CFOs Are Counting Every Token: Tokenmaxxing Gives Way to Value-Maxing

On Friday, 9 October 2026, Bloomberg Tech took up a topic that is coming up more and more often in corporate planning meetings: finance chiefs are recounting every token. The era of "tokenmaxxing," when employees were pushed to use AI as much as possible and rising consumption was passed off as progress, is ending. "Value-maxing" is taking its place: the question is no longer how many tokens were burned, but what the company got in return. AI spending still keeps CFOs awake at night. Host Ed Ludlow discussed this with Nina Trentmann, author of the CFO Briefing newsletter.

The show's announcement gives no specific figures or company names, so below we rely only on the stated gist and walk through the logic of the shift.

From tokenmaxxing to value-maxing

The first wave of enterprise generative AI adoption rested on a simple idea: the more people work with models, the faster the company learns. That produced internal activity leaderboards, per-employee quotas, and rewards for heavy usage. Tokens became an engagement metric, even though by nature they are a line item on a vendor's invoice.

Once the bills grew, the metric stopped working. A finance chief does not see an "innovation culture"; they see a variable expense that is hard to forecast and grows with employee enthusiasm. Hence the turn to value-maxing: every use of AI should be tied to a measurable result.

Why tokens are hard to budget

Classic SaaS is straightforward: a price per seat, a fixed contract, a predictable line in the budget. With AI services billed per token, cost depends on prompt length, the model chosen, the number of calls, and how employees and agents chain those calls together. One long context or one autonomous agent running a model in circles can eat through a budget in days.

As a result, finance teams face several problems at once:

• there is no familiar unit of account: a token is not a task, and a task is not a result;
• costs are scattered across departments, contractors, and employees' "shadow" subscriptions;
• different models cost different amounts, and the choice between them is often made by someone other than the person paying;
• the benefit of AI is often expressed in employee time, which is hard to convert into money.

What metrics CFOs are starting to demand

The logic of value-maxing points to a set of indicators that finance teams typically ask AI project owners for. This is our interpretation, not a retelling of data from the broadcast:

• cost per completed task instead of cost per token;
• the share of requests that produced a result that was actually used, rather than a discarded draft;
• AI spending broken down by department and product;
• the ratio of expensive to cheap models in the workload;
• comparison with the baseline cost of the same work done without AI.

What finance and product teams should do

The first step is mundane: consolidate all AI spending into a single picture. While some of the spend runs through corporate APIs, some through personal subscriptions, and some inside cloud contracts, it cannot be managed.

The second step: assign an owner to each major use case and require a business metric from them. Examples include ticket handling time, the number of tickets closed, or the speed of report preparation. If there is no metric, that is a reason to cut limits, not expand them.

The third step: route workloads across models. Hand routine tasks to cheaper options, and keep expensive models for cases where quality affects the outcome. The fourth: set limits and alerts for agentic scenarios in which a model calls itself.

The trend is clear: AI in companies is moving from the "experiments" category to the "operating expenses" category, which means it faces the same requirements as any other budget line. The winners will be those who can explain exactly what they are paying for.