You Cannot Govern What You Cannot See: What Enterprise Tokenomics Looks Like in Practice

KEY INSIGHT: A conversation with Sonali Niswander, Senior Vice President of Global Technology (AI, Cloud and Data) at MetLife, surfaced a pattern showing up across the Fortune 100 and FTSE 100: managing AI token spend has become a board-level concern, and leadership is pulling ahead of practitioners in asking about it. The organizations getting traction are working in three stages, consumption visibility first, then optimization, then value realization, and they are shifting the cost conversation left to design time rather than treating it as a chargeback problem after the fact. The recurring ask from these leaders is not a vendor tool. It is a vendor-neutral standard for measuring the cost and value of AI. That is the gap the community is positioned to fill.

---

The room sounds like 2017 again

In early 2026, most of my conversations with enterprise technology leaders were about FinOps going beyond cloud: shifting left, taking on broader scopes, and expanding the mission beyond public cloud cost. Useful conversations, but familiar ones.

That changed fast. When I sat down for a fireside chat with Sonali Niswander, Senior Vice President of Global Technology at MetLife, we scrapped the planned questions and went straight to tokens. Her remit spans cloud, AI, developer tooling, DevSecOps, enterprise architecture, and technology governance. MetLife operates across 40 countries and six regions, giving her a genuinely global purview. When someone in that seat says the questions need to change, the questions need to change.

What she described from the executive forum earlier that day was striking in its ordinariness. The conversations were real and practical. Everyone in the room was wrestling with similar problems, similar questions, and the same lack of shared answers. It felt eerily similar to cloud in 2017, when little was defined and everything was inconsistent.

One thing was different, and it matters: the AI value conversation is no longer contained to the CIO and CFO. CISOs are now part of it as AI and security intersect. CEOs and boards are asking questions as well. My own path into this topic started at a global CIO event where three CIOs from large financial organizations used the word “tokenomics” without prompting. What began as a practitioner concern is becoming an executive concern.

The clearest ask in the room: a neutral standard

When I asked Sonali what she took away from a room full of her peers, her answer was not about a product. It was about a gap.

Frontier model providers keep releasing new models. Meanwhile, enterprises are building their own cost frameworks and their own ways of measuring value, often in isolation and from scratch. What these leaders want is a vendor-neutral set of standards and frameworks so that every organization does not have to solve the same problem privately without a shared language.

That is exactly the role a neutral, community-owned discipline can play: not to sell a methodology or pick a winner among model providers, but to give practitioners a common vocabulary and measurement approach that works across vendors. The market has been here before with cloud. The FOCUS specification exists because the industry decided cost data should be comparable across providers rather than fragmented across bespoke billing formats. AI consumption and value measurement now present a similar coordination problem.

A working definition of enterprise tokenomics

In this conversation, enterprise tokenomics means managing the cost, consumption, and value of AI tokens—and connecting that activity to measurable business outcomes.

When I pushed Sonali on whether tokenomics is different from FinOps, her answer was measured. It is not a replacement for FinOps. It adds new dimensions to the practice.

Inside MetLife, that work progresses through three stages:

1. Visibility: Understand where and how AI consumption is occurring.
2. Optimization: Match models, controls, and architecture to the workload.
3. Value realization: Connect consumption to measurable business outcomes.

This is one enterprise’s operating model, not a prescribed framework. But the sequence is worth studying because it maps cleanly to how FinOps has always worked.

Stage one: consumption visibility

AI consumption appears from many places at once: internal AI platforms, SaaS products, vendor APIs, and tools used across technical and non-technical teams. A unified view does not come out of the box. The first investment, therefore, is instrumentation. An organization must be able to see where consumption is happening before it can manage that consumption coherently.

You cannot govern what you cannot see. That line is the whole discipline in six words.

Stage two: optimization

Once you can see consumption, the question becomes whether you are using AI efficiently. Two moves stood out in the conversation.

The first is model routing. The question is shifting, thankfully, from “Are we using the latest and greatest model?” to “Are we using the right model for the right task?” That routing intelligence can be built into the platform so that cost-aware decisions are made by the system rather than independently by every user.

The second is the use of evaluations at design time. The tradeoff between cost, performance, and accuracy should not be an afterthought. Engineers and business partners need that context while they are designing the solution, not after implementation when the chargeback arrives. Cost context at inception, not after the chargeback lands.

Stage three: value realization

This is the hardest stage, and Sonali was candid about it.

Tying token spend to business outcomes requires correlating data across multiple systems. Most organizations are still early in that work, and it also requires distinguishing between two kinds of value measurement.

Efficiency use cases often have an existing baseline. A process previously took a known number of hours or cost a known amount. If AI reduces that time or cost, the comparison can produce a credible business case.

Growth and net-new use cases are harder because AI may enable something the organization could not do before. There is no historical baseline.

In those cases, MetLife starts with a value hypothesis. Teams run structured, time-bound pilots, define success up front, and measure leading indicators to determine whether the hypothesis is being confirmed. The result is not only cleaner measurement, but also preventing pilots from running indefinitely.

A pilot produces evidence or the organization moves on, which is its own form of cost control.

Unit economics, all over again

The value problem is where the FinOps déjà vu is sharpest. We spent years learning to tie cloud cost to a unit of business value. AI raises the same question with higher stakes, because so much AI is customer-facing and should tie to an outcome.

The measurement rigor described above is not a reporting exercise bolted on at the end. It helps decide whether a workload continues receiving investment. That reframes the FinOps role from cost reporter to a partner in the investment decision itself.

The speed problem is a governance problem

Every enterprise leader I talk to is under the same pressure. CEOs want speed. CFOs want efficiency to fund the next round of AI investment. CISOs want a control environment they can defend. And underneath all of it, there is bottoms-up pressure, because these tools are no longer just for engineers.

That last point deserves emphasis. At the Linux Foundation, after the engineering teams, some of the most aggressive AI adopters are in legal. Marketing, sales, and customer success are close behind. Technology leaders spent a decade learning to govern engineers spinning up cloud resources. Now they have to extend guardrails to people who have never touched an infrastructure control, while recognizing that those same non-technical users may be driving genuinely transformational work.

The resolution she offered is not to slow experimentation. It is to build the guardrails into the platform. Spend limits and caps live in the system. Model access is shaped by persona and workload classification, so the engineer working on something complex can reach for the heaviest reasoning model while a routine task is routed to something cheaper and prescribed. Not everyone needs to become a model expert, and they should not have to be. The governance is in the plumbing, and it is paired with constant education about when a deterministic, rules-based approach beats a probabilistic model in the first place.

Speed and governance are not opposites here. Built well, governance is what lets an organization go faster with confidence.

The pendulum, and the workload-classification lever

There is an iron triangle in the room right now: speed, governance, and value. The instinct to slow down and get it right is real, and so is the pressure to move. Her answer was that the right balance depends on the problem. Not everything needs speed. In a domain where accuracy is the differentiator, getting it right beats getting it fast. The target is not a single org-wide setting. It flexes with the workload.

That connects to a lift-and-shift lesson from the cloud era. In cloud migrations, we talked constantly about refactoring and modernization, and then a lot of the time we just rehosted. The AI parallel is using AI as a more expensive way to do something we could already do, instead of for genuinely net-new value. The discipline that guards against this is workload classification, done early, so the choice to use AI at all is deliberate and the model matched to the task.

What leaders should do right now

I closed by asking what a leader in this seat should be doing today to prepare for the CEO question that is coming, if it has not already arrived: how are we managing AI spend and value?

Her sequence was the same three stages, in order:

First, get visibility into how you are consuming AI tokens, across every vector where consumption is happening. You cannot govern what you cannot see.

Second, build the guardrails into your platforms, and give the tradeoff decisions to the engineers and business partners who are closest to the work.

Third, remember that the entire point is value. All of this exists to meet business outcomes, so keep the balance between speed and governance anchored to the outcome you are trying to produce.

None of that requires a specific vendor’s tool or a proprietary methodology. It requires visibility, a governance model, and a measurement framework the whole organization can trust. The through-line from cloud to AI is that the discipline travels. The unit changed from an instance-hour to a token, but the practice of tying consumption to value is the same practice, and the community already knows how to build it.

That is the opportunity in front of us. The market is not short on AI cost tools. It is short on a neutral, vendor-independent, community-built standard for what AI cost and value even mean. We have done this before. Time to do it again.

---

This article is an editorial synthesis of a live fireside conversation between J.R. Storment and Sonali Niswander from FinOps X 2026. The framing and interpretation are J.R.’s. Sonali’s comments reflect MetLife’s approach and are presented as an illustrative enterprise example, not as a MetLife- or Tokenomics Foundation-endorsed standard. An earlier version of this article was originally published by J.R. Storment on LinkedIn.