AIOps
When the AI bill grows faster than the return
As agents begin to carry out tasks, query tools, and make decisions at scale, the cost of intelligence stops being only an infrastructure matter. Governance, architecture, and context also begin to determine how much value we can extract from the resources AI consumes.
Intelligence is getting cheaper. The bill, not always.
For much of the recent evolution of artificial intelligence, we have seen a well-known technology dynamic: models became more capable, infrastructure evolved, and the unit cost of inference fell quickly. It would be natural to imagine that this combination would produce an equally predictable trajectory for companies, with more intelligence available at a progressively lower cost.
In practice, one is not necessarily leading to the other.
As AI stops being used only for one-off interactions and begins to take part in complete processes, the consumption pattern changes. A user can ask a question and get an answer after a single interaction with the model. An agent, on the other hand, may interpret a goal, break down a task, query tools, retrieve information, evaluate intermediate results, repeat a step, call another model or agent, and only then produce an answer or execute an action. When this behavior starts to happen at scale, the economics of AI begin to change.
This effect already shows up in the most recent research. In a McKinsey survey on AI FinOps, 93% of qualified respondents reported having exceeded their AI budgets, while only 20% to 25% of companies had mature AI FinOps practices [1]. The same survey indicates that AI spending increases almost fourfold as organizations move from isolated cases to broader adoption [1]. The problem, therefore, begins to go beyond the price of a given model and involves managing a new category of consumption spread across applications, processes, and business units.
McKinsey’s own experience helps size this transformation. In May 2026, the firm processed approximately five trillion tokens per month, with about 10% of users accounting for 65% of consumption [2]. It is a specific case and should not be generalized, but it illustrates how expanding access can produce demand patterns that are hard to anticipate.
With agents, that predictability can decrease even further. A study by the Stanford Digital Economy Lab analyzed agentic programming tasks run by eight frontier models and found consumption roughly a thousand times higher than in code reasoning and code chat tasks. Runs of the same task varied by up to thirty times in consumption, and using more tokens did not mean achieving greater accuracy. In many cases, the best performance appeared at intermediate levels of consumption [3].
The result needs to be interpreted within the scope of the research, which analyzed programming tasks and does not aim to represent every kind of agentic application. Even so, it reveals an important characteristic: with agents, cost is no longer tied only to the final answer and comes to reflect the whole trajectory taken to reach it.
When AI starts deciding on its own when it needs to reason, query a tool, or involve another agent, it also takes part in the decisions that determine how much intelligence will be consumed.
This characteristic tends to gain importance as autonomy increases. If today a person decides when to query a model, in an agentic operation the system itself may decide when it needs to seek more information, validate a hypothesis, or trigger another resource. Intelligence thus also becomes a resource consumed autonomously by machines.
The problem is not only how much AI costs, but how much value that consumption produces
It would be easy to conclude that the answer is simply to reduce tokens. That reading, however, may lead us to confuse efficiency with cost reduction regardless of the result. The goal should not be to consume less intelligence, but to use it where it actually produces value.
A Deloitte survey of 1,854 executives in Europe and the Middle East helps show this difference between investment and return. Among respondents whose organizations were already using agentic AI, a group that made up 57% of the sample, only 10% said they were getting significant, measurable return at that time. Half expected to reach return within three years and another third estimated a horizon between three and five years [4].
This does not mean agentic AI is failing. Technologies capable of changing processes and operating models naturally may take time to produce deeper effects. The figure shows, however, that investment growth and value capture do not necessarily happen at the same speed, as we also explored in Your company has already invested in AI. But is it getting all the return possible?.
The challenge is not simply to spend less on AI. It is to prevent the consumption of intelligence from growing without the organization being able to relate it to the value produced.
That gap is where the economic governance of AI gains importance. Cloud computing already led many organizations to develop FinOps practices to understand who consumes resources, how much that consumption costs, and what result is being produced. AI adds a new layer because model choice, context size, number of calls, additional attempts, tool use, memory, and chains of agents can lead two architectures to solve similar problems using very different amounts of resources.
The 2026 IBM Institute for Business Value survey reinforces this difficulty. Among 2,000 technology executives surveyed, 85% said they did not have complete, real-time visibility into AI spending, while 84% had not yet fully operationalized the financial management of those resources. At the same time, participants expected their organizations to be using, on average, 1,661 AI agents in 2027, 38% above the level then observed [5].
The combination deserves attention: organizations are preparing to expand the number of systems capable of consuming AI resources relatively autonomously while many are still developing mechanisms to understand that consumption.
That is why the ROI of artificial intelligence needs to be viewed beyond the number of tokens, especially when operational context protects AI investment. An agent that uses more resources to correctly resolve a critical incident can produce far more value than another extremely economical one that delivers barely useful recommendations. Likewise, consuming millions of tokens to repeat investigations or produce analyses that generate no result is still waste, even when the unit price of inference is falling.
The economic question thus comes closer to the operational one: we need to understand not only how much AI is consuming, but why it is consuming and what we are getting in return.
Context and governance also begin to be part of the economics of AI
This discussion connects with some themes I have been exploring in previous articles. When we talk about operational context, we usually think first of the quality of the decision. When we discuss governance, we think about security, limits of autonomy, permissions, and accountability. These aspects remain fundamental, but they also begin to acquire an economic dimension.
McKinsey identifies several factors that influence the cost of agentic systems, among them maintaining context throughout execution, successive refinements, different execution paths, model choice, orchestration among agents and tools, and the very way information is structured and presented to the model [6]. In another study, the firm points to mechanisms such as caching, reducing unnecessary context, selecting the right model for each task, and limits on retries, tool calls, and agent loops as ways to optimize consumption [1].
This suggests that the economic efficiency of AI will not be solved only by negotiating lower prices with model vendors.
Consider an agent responsible for investigating an operational incident. If it receives an enormous amount of unstructured data, it may need to query different sources, process long histories, test hypotheses, and repeat calls until it builds a reasonable understanding of the problem. If it receives insufficient information, it may save resources and, at the same time, miss elements needed to decide correctly. The challenge is not simply to provide more or less context, but to make available what is relevant for that decision.
This distinction will be particularly important in complex IT and Telecom environments, where an investigation may involve thousands of events, different platforms, topologies, recent changes, and dependencies between components. Indiscriminately putting all that content into a model’s context window will hardly be efficient, just as expecting each agent to rebuild these relationships alone with each new occurrence.
Context is not only what helps an agent decide better. In an agentic operation at scale, the way that context is built and made available also becomes part of the discussion about efficiency.
This does not mean that a context platform automatically reduces token consumption, nor does it allow establishing a direct relationship between operational context and a given percentage saving. The research consulted does not support that conclusion.
What it does support is that the way we structure information, design workflows, select models, and govern agent behavior affects the costs of agentic systems [1][6]. At scale, this requires visibility into which agents are consuming resources, in which processes, using which models, with how many calls, and with what results, which points to the discipline of AgentOps: who will operate AI agents in production?
The challenge is no longer simply controlling the bill and becomes relating consumption, performance, and result, making it possible to invest more where intelligence produces value and fix what only consumes resources.
The answer should not be to limit that autonomy indiscriminately, but to build the visibility, context, and governance mechanisms needed for it to grow accompanied by results. In the end, the most important question may not be how much a million tokens cost, but how much value we can produce with them.
From vision to practice: ARGUS
As AI agents begin to run activities continuously, tracking their efficiency requires understanding not only the resources consumed, but also the tasks performed, the results produced, and the context in which they acted.
This discussion connects with the proposal of ARGUS, although it is important to distinguish its different dimensions.
Through AgentOps and AI Gateway, ARGUS seeks to offer visibility into agent execution and governance mechanisms over their activity, including available information on calls, resource usage, performance, and costs in the integrated environments.
Its AIOps capabilities add another perspective: the operational context in which these activities happen. By relating information from infrastructure, applications, services, and existing tools, the platform can help better understand the circumstances in which agents make decisions and execute actions.
This combination creates conditions for a more consistent analysis of agentic operation. However, observing consumption and behavior does not, in itself, demonstrate cost reduction or increased financial return. Those results depend on metrics, goals, and assessments specific to each organization.
ARGUS should not be presented as a complete AI FinOps solution. Its contribution lies in the observability and governance of agents, complemented by operational context, offering information that can support decisions about efficiency, control, and the evolution of artificial intelligence use.
Learn about ARGUS and see how it works in practice.
About the author
Paulo Florêncio de Menezes is an entrepreneur, partner, and co-founder of companies such as Target Solutions, BiPTT, and ARGUS.
He holds a degree in Data Processing from the Federal University of Amazonas (UFAM), an MBA in Business Management from Fundação Dom Cabral (FDC), and a Professional Master’s in Economics from Fundação Getulio Vargas (FGV).
In his Master’s research, he developed the IBEOP-IA (Brazilian Index of Potential Occupational Exposure to Artificial Intelligence), investigating the potential exposure of Brazilian occupations to AI capabilities by combining occupational data, artificial intelligence patents, and natural language processing techniques.
References
[1] McKinsey & Company. The cost of intelligence: How CIOs can manage AI demand at scale. McKinsey Quarterly, July 20, 2026.
[2] McKinsey & Company. Is that AI agent worth it? Agentic economics and the modern operating model. McKinsey Quarterly, July 13, 2026.
[3] Bai, Longju et al. How Do AI Agents Spend Your Money? Analyzing and Predicting Token Consumption in Agentic Coding Tasks. Stanford Digital Economy Lab, Working Paper, April 14, 2026.
[4] Deloitte. AI ROI: The paradox of rising investment and elusive returns. Deloitte Global, October 22, 2025. Survey of 1,854 executives in Europe and the Middle East.
[5] IBM Institute for Business Value. The 2026 Tech Leader Study: Redefining the tech leader’s mandate: Building the IT foundation for agentic AI at scale. June 2026. Global survey of 2,000 C-level technology executives.
[6] McKinsey & Company. Cost versus value: Managing agentic AI system performance. McKinsey Quarterly, 2026.