For most of AI’s recent history, humans decided what needed to be done and used models to accelerate parts of the work.
That division is beginning to change.
AI systems can now write and execute code, use tools, delegate work to other agents, run experiments, evaluate results, and continue working for hours with limited human intervention. Anthropic, in a recent paper, argues that a growing share of AI development is already being delegated to AI itself.
The company is not claiming that full recursive self-improvement has arrived. But it sees a progression from chatbots that suggest code to autonomous agents capable of completing increasingly long and complex tasks.
The length of software tasks Claude can reliably complete has moved from approximately four minutes of human work in March 2024 to twelve hours in 2026. By May 2026, Claude authored more than 80% of the code merged into Anthropic’s production codebase, while the typical engineer was merging eight times more code per day than in 2024.
Research appears to be following the same path.
In one Anthropic experiment, a group of Claude-powered agents was given an open research problem. The agents proposed hypotheses, ran experiments, shared findings, and iterated with little human involvement. They recovered 97% of the available performance improvement, compared with 23% for two human researchers working for a week.
The agents also accumulated 800 hours of work and consumed approximately $18,000 in compute.
That figure points to a second-order challenge in the automation of knowledge work:
Why More Capable AI Creates Less Predictable Costs
Anthropic describes a future in which humans increasingly provide the objective while AI systems determine the method.
But that greater autonomy also makes its economics harder to control.
A chatbot typically responds to a defined request. An autonomous system can decide how to approach the task, break it into subtasks, call tools, consult other models, run experiments, evaluate its own output, and retry when something fails.
Each additional step creates another source of token and infrastructure consumption.
This means that AI costs may not grow in proportion to its capabilities. A system that is twice as capable could explore ten times as many alternatives, operate for much longer, or initiate hundreds of model calls before reaching a result.
Now the challenge is understanding the economics of an entire chain of autonomous decisions, and ensuring that the cost of reaching the outcome does not exceed the value of the outcome itself.
From the Cloud Reconsideration to the Tokenomic Reconsideration
This is not the first time organizations have discovered that pay-as-you-go technologies can create difficult economics at scale.
AI is beginning to follow a similar economic pattern to the cloud.
As Chris Reed, senior director of IT finance at Priceline, told The Wall Street Journal: “With AI, you’re putting the credit card in the hands of the end user. If you have no control over that, or if the end user is not educated enough, they’re going to run up that tab.”
We have already seen a similar maturation cycle in cloud.
During the first migration wave, organizations prioritized speed, scalability, and access to infrastructure. Over time, however, many began reconsidering whether every workload belonged in the public cloud.
As explored in Inclusion Cloud’s AXIS report, Is the All-to-the-Cloud Era Over?, this cloud reconsideration is not a rejection of cloud. It is an effort to align each workload with the architecture, operating model, risk profile, and economics that make the most sense.
AI may now be entering its own reconsideration phase.
The first stage focused on adoption: put AI in the hands of as many people as possible and identify as many use cases as possible.
The next stage will require a tokenomic reconsideration.
Organizations will need to decide which tasks justify frontier models, which can be routed to smaller systems, which should rely on deterministic software, and which may not produce enough value to justify AI at all.
Just as the cloud discussion evolved beyond “cloud or no cloud,” the AI discussion must move beyond “use AI or do not use AI.”
Now, the question is:
Which mix of models, agents, infrastructure, and human oversight delivers the best return?
Choosing a Model Is Also a Financial Decision
That question leads directly to one of the central debates in AI FinOps: choosing the right model for each task.
The largest model may deliver the strongest performance, but using it for every classification, extraction, summary, or routing decision would be unnecessarily expensive.
A smaller model may be sufficient for many of those tasks.
The opposite mistake is also possible.
A cheaper model that fails frequently may require longer prompts, more retries, additional verification, or greater human intervention. A frontier model that completes the task correctly on the first attempt could ultimately cost less.
BCG argues that organizations should optimize for the cost of a successful outcome at the required level of quality, latency, and risk, rather than simply selecting the lowest token price. Its framework also includes the human effort required to initiate, review, correct, and approve the output.
A company might therefore use:
- A lightweight model to classify an incoming document.
- A stronger model to interpret an ambiguous exception.
- Deterministic software to validate a calculation.
- A frontier model for a complex, high-value decision.
Human review when the cost of an error exceeds the savings from automation.
The key challenge for CIOs and CFOs is to find the right balance between model capability, operating cost, and business risk.
That requires looking beyond the price of an individual token. The real question is how much the complete workflow costs once retries, verification, human intervention, and the consequences of failure are included.
Tokenmaxxing Was a Bad Implementation of a Useful Idea
Most organizations can eventually determine how many tokens they consumed.
Far fewer can explain what those tokens produced.
That gap sits at the center of the tokenmaxxing debate.
The original idea was understandable: encourage employees to use AI, make experimentation visible, and measure adoption through usage dashboards. In an early market, higher token consumption could indicate that teams were testing new tools, discovering use cases, and building familiarity with the technology.
The problem began when usage became the objective rather than the signal.
Once teams or employees are ranked by token consumption, the incentive is no longer simply to find valuable applications. It is to generate more AI activity.
That can mean using AI for low-value tasks, repeating work that does not need to be repeated, or adopting personal productivity shortcuts that never improve a broader process. It can also encourage employees to consume tokens simply to appear active on a dashboard.
In that environment, a high token count might mean that an agent processed thousands of documents, resolved customer cases, accelerated software delivery, or produced a useful research finding.
It might also mean that the agent carried an unnecessarily large context, duplicated work, repeatedly called the wrong tool, or became trapped in a loop.
The consumption may look similar. The economics are not.
Tokenmaxxing was therefore not a bad idea because it gave people room to experiment. It was a bad implementation because it attached incentives to activity without defining the business outcome that activity was supposed to produce.
A better approach preserves experimentation but adds direction.
Instead of asking teams to consume more AI, organizations can give them a specific problem, a controlled token budget, and a clear way to measure success. A pilot might ask how AI can reduce document-processing time, lower support escalations, accelerate a development cycle, or improve maintenance planning.
For software development, the relevant metric may be the cost of code that passes review and reaches production, not the amount of code generated.
For customer service, it may be the cost of resolving a case without reopening or escalation.
For document processing, it may be the cost of extracting and validating a document correctly.
AI FinOps must also account for the human work that remains around the system: preparing the workflow, reviewing outputs, correcting mistakes, managing exceptions, and recovering when an autonomous action produces an undesirable result.
An agent that is inexpensive to run but requires constant supervision may be less economical than a more capable system with a higher token price.
The lesson from tokenmaxxing is not that companies should restrict experimentation. It is that freedom to experiment needs the right economic guardrails.
Otherwise, organizations may end up spending more to demonstrate AI activity than they can justify through the value created.
AI FinOps Is Neither Purely Financial nor Purely Technical
Managing AI costs requires more than setting token limits or reviewing a monthly dashboard.
The final bill is shaped by decisions made across the entire lifecycle of an AI system: which process is automated, which model is selected, how much context it receives, how the agent is designed, how many tools it can call, how often it retries, and how much human review is still required.
A poorly designed agent may take unnecessary steps, repeatedly retrieve the same information, call an expensive model for simple tasks, or continue working after it has stopped making meaningful progress.
Those are technical decisions with direct financial consequences.
At the same time, optimizing only for lower consumption can create a different problem. A cheaper architecture may produce more errors, require additional supervision, or expose the company to unacceptable operational risk.
This is particularly important in use cases where an incorrect answer can trigger financial, regulatory, safety, or customer consequences.
Before introducing a generative AI agent, organizations should therefore ask whether generative AI is actually the right technology for the problem. Some use cases may be better served by deterministic automation, traditional machine learning, predictive AI, or a combination of these approaches.
As we explored in our guide to Predictive AI versus Generative AI, the right choice depends on what the system is expected to do. Generative AI is useful when the task involves language, interpretation, synthesis, or flexible reasoning. Predictive models may be more appropriate when the objective is to forecast a defined outcome from structured historical data. Deterministic software remains preferable when the rules are stable and errors cannot be tolerated.
This is why business processes should be reviewed before they are automated with agents:
- Is this the right process to automate?
- What business outcome should improve?
- How much is that outcome worth?
- Which tasks need AI, and which do not?
- Which model should perform each step?
- What level of autonomy is appropriate?
- How will errors be detected and corrected?
- At what point should the workflow stop?
- Who owns the budget and the outcome?
Answering those questions requires collaboration across the organization.
CIOs and CTOs understand the architecture, integrations, model behavior, and operational risks. CFOs understand budgets, margins, forecasts, and return expectations. Business leaders understand whether the system is improving the process it was designed to support.
None of those functions can manage AI economics alone.
Companies are already formalizing this shared responsibility. Some are placing AI spend under existing FinOps teams, while others are creating roles that combine AI operations, engineering, governance, and cost control. CVS Health, for example, advertised an AI Ops Engineering leadership position that included GPU cost governance and cost reduction.
The role of AI FinOps is therefore broader than reducing the bill. It is to connect technical decisions with financial accountability and business outcomes before, during, and after implementation.
Token Reconsideration: The Next Five Years Will Make AI Costs Harder to See
Anthropic’s research points toward a future in which AI systems take on progressively longer and more complex tasks.
As these systems become more autonomous, their consumption will also become less visible.
The organization sees the final outcome and, eventually, the invoice. What happened between those two points can be much harder to track.
That opacity can become a serious financial problem.
Several organizations have already exhausted AI budgets much earlier than expected, while others are introducing token caps, showback systems, cost dashboards, and model-routing strategies. The concern is real: companies can accumulate millions of dollars in AI spending while still struggling to demonstrate a clear return.
Anthropic describes a future in which AI can perform more of the work involved in building software, running experiments, and improving AI systems themselves.
If that trajectory continues, AI FinOps may become one of the most important disciplines of the next five years.
Make AI Spending More Predictable
Inclusion Cloud helps organizations review AI use cases, model choices, agent architectures, governance, and consumption patterns before costs become difficult to control.
Our consultants can help you identify inefficient workflows, select the right technology for each task, establish cost and outcome metrics, and design controls that make AI spending more visible and predictable.
Book a discovery call to evaluate your current AI spending model and define the right next steps for greater control and accountability.
