is materially higher than in functions where comparable analytical capacity already exists.

The underlying data is well suited to AI-driven analysis. Claims data, remittance files, EDI 835 transactions, general ledger detail, and payer contract terms are structured, high-volume, and consistently formatted relative to unstructured data in other domains. Structure alone is not sufficient. Interpreting this data correctly requires the tribal knowledge and institutional context that experienced finance teams carry but rarely document: which payer relationships have unusual terms, which cost centers have known reporting quirks, which historical adjustments still affect current numbers. Incorporating that context into an analysis has historically been slow and dependent on the availability of the specific people who hold it. AI systems can now capture this context in a persistent layer that sits alongside the underlying data, making it consistently available to every analysis rather than confined to institutional memory. The combination of well-structured data and accessible customer context is what makes the rate at which token consumption converts into usable financial output correspondingly high.

A substantial backlog of high-value analysis remains unaddressed. Health system finance has not yet been subject to a generation of AI-native tooling comparable to what other enterprise functions have absorbed. Significant value-generating analyses remain undone, not because they are conceptually difficult, but because the analytical capacity to perform them has not existed. Directing AI capability at this backlog produces returns that are disproportionate to the computational cost involved. It may cost $100 in compute to produce $20 million in actionable financial opportunity for the system. Both the platform and the health system can win without a complex token-based pricing model. 

Taken together, these conditions change the underlying economics. At Northwestern Medicine, a coordinated, enterprise-wide view of financial data compressed an analysis that previously required two to three days into 90 minutes and surfaced tens of millions of dollars of margin opportunity. That outcome was delivered under a flat subscription fee. It did not require Northwestern to build a token governance function, because the underlying token economics had already been solved on their behalf.

The Question CFOs Should Be Asking

For CFOs evaluating AI-native finance platforms, the relevant diagnostic question is not "how do I build the internal discipline required to manage this vendor's token costs." It is "does this vendor's commercial model—and the value it provides—require me to build that discipline at all."

Tokenomics, as Deloitte describes it, is helpful guidance for organizations purchasing raw model access and assuming responsibility for the application layer built on top of it. It is the wrong framework for evaluating a platform architected to absorb that cost volatility structurally, rather than pass it through. The health system CFOs who extract the most value from AI over the next several years are unlikely to be those with the most sophisticated token governance function. They are more likely to be those who selected partners for whom token governance was never the client's responsibility to begin with.

‍