AI has already demonstrated its ability to automate processes, improve productivity and accelerate decision-making in companies. However, as companies try to scale AI and extend its use on a large scale, a new question arises: is it profitable to do so? And the main obstacle to deploying AI on a massive scale is no longer technological, but economic.
The 2025 State of AI Cost Management report, prepared by Mavvrik and Benchmarkit, indicates that 80% of companies recognize deviations greater than 25% in their AI cost forecasts, while almost 1 in 4 underestimated spending by more than 50%. Furthermore, according to a study by the MIT NANDA initiative, 95% of generative AI pilots fail to translate into tangible and quantifiable business benefits.
Behind this difficulty in scaling AI profitably, the three mistakes set out below are often repeated.
First mistake: using more powerful models than necessary
Many organizations use the most advanced models for virtually any task, even when much of the work could be solved by lighter, more efficient systems. The result is oversized automation that ends up directly affecting the profitability of the projects. As the volume of use increases, infrastructure costs grow and, in some cases, do so faster than the value generated by the technology itself. For years, we have been obsessed with having increasingly powerful models, but the real challenge is to use them intelligently.
Second mistake: trying to scale pilots designed for controlled environments
Many AI projects show promising results during initial testing, but encounter difficulties when it comes time to deploy them across the organization. The reason is that a pilot operates in a controlled environment, with few users and seemingly affordable costs. However, when AI begins to be widely used, there are exceptions, retries, multiple steps of reasoning, and a volume of operations much higher than expected.
Additionally, unlike traditional software, the costs of generative AI are variable. The same process can require very different resources depending on the complexity of each query, making financial planning and estimating the return on investment difficult.
As a consequence, some organizations end up limiting access to certain tools to contain spending, reducing precisely the transformative impact they sought to achieve.
Third mistake: building entire processes around a single model
A large part of the market continues to focus its efforts on choosing better models or looking for cheaper alternatives. However, the true optimization potential lies in the way work is designed and executed.
In many organizations, entire processes are delegated to a single AI model. This forces the use of advanced technology throughout the entire process, even in repetitive or simple tasks that could be solved by more efficient and economical systems. As a result, companies end up paying for complex reasoning capabilities at stages where they do not really provide differential value, causing costs to increase as the volume of use grows.
Optimize processes before models as a great solution
The key to scaling AI profitably is to stop thinking only about models and start designing architectures focused on the efficiency of each process. This involves dividing workflows into specific tasks and assigning the most appropriate technology to each. In this way, the most advanced models are reserved for those activities where they really add value, while the rest of the process can be executed using smaller specialized models, traditional automations or code-based systems, with significantly lower costs.
In many organizations, entire processes are delegated to a single AI model
When each task uses only the resources it needs, costs stop growing at the same rate as the volume of use, making it possible to scale AI sustainably. Additionally, this approach provides greater flexibility: by decoupling processes from a specific model, companies can incorporate new technologies, combine different models or replace existing solutions without having to completely redesign their operations.
Analysis by David Villalón, CEO and co-founder of Maisa
