How to implement artificial intelligence in your company

In the corporate world, artificial intelligence has ceased to be an aspirational concept and has become a concrete tool for efficiency and competitive advantage. At a high managerial level, there is a conviction that artificial intelligence is transforming their business and they must get on that wave. After visiting many clients in different Latin American countries, I can say that the when is now, the unanswered questions these executives have are, where and how to start? My vision is that the starting point should not be a megaproject that encompasses the entire organization, but a specific, measurable use case with a tangible impact on a business area.

The starting point

In the world of technology, everything starts with a specific use case. It's not about transforming the entire organization overnight, but about identifying a concrete situation where artificial intelligence can generate a significant and measurable improvement. For example, in a service company, implementing an AI model that automatically manages frequently asked questions can reduce response times and free up human agents for higher-value tasks. This type of initiative is limited, measurable, and with visible results in weeks, which makes it an excellent first step. From there, the organization gains confidence, experience, and metrics that allow it to scale towards more sophisticated scenarios, such as the personalization of the customer experience, demand prediction, or logistics optimization. The essential thing is to start with a success case, consolidate learning, and then gradually expand to new fronts. To achieve this, it is key to form a multidisciplinary team, considering people from technology, business, and management control, led by a Chief AI Officer (CAIO), who acts as a bridge between the business strategy and the technological implementation. In this journey, data is the most valuable input: without quality, consistency, and governance, no artificial intelligence model will be able to produce reliable results. Therefore, before choosing a use case, senior management must assess whether it has sufficient and good quality data, and whether the infrastructure and talent exist to transform it into actionable intelligence. The technological decision: LLM vs. SLM Once the use case has been chosen, the inevitable discussion turns to the technological architecture. Here, two approaches appear: Large Language Models (LLMs), capable of processing information on a large scale with sophisticated results, and Small Language Models (SLMs), which are lighter, more economical, and easier to train with specific company data. The LLMs offer power and versatility, but they bring with them higher computing costs and confidentiality concerns when they rely on external providers. In contrast, SLMs can be run in more controlled environments, even on-premises, with predictable costs and trained in more reduced contexts, but deeply relevant to the organization. The decision should not be binary. A hybrid strategy can leverage the scale of an LLM for general tasks (summaries, broad content generation) and a specialized SLM for critical functions with sensitive data (financial processes, customer information, internal operations). The Cybersecurity Angle In cybersecurity, the LLM vs. SLM discussion becomes critical. SLMs are manageable, but running them on-prem or in private clouds implies an additional responsibility: securing the entire stack that supports them. This includes protecting the hardware, the training and inference data pipelines, the exposed APIs, and the update mechanisms. In practice, this means integrating them with Zero Trust, EDR/NDR, robust IAM, network segmentation, and end-to-end encryption solutions. Without this scaffolding, an SLM can turn into an attack vector instead of a defense. The LLMs, on the other hand, can rarely be deployed in an organization's private cloud. Large providers (such as OpenAI or Anthropic, Gemini, xAI) offer them under an "as a Service" model, without access to the base model. The alternative is to use open source LLMs (LLaMA 3, Falcon, Mistral) on their own GPU clusters in private or hybrid infrastructures, with high hardware and maintenance costs. When consumed as a service in the public cloud, LLMs introduce the problem of information leakage: every time data is sent, there is a risk that sensitive information will be retained or indirectly used in future training. Even if the provider claims anonymization, compliance with strict regulations (GDPR, HIPAA, PCI-DSS) is not always guaranteed. In fact, there have already been documented incidents where employees leaked confidential customer data by using it in public LLMs.

Practical Strategy

1. SLM on-prem or private cloud: trained with internal data and protected with a robust cybersecurity stack.
2. Public LLMs as a service: restrict their use to non-critical cases, applying data sanitization policies before the query.
3. Open source LLMs in private clouds: intermediate path for regulated sectors, although with significant investment in infrastructure and talent.

In the corporate world, artificial intelligence has ceased to be an aspirational concept and has become a concrete tool for efficiency and competitive advantage. At a high managerial level, there is a conviction that artificial intelligence is transforming their business and they must get on that wave. After visiting many clients in different Latin American countries, I can say that the when is now, the unanswered questions these executives have are, where and how to start? My vision is that the starting point should not be a megaproject that encompasses the entire organization, but a specific, measurable use case with a tangible impact on a business area.

The starting point

In the world of technology, everything starts with a specific use case. It's not about transforming the entire organization overnight, but about identifying a concrete situation where artificial intelligence can generate a significant and measurable improvement. For example, in a service company, implementing an AI model that automatically manages frequently asked questions can reduce response times and free up human agents for higher-value tasks. This type of initiative is limited, measurable, and with visible results in weeks, which makes it an excellent first step. From there, the organization gains confidence, experience, and metrics that allow it to scale towards more sophisticated scenarios, such as the personalization of the customer experience, demand prediction, or logistics optimization. The essential thing is to start with a successful case, consolidate learning, and then gradually expand to new fronts. To achieve this, it is key to form a multidisciplinary team, considering people from technology, business, and management control, led by a Chief AI Officer (CAIO), who acts as a bridge between the business strategy and the technological implementation. In this journey, data is the most valuable input: without quality, consistency, and governance, no artificial intelligence model will be able to produce reliable results. Therefore, before choosing a use case, senior management must assess whether it has sufficient and good quality data, and whether the infrastructure and talent exist to transform it into actionable intelligence. The technological decision: LLM vs. SLM Once the use case is chosen, the inevitable discussion revolves around the technological architecture. Here, two approaches emerge: Large Language Models (LLMs), capable of processing information on a large scale with sophisticated results, and Small Language Models (SLMs), which are lighter, more economical, and easier to train with specific company data. LLMs offer power and versatility, but they come with higher computing costs and confidentiality concerns when relying on external providers. In contrast, SLMs can be run in more controlled environments, even on-premises, with predictable costs and trained in more reduced contexts, but deeply relevant to the organization. The decision should not be binary. A hybrid strategy can leverage the scale of an LLM for general tasks (summaries, broad content generation) and a specialized SLM for critical functions with sensitive data (financial processes, customer information, internal operations). The angle of cybersecurity In cybersecurity, the LLM vs. SLM discussion becomes critical. SLMs are manageable, but running them on-prem or in private clouds implies an additional responsibility: securing the entire stack that supports them. This includes protecting the hardware, the training and inference data pipelines, the exposed APIs, and the update mechanisms. In practice, this means integrating them with Zero Trust solutions, EDR/NDR, robust IAM, network segmentation, and end-to-end encryption. Without this scaffolding, an SLM can turn into an attack vector instead of a defense. LLMs, on the other hand, can rarely be deployed in an organization's private cloud. Large providers (such as OpenAI or Anthropic, Gemini, xAI) offer them under an "as a Service" model, without access to the base model. The alternative is to use open source LLMs (LLaMA 3, Falcon, Mistral) on your own GPU clusters in private or hybrid infrastructures, with high hardware and maintenance costs. When consumed as a service in the public cloud, LLMs introduce the problem of information leakage: every time data is sent, there is a risk that sensitive information will be retained or indirectly used in future training. Even if the provider claims anonymization, compliance with strict regulations (GDPR, HIPAA, PCI-DSS) is not always guaranteed. In fact, there have already been documented incidents where employees leaked confidential customer data by using it in public LLMs. Practical Strategy SLM on-prem or private cloud: trained with internal data and protected with a robust cybersecurity stack.
Public LLMs as a service: restrict their use to non-critical cases, applying data sanitization policies before the query.
Open source LLMs in private clouds: intermediate path for regulated sectors, although with significant investment in infrastructure and talent.