The real cost of AI in IT companies

Column

Artificial Intelligence

May 22, 2026

The real cost of AI in IT companies

The next cycle of technology in organizations should be less marked by the euphoria of replacement and more by the discipline of operational sustainability

The accelerated incorporation of artificial intelligence in technology companies has been presented as an almost inevitable path to increase productivity, reduce costs, and replace part of human labor in technical, administrative, and often even creative tasks, depending on the situation.

This narrative has guided internal reorganization business decisions, team reductions, creation of products based on automation, and adoption of generative tools on a large scale in recent years. However, a central economic tension is becoming more evident: artificial intelligence used at scale is not free, it is not unlimited, and it depends on increasingly expensive computational infrastructure.

This point appears concretely in the change of billing for GitHub Copilot. In April 2026, GitHub announced that all Copilot plans would switch to usage-based billing starting in June 2026, with consumption calculated by AI credits associated with input, output, and cache tokens. The company itself justified the change by stating that the agent’s use, with long sessions and multiple programming steps, had significantly increased computing and inference demands, making the previous model of premium requests unsustainable (Rodriguez, 2026).

The movement in question is important because it reveals a phase change. During the initial period of generative AI expansion, many tools were offered in simplified commercial models, with monthly subscriptions, unclear limits, and strong implicit usage subsidies.

As users began to delegate longer, more complex tasks closer to real professional work, the cost ceased to be marginal. AI stopped operating solely as an individual productivity complement and began to represent a line of variable expense, dependent on usage intensity, context size, chosen model, and the complexity of the requested operations.

Operating cost

The first sign of this transformation is in the review of the pricing models of AI-based development tools themselves. Cursor, one of the platforms most associated with the use of AI in programming, changed its “Pro” plan in 2025 to include a monthly quota of US$ 20 in the use of frontier models, calculated based on API pricing. The company explained that newer models can spend more tokens per request on longer horizon tasks and that more difficult requests can cost an order of magnitude more than simple requests (Cursor, 2025).

This data is decisive for IT companies, because it dismantles the idea that AI has a fixed and predictable cost in any scenario. The more the organization tries to use AI to solve complex problems, such as system refactoring, analysis of large repositories, test generation, code review, and automation of complete workflows, the greater the computational consumption tends to be. The cost does not grow only with the number of users, but with the depth of the tasks delegated to the model. Thus, automation ceases to be just a promise of savings and begins to require technical accounting.

Anthropic’s documentation for Claude Code follows the same direction. According to the company, Claude Code charges based on token consumption per API, and costs per developer vary according to the model used, codebase size, and usage patterns, such as multiple instances or automation. In corporate deployments, Anthropic reports an approximate average cost of US$13 per developer per active day and between US$150 and US$250 per developer per month (Anthropic, 2026).

These values do not mean that the use of AI is unfeasible. They indicate, however, that its adoption needs to be treated as a recurring operational cost, and not as a simple software subscription. For companies that reduced teams believing that generative tools would absorb a substantial part of the technical work, the bill may become more complex. The savings obtained with fewer professionals may be partially offset by expenses for inference, human review, rework, integration, security, and governance.

Infrastructure and cost pass-through

The problem is not limited to the tools used at the edge. It begins in the infrastructure of central AI companies. In February 2026, Reuters reported that OpenAI projected approximately $600 billion in total computing spending by 2030, according to a source familiar with the matter.

The same report indicated that the company’s inference costs would have quadrupled in 2025, with a drop in adjusted gross margin from 40% in 2024 to 33% in 2025, according to a report attributed to The Information (Reuters, 2026).

Although these numbers should be interpreted as estimates and information reported by sources, they point to a structural aspect: large-scale generative AI requires intensive capital, chips, energy, data centers, high-capacity networks, and long-term infrastructure contracts.

The cost does not disappear. At some point, it needs to be absorbed by suppliers, investors, or customers. When the subsidy cycle diminishes, the tendency is for part of this cost to be transferred to companies that depend on AI platforms to operate their own products and services.

In this context, the risk is not the disappearance of artificial intelligence, but the weakening of business models built on the hypothesis of cheap, abundant AI capable of replacing professionals on a large scale. Companies that sold solutions with strong dependence on calls to models external may face margin compression, should the cost per use increase.

Similarly, organizations that have internalized the idea that they could operate technical areas with very small teams may find that the payroll reduction has been offset by new invisible costs, distributed among platforms, APIs, credits, monitoring, review, and correction of automated outputs.

Productivity and substitution

The second point of tension lies in the difference between apparent automation and effective productivity. Many IT companies have adopted AI tools based on the idea that writing code faster equates to delivering better, cheaper software with less dependence on specialized professionals.

This equivalence is limited. Software development involves architecture, requirements understanding, maintenance, testing, documentation, security, integration, user communication, and technical responsibility. AI can accelerate parts of this process, but it can also introduce rework, inconsistencies, fragile solutions, and increased technical debt.

A study by Becker, Rush, Barnes, and Rein (2025), published in a preliminary version on arXiv, evaluated experienced developers working on their own open-source repositories. The authors analyzed 246 tasks and found that when the use of AI tools was permitted, developers took 19% more time to complete the activities. The result contradicted the expectation of the participants themselves, who before the study estimated a 24% reduction in time with AI (Becker et al., 2025).

This result should not be generalized to all programming scenarios. It refers to a specific context, with experienced developers, mature codebases, and tools available in early 2025. Nevertheless, the study is relevant for questioning the assumption of automatic and universal productivity gains. In real systems, the time saved in code generation may be consumed by reading, validating, adjusting, testing, correcting, and discarding inadequate responses. For companies that reduced teams by betting on direct replacement, this difference between apparent production and reliable delivery can become an operational problem.

The likely consequence is not the elimination of AI in technology companies, but the recomposition of the role of human labor. Technical professionals tend to continue using AI, but in a more selective, supervised, and economically controlled way.

The promise of replacing entire teams with tools tends to find limits when the inference cost increases, when critical systems require constant review, and when the delivery of value depends on technical responsibility. In this scenario, companies that have confused automated artifact generation with organizational capacity may need to rebuild teams, rehire professionals, or create roles focused on the supervision, validation, and integration of AI-based systems.

Return on Investment

The third element is the difficulty in demonstrating consistent economic return. In 2024, Gartner estimated that at least 30% of generative AI projects would be abandoned after the proof of concept by the end of 2025, due to poor data quality, inadequate risk controls, rising costs, or unclear business value. The consultancy also noted that organizations faced challenges in justifying substantial investments in generative AI when productivity gains did not directly translate into financial benefit (Gartner, 2024).

In 2025, Gartner projected that over 40% of agent AI projects would be canceled by the end of 2027, again due to rising costs, undefined business value, or insufficient risk controls. The analysis also warned that many projects were still in the experimental stage and were often driven by market enthusiasm, without clarity on the cost and complexity of large-scale deployment (Gartner, 2025).

The report “The GenAI Divide: State of AI in Business 2025”, by MIT NANDA — based on interviews with leaders from 52 organizations in various locations around the world, collected at four major industry conferences —, reinforces this diagnosis. According to Challapally, Pease, Raskar, and Chari (2025), despite estimated business investments between US$ 30 billion and US$ 40 billion in generative AI, 95% of the analyzed organizations have not achieved measurable returns.

The report also pointed out that tools like ChatGPT and Copilot are widely explored and deployed, but tend to improve individual productivity more than aggregate financial performance (Challapally et al., 2025). This point is central for IT companies that have started building products and processes on AI. The question is no longer whether the tool impresses in demonstrations, but whether it sustains margin, quality, recurrence, and scale.

During the excitement phase, AI allows for rapid prototyping, reduced initial development time, and low-friction creation of texts, code, interfaces, and analyses. However, when the product needs to operate with real customers, fulfill contracts, handle exceptions, maintain security, and preserve quality, the initial savings can turn into ongoing expenses.

Economic governance

In the business context, the adoption of AI now requires economic governance as rigorous as technical governance. It is not enough to ask if a task can be automated. It is necessary to evaluate how much it costs to automate it, what the cost of human review is, what the risk of error is, what the dependence on external suppliers is, what the frequency of use is, and what the impact on the organization’s internal capacity is. Companies that dispense with technical knowledge in favor of automation may lose precisely the competence needed to evaluate, correct, and integrate the systems they intend to use.

The most plausible trend is a tougher selection of use cases. Repetitive, delimited, and low-risk activities should continue to be automated. Complex, ambiguous, and critical processes tend to remain dependent on qualified professionals, even if supported by AI.

The strategic error of IT companies will not be the use of artificial intelligence, but structuring costs, teams, and products based on the assumption that AI will replace specialized work without loss of quality, without increased risk, and without economic pressure.

Given this set of signals, the next cycle of AI in technology companies tends to be less marked by the euphoria of substitution and more by the discipline of operational sustainability. Generative tools will remain relevant, but the idea of unlimited automation tends to give way to models of controlled use, budget per team, consumption limits, return analysis, and intensive human review. AI should not disappear from IT companies. What tends to collapse is the fantasy that it will allow operating complex companies with few professionals, low marginal cost, and automatically increasing productivity.

In conclusion, the central point is not to deny the importance of AI, but to understand that its adoption enters an economically more realistic phase. As vendors pass on costs, adjust plans, limit intensive uses, and organize pay-per-use billing, companies that have built their strategy on cheap AI may face cost recomposition and the need for technical rehiring.

The near future tends to favor organizations capable of combining automation with human competence, and not those that have replaced professional structure with uncritical dependence on external platforms. In this sense, AI does not eliminate the need for qualified people. On the contrary, the more expensive, complex, and strategic it becomes, the greater the need for professionals capable of deciding when, how, and to what extent it should be used.

To access the references of this text click here.

Who wrote this column

Maurício Acconcia Dias

Possui graduação em Ciência da Computação pela Universidade Federal de Lavras, mestrado e doutorado em Ciências da Computação e Matemática Computacional pela Universidade de São Paulo e MBA em Data Science Analytics pela USP/Esalq. Atua com desenvolvimento de hardware para sistemas inteligentes aplicados à robótica. É consultor em Data Science & Analytics e Desenvolvimento de sistemas embarcados e orientador do MBA USP/Esalq.

You may also like

November 16, 2026

Exploratory modeling of the effect of preemptive cytomegalovirus treatment in AIDS patients

Cytomegalovirus (CMV) is highly prevalent and causes severe disease in immunocompromised individuals, such as HIV patients. Given the scarcity of robust data on preemptive treatment for this infection, an exploratory model was developed. The objective was to identify potential subgroups that would benefit from this approach. A retrospective observational study was conducted, analyzing 176 medical records of patients admitted to a tertiary hospital between 2017 and 2022. After applying inclusion and exclusion criteria, 116 patients were divided into a preemptive treatment group (19) and a control group (97). Machine learning techniques, such as LinearDML and CausalForestDML, and variable analysis by Feature Importance and SHAP were used. Variable analysis indicated that high CMV viremia was associated with death, with an Odds Ratio of 1.85 (p = 0.004). The model estimated an increase in overall mortality associated with preemptive treatment (ATE of 48%, p = 0.0180), possibly reflecting the greater baseline severity of the treated patients. However, heterogeneity analysis identified that older patients showed a relatively more favorable estimated effect. Limitations included the small sample size and the retrospective and observational nature. Nevertheless, with an expanded evaluation group, the model could serve as a basis for building clinical decision support tools.

Keywords: Machine learning; Double machine learning; Causal inference; Mortality.

November 16, 2026

Regression models for personal hygiene and personal care product expenses in the Southeast region

Regression models were evaluated to correlate expenses with personal hygiene and care products of the population in the Southeast Region of Brazil. Data from the 2017-2018 Family Budgets Survey (POF), conducted by the Brazilian Institute of Geography and Statistics, were used. The models were estimated by ordinary least squares and verified for the assumptions of normality of error terms, multicollinearity, and heteroscedasticity. Linear regression, nonlinear regression with Box-Cox transformation, and linear regression with dummy variables were evaluated. Coefficients statistically different from zero were obtained for all three models, confirmed by the p-values of the F and t tests, and the adherence of residuals to normality was graphically assessed. However, heteroscedasticity problems were observed in all cases, indicating the omission of relevant variables from the model. The low R² values obtained, between 2.19% and 5.56%, suggested that the dependent variable was poorly explained by the models. Despite this, a growing trend in expenses for personal hygiene and care as a function of disposable income per capita was observed. For future studies, it is recommended to evaluate other explanatory variables to improve the model’s adjustment coefficient and eliminate heteroscedasticity.

Keywords: Family expenses; Ordinary least squares; Per capita income; Box-Cox transformation; Dummy variables.

November 16, 2026

Cognitive Accessibility and ASD: Design Guidelines for Augmentative and Alternative Communication Applications

Autism Spectrum Disorder (ASD) imposes specific challenges on the design of digital interfaces, affecting cognitive processing, sensory regulation, and functional communication. In this context, Augmentative and Alternative Communication (AAC) is an essential resource to support individuals with oral language limitations. This study aimed to analyze the usability and accessibility of AAC applications for the neurodivergent public. The research was characterized as descriptive, with a quantitative approach, and was based on the systematic observation of the interfaces of the Brazilian applications Livox, Matraquinha, and Expressia. A structured checklist, developed from the convergence of Human-Computer Interaction (HCI) guidelines, Applied Behavior Analysis (ABA), and the GAIA Guide, was used as a data collection instrument. The results revealed that Livox and Expressia showed greater technical compliance and learning support, while Matraquinha stood out for its semantic literalness, but with limitations in customization and architecture. A generalized critical gap was identified in the proactive detection of frustration, highlighting opportunities for strategic innovation for the digital business market and for the advancement of neuroinclusive design.

Keywords: Neuroinclusive Design; Digital Inclusion; Strategic Innovation; Assistive Technology; Usability.

November 16, 2026

Multidimensional Poverty and Bolsa Família Program: challenges and perspectives of inclusion through public policies

The concept of poverty has evolved over the 20th and 21st centuries, incorporating a more inclusive perspective that transcends monetary income. The Multidimensional Poverty Index (MPI) has emerged as a complementary tool to characterize poverty and guide programs like Bolsa Família, aiming to improve their effectiveness and targeting. The study aimed to create a national MPI using data from the Continuous National Household Sample Survey (PNADC) of 2024, seeking to identify the correlation between receiving the benefit and multidimensional poverty, as well as to detect underserved populations. The Alkire-Foster methodology was used to calculate the MPI, decomposing the results into groups of beneficiaries and non-beneficiaries of PBF. Subsequently, a Generalized Linear Model with a logit function was applied to analyze the correlations. The results revealed that 4.84% of the Brazilian population lived in multidimensional poverty, with an incidence three times higher among Bolsa Família beneficiaries. Two profiles of multidimensional poverty and two potential blind spots in PBF coverage were identified, including 9.0% of non-beneficiary multidimensional poor in extreme monetary poverty and 25.2% in low-income vulnerability. Being a PBF beneficiary was associated with a two-fold higher chance of being classified as multidimensionally poor. It was concluded that the MPI is a valuable tool for designing public policies, increasing targeting capacity and demanding actions to address structural problems that perpetuate poverty.

Keywords: Alkire-Foster; MPI; Public policies.

November 16, 2026

The influence of People Management on the Health and Well-being of collaborators

In a dynamic organizational scenario, understanding the influence of People Management practices on employee health and well-being has become essential. The study aimed to analyze how People Management practices influence employees’ perception of health, well-being, and quality of life in various organizations. The methodology adopted was descriptive, with a quantitative approach, and employed the survey technique. Data were collected through a structured questionnaire with 15 sociodemographic questions and Likert scale items, disseminated for 40 days on social media. The sample consisted of employees from different companies, selected for convenience, with voluntary and anonymous participation. The results revealed that people management practices impacted motivation, engagement, and the perception of quality of work life, but were not consistently perceived by all employees. It was observed that organizational actions focused on health and well-being were recognized in a moderate and non-equivalent manner, with the presence of neutral and dissenting perceptions. It was concluded that, although People Management practices have strategic potential, their effectiveness depends on greater structuring, continuity, and communication. The study contributed to highlighting gaps between the adopted practices and employee perception, providing subsidies for the improvement of organizational strategies and for the production of academic knowledge in the area.

Keywords: Organizational well-being; Organizational climate; Engagement; Motivation; Quality of work life.

November 16, 2026

Optimization of Delivery Logistics Network through Mathematical Modeling and Clustering

Efficiency in distribution logistics constitutes an essential competitive advantage for the reduction of operational costs and the financial sustainability of contemporary organizations. The general objective of this work was to optimize the delivery network of a company, aiming at cost reduction through the decrease of mileage traveled. The methodology employed was based on the principles of Operational Research and was structured in three sequential phases: geographical allocation of clients to depots respecting Federal Unit restrictions, clustering of stores using the K-Means algorithm, and routing of the heterogeneous fleet using the mathematical formulation of the Vehicle Routing Problem. The main results obtained demonstrated a reduction in the total distance traveled, which represented an absolute saving of 5,116 km and a percentage gain of 3.24%. The model validated itself as a robust tool, capable of optimizing vehicle occupancy and eliminating subjectivity in logistical planning. It was concluded that the use of advanced mathematical modeling allowed operational restrictions to be converted into a measurable competitive advantage. The solution validated the viability of data-based decision support systems for the automation and optimization of large-scale logistical flows.

Keywords: Clustering; Logistics optimization; Operational Research; Cost Reduction; Vehicle routing.