How much does an AI agent cost to run beyond tokens?

The use of artificial intelligence agents is advancing in companies. What used to be a one-off interaction with an AI model is starting to become processes in which agents query information, use tools, make decisions, and carry out tasks with different levels of autonomy.

This evolution raises an important question for managers: how much does it actually cost to keep these agents running?

According to Gartner projections published in September 2026, global spending on artificial intelligence is expected to reach US$2.67 trillion in 2026, growth of 49.5% over the previous year. Within that market, spending on AI agents and assistants is expected to reach US$29.2 billion in the same year [1].

The growth in investment increases the need to understand not only how much a company spends on artificial intelligence, but how much each agent in operation costs and what result it delivers.

The answer is not only in the price of tokens. An agent may make several calls to models, query corporate systems, repeat steps, and involve other agents before completing a request. The effective cost of that task can be much higher than the price of an isolated call to the model.

To understand this difference, companies need to look beyond the invoices from AI vendors.

The cost of an agent goes beyond token consumption

Token consumption remains one of the main references for calculating the cost of using language models. In general, vendors price differently depending on the model used, the volume of information processed, and the characteristics of each operation.

However, an AI agent does not just generate an answer. Depending on the task, it may query knowledge bases, run tools, access applications, and repeat steps to verify or correct a result.

Consider an agent responsible for analyzing a support request. To complete the work, it may query the customer’s history, check information in a corporate system, request additional data, and forward an action to another application. Each of these steps can add resource consumption.

The AWS Well-Architected Agentic AI Lens highlights that the costs of agentic systems can involve reasoning cycles, tool calls, memory operations, and coordination among agents [2].

This means that two agents using the same model can have quite different operating costs, depending on the complexity of the tasks, the number of interactions, and the efficiency of their execution.

The challenge grows when these agents start operating continuously, handling thousands of requests or taking part in processes that involve different areas of the company.

Why measure cost per completed task

Tracking token consumption is necessary, but not sufficient to assess an agent’s economic efficiency.

Imagine two agents responsible for the same activity. The first consumes fewer tokens per run, but often needs to repeat steps or request corrections. The second uses more resources initially, but completes a larger share of tasks without rework.

If the company looks only at the individual cost of each call, it may conclude that the first agent is more economical, even though the second has a lower cost per valid result.

The most relevant indicator is no longer only how much the agent consumes and now includes how much it costs to deliver a completed task with the expected quality.

In its guidance on governance and continuous cost optimization, AWS recommends relating agent consumption to indicators such as cost per decision and cost per completed task, making it possible to assess the efficiency of operations and not only the volume of resources used [3].

In practice, this assessment can combine token consumption, the cost of the tools used, the number of runs, additional attempts, and the task completion rate.

It is not about abandoning traditional metrics, but about complementing them with indicators that help better understand the efficiency of the operation.

This distinction also helps separate expense reduction from value creation, as we discussed in when the AI bill grows faster than the return and in return on AI investment. A cheaper agent is not necessarily better if it delivers inferior results, increases rework, or requires excessive supervision.

Deloitte’s State of AI in the Enterprise 2026 survey shows that taking artificial intelligence initiatives from experimentation to production remains a challenge. Only about a quarter of the organizations surveyed had put at least 40% of their AI pilot projects into production [4].

The figure does not measure agent costs directly, but it helps put into context the importance of establishing economic and operational indicators before expanding their use.

How to control costs when agents begin to scale

During a pilot project, tracking the consumption of a few agents can be relatively simple. The scenario changes when different departments begin to use agents, each with its own goals, models, tools, and execution patterns.

In that environment, the aggregate cost of the infrastructure can hide important differences. One agent may consume more resources than expected because it performs more complex tasks. Another may show waste by repeating calls, using unnecessary tools, or staying in execution loops without producing useful results.

In an analysis published in August 2026, Gartner highlights that the challenge of AI investment lies not only in the volume of resources allocated to the technology, but in the discipline used to avoid waste, improve performance, and direct spending toward measurable results [5].

This concern gains importance as agents begin to take part in corporate processes. Without visibility into individual consumption and the cost of tasks, it becomes harder to identify inefficiencies and evaluate alternatives.

AWS also recommends tracing costs across workflows that involve multiple agents. This approach makes it possible to distinguish coordination, execution, and tool usage expenses, instead of looking only at the total cost of an operation [6].

Visibility into these components helps identify which agents and processes concentrate spending, compare alternatives, and investigate deviations before they turn into relevant budget problems.

It also makes it possible to set consumption limits suited to each activity. An agent that supports a critical decision may justify a higher cost than another dedicated to simple, repetitive tasks.

Controlling costs does not necessarily mean consuming fewer resources, but using the right resources to deliver each result with quality and predictability.

That is the point at which cost management becomes part of operating agents in production, and not only of the financial analysis done at the end of the month.

From vision to practice: ARGUS

As AI agents begin to run activities across different systems and departments, tracking their behavior and consumption becomes essential to keep the operation under control.

ARGUS, Target Solutions’ Operational and Agentic Intelligence platform, uses its AgentOps capabilities to support the monitoring, observability, and governance of AI agents.

The platform makes it possible to track token consumption and the costs associated with each agent individually, offering visibility to identify usage patterns, analyze expenses, and support optimization decisions.

With the support of the AI Gateway layer, ARGUS also contributes to applying policies and controls over the use of AI resources.

The goal is to let companies expand the use of agents with greater operational and financial predictability, without losing visibility into what happens in their operation.

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] Gartner. Gartner Forecasts Worldwide AI Spending to Grow 49.5% in 2026. September 16, 2026.

[2] Amazon Web Services. AWS Well-Architected Framework: Agentic AI Lens. Technical documentation on architecture, operations, and cost optimization of agentic systems.

[3] Amazon Web Services. Agent Cost Governance and Continuous Optimization. AWS Well-Architected Agentic AI Lens, AGENTCOST07.

[4] Deloitte. State of AI in the Enterprise 2026. Survey on adoption, scale, and results of artificial intelligence in organizations.

[5] Gartner. The Real Threat to AI ROI Is Unchecked AI Spend. August 31, 2026.

[6] Amazon Web Services. Implement Distributed Cost Tracing for Multi-Agent Workflows. AWS Well-Architected Agentic AI Lens, AGENTCOST05-BP02.