Its foundations encompass disciplines
such as mathematics, logic, statistics,
and computer science, applied through
different technological techniques
and architectures.
Artificial intelligence does not
constitute a single technology,
nor is it limited to current
generative models. It encompasses
approaches based on knowledge and
rules of logical inference, as well
as machine learning, deep learning,
and generative models.
These approaches differ in how
they operate, their data requirements,
their processing capabilities,
and their conditions of application.
In November 2023,
the OECD approved
an update to its definition of
an AI system, explained in a
memorandum published in
March 2024.
The definition characterizes
these systems by their ability
to infer, from the information
they receive, how to produce
outputs such as predictions,
content, recommendations,
or decisions that can influence
physical or virtual environments.
It also recognizes varying
degrees of autonomy and
adaptability after deployment.
An essential technological
distinction is that
not every computer program
or automated process
constitutes artificial intelligence
.
Conventional systems can perform
operations through predefined
instructions and rules without
incorporating learning or inference
mechanisms characteristic of
certain AI techniques.
However, artificial intelligence
can also use formalized rules
and knowledge, so the distinction
does not depend exclusively
on the presence of machine learning.
A distinction must also be made
between an AI model
and an AI system.
A model is a computational component
that can be incorporated into
a broader application, whereas
a system encompasses the elements
necessary for its operation and use,
including software, interfaces,
processing mechanisms, and other
technological components.
This distinction makes it possible
to identify more precisely
which technology an organization
actually uses.
The capabilities of AI systems
are neither uniform nor do they,
by themselves, guarantee
correct results.
Their performance depends on
factors such as the methods used,
the quality and relevance of data,
the characteristics of the model,
operating conditions, and
the context of use.
Some systems may produce
incorrect results, exhibit bias,
or behave differently when
confronted with situations
not considered during their
development or evaluation.
Not all systems possess
the ability to continue learning
after deployment.
These foundations establish
an essential technological
distinction: artificial intelligence
does not constitute a single system,
nor does it follow a universal
operating architecture.
Its capabilities and outputs
depend on the computational
methods employed and the interaction
among its different components.
Understanding these elements
and their functions is essential
for examining how artificial
intelligence systems operate
and identifying their technical
possibilities and limitations.
Components and Operation
of Artificial Intelligence Systems:
The operation of an artificial
intelligence system depends on
the interaction among different
computational components,
whose organization varies
according to the technology
employed and the functions
the system is intended to perform.
Fundamental elements include
data, algorithms, models,
processing mechanisms, and the
technological infrastructure
that enables their operation.
These components do not constitute
a universal architecture,
nor do they necessarily participate
in the same way in every AI system.
Data provide
information that systems can use
to develop or perform specific
functions. In machine learning
systems, datasets make it possible
to identify relationships and
regularities through training
procedures.
Their quality, representativeness,
origin, and preparation influence
model behavior.
However, a distinction must be
made between the data used
to train a model and the data
it subsequently receives
during operation.
Algorithms are
defined computational procedures
for processing information
and performing operations.
In artificial intelligence,
they may be used to perform
logical inference, search
for solutions, optimize parameters,
recognize patterns, or produce
outputs.
Knowledge-based and rule-based
systems represent information
through formal structures
and use inference procedures
to derive conclusions.
Machine learning systems,
in turn, employ mathematical
and statistical methods
to fit models to data.
An AI model
is a computational representation
used to perform specific tasks.
In machine learning, its structure
and parameters enable it to process
new inputs according to relationships
identified during training.
Other approaches may use explicit
representations of knowledge
and logical relationships.
Therefore, the terms algorithm
and model are not equivalent:
an algorithm describes processing
or learning procedures, whereas
a model constitutes a representation
that those procedures can use.
Training is the
process through which a machine
learning model adjusts its
parameters using data and
optimization methods.
During development, separate
datasets may be used for
training, validation, and testing
to evaluate the model's ability
to operate on information
not used to adjust its parameters.
Not all artificial intelligence
systems require this procedure,
and an organization may use
models previously trained
by third parties without
participating in their
original development.
Inference
encompasses the procedures
through which a system produces
outputs from input information
and the mechanisms or
representations it uses.
In a previously trained
machine learning model,
this operation makes it possible
to generate predictions,
classifications, recommendations,
or content from new data.
Ordinary inference does not
necessarily involve modifying
the model's parameters
or introducing a new
learning process.
In symbolic systems, inference
may consist of applying
logical rules to
formalized knowledge.
Technological infrastructure
comprises the resources needed
to execute and connect these
components, including processing
equipment, memory, storage,
software, interfaces, and
communication mechanisms.
Depending on the architecture,
operations may be performed
on local devices, proprietary
servers, or infrastructure
operated by external providers.
Furthermore, operating a complete
system may require additional
processes for collecting
and preparing data, integrating
software, evaluating results,
and performing operational
maintenance.
These differences establish
a fundamental technological
distinction:
a model, an algorithm, and
a complete artificial intelligence
system do not necessarily
constitute the same technological
object. The effective operation
of a system depends on how
its components are organized,
which operations each performs,
and the technical conditions
under which they are used.
Understanding this structure
provides the foundation
for examining the different
ways in which artificial intelligence
can be incorporated into
an organization's processes
and applications.
Modes of Business Use and Integration
of Artificial Intelligence:
The incorporation of artificial
intelligence into business
encompasses different modes
of technological use and integration,
determined by each organization's
operational needs, the characteristics
of its information systems,
and the degree to which AI
participates in its processes.
These modes differ fundamentally
in terms of infrastructure,
access to information,
execution capabilities,
dependence on providers,
and responsibilities
for technological administration.
A first mode involves the
individual use of external
artificial intelligence tools
through commercial applications
that employees access directly
to draft documents, analyze
information, generate content,
or support specific
professional tasks.
In these cases, the company
uses the capabilities of
a technological service,
but has not necessarily integrated
its AI components into
its own information infrastructure.
The main operations are performed
on systems administered
by the provider, while the
organization retains control
over the activities it assigns
to its employees.
Nevertheless, this mode may
also involve processing
business information, meaning
that its external nature does not
eliminate the need to understand
the conditions governing
data use and processing.
A second mode consists of
using AI functions embedded
in existing business platforms
,
such as customer relationship
management systems,
administrative applications,
communication tools,
accounting software,
or information analysis solutions.
In these cases, AI may operate
as a component of software
the company already uses,
without requiring the development
or installation of its own model.
Its participation may be limited
to providing recommendations
or may extend to automated
functions within workflows.
The actual scope of integration
depends on the enabled functions,
the data it can access,
and the operations it can perform.
A third mode involves the
development or adaptation
of business applications
that incorporate AI models
through application programming
interfaces (APIs)
.
Under this architecture,
a company may develop
its own software and connect it
to models managed by
external providers.
For example, a corporate
customer service system
may receive requests,
transmit relevant information
to an AI service,
process its outputs,
and use them to classify inquiries,
prepare responses,
or update internal records.
The company controls
the programming and connections
it develops, but model processing
may remain dependent on
external infrastructure.
These applications may also
incorporate mechanisms
for retrieving business information,
such as
retrieval-augmented
generation (RAG)
.
This technique makes it possible
to retrieve relevant documents
or data from authorized sources
and provide them as context
to a model, enabling it
to generate responses related
to that information.
Its use does not necessarily
involve retraining the model
or permanently transferring
the company's knowledge to it.
Temporary access to information
during an operation must be
distinguished from training
or modifying the model's parameters.
A fourth mode consists of
deploying AI models
on infrastructure managed
by the organization itself
,
whether through local servers,
cloud computing services,
or specialized technological
environments.
The company may install
and configure previously
developed models, adapt
certain components, or,
when it has the necessary
resources, develop and
train specialized models.
This mode offers different
possibilities for configuration
and operational control,
but requires an assessment
of processing resources,
maintenance, security,
updates, and the technical
personnel needed.
Managing a model on
company-controlled infrastructure
does not necessarily imply
ownership of the rights
to its original technology,
nor does it, by itself,
guarantee independence
from third parties.
These modes may also incorporate
agent-based AI systems
capable of using software
tools to execute sequences
of operations.
Unlike an application intended
solely to generate information,
an agent-based system may connect
to databases, email systems,
administrative applications,
or business services
to retrieve records,
prepare transactions,
and execute specific actions.
Its operational capabilities
depend on the implemented
architecture, available tools,
and assigned permissions.
Autonomous execution is not
a mandatory characteristic
of all agents, and sensitive
operations can be configured
to require prior human authorization.
A distinction must also be made
between
conventional computer automation
and automation incorporating
artificial intelligence
.
A business process may execute
previously programmed instructions
without using AI models.
It may also combine conventional
mechanisms with intelligent
components that classify
information, interpret requests,
or generate results subsequently
used by other programs.
Therefore, the presence
of automation does not,
by itself, demonstrate
the existence of artificial
intelligence, just as
the use of an AI model
does not necessarily mean
that the system can execute
actions on a company's resources.
From an operational standpoint,
distinguishing among these modes
requires identifying
which components belong
to the company, which are
administered by third parties,
where data are processed,
which systems exchange information,
and which operations can be performed
without human intervention
.
It is also necessary to determine
who configures the applications,
manages permissions,
modifies connections,
and is responsible for
their technical operation.
Contracting a platform,
developing a proprietary
application, and installing
a model represent different
configurations whose scopes
must not be confused.
These modes establish
a fundamental business distinction:
using artificial intelligence
does not necessarily mean
developing a proprietary system,
administering a model,
or integrating AI capabilities
into corporate infrastructure.
Nor is there an automatic
relationship between
the complexity of an implementation
and its business value.
The appropriate mode depends
on the required functions,
available resources,
the characteristics of
business processes,
and the technical conditions
of use.
Once these forms of incorporation
have been identified,
the organization's technological
infrastructure must be examined
to determine which systems
it owns or uses,
how they relate to one another,
and the actual operating conditions
of their artificial intelligence
components.
Assessment of Technological
Infrastructure, the Need for
Artificial Intelligence,
and Corporate Controls:
The decision to incorporate
artificial intelligence
into a company must begin
with an understanding
of its operational needs
and the capabilities
of its existing
technological infrastructure.
The commercial availability
of AI tools does not,
by itself, justify their adoption.
Before recommending implementation,
it is necessary to determine
which problems the organization
seeks to solve, what its
objectives are, and whether
it already possesses adequate
technological resources
to achieve them.
The initial assessment includes
a diagnosis of business
processes and the conventional
technology currently in use
.
This examination must identify
the applications, databases,
administrative systems,
computing resources,
and external services
involved in the company's operations.
It must also establish their
capabilities, limitations,
maintenance conditions,
and potential deficiencies.
An organization may encounter
operational difficulties
that do not require
artificial intelligence,
but rather improvements
in programming,
conventional automation,
application integration,
information organization,
or employee training.
Once the needs have been identified,
a
comparative assessment
of available technological
alternatives
must be conducted.
The organization must examine
whether its problems can be resolved
by optimizing existing systems,
developing conventional software,
or incorporating new AI capabilities.
This comparison must consider
expected effectiveness,
acquisition and maintenance costs,
required human resources,
technological dependencies,
and associated risks.
The outcome may justify
the incorporation of
artificial intelligence
or conclude that a conventional
solution is more appropriate.
When the use of AI is considered
appropriate, an assessment must
determine
which technological mode
meets the identified needs
.
The company may require
an external consultation tool,
a function embedded
in an existing application,
a system developed through
API connections,
or a broader implementation
integrated into its processes.
In each case, it is necessary
to identify the data used,
the processing infrastructure,
connections between systems,
execution capabilities,
and the technical responsibilities
of both the provider
and the organization.
However, the assessment must not
be limited to projects involving
the introduction of new systems.
Many organizations
already use artificial intelligence
through commercial consultation
and productivity tools
,
such as ChatGPT, Claude,
or Copilot, provided
by the company or used
by its employees.
This practice constitutes
a form of business use of AI,
even when it does not involve
developing proprietary models
or directly integrating them
with corporate applications.
Therefore, a distinction
must be made between
the workplace use
of external tools
and the implementation
of systems connected
to an organization's
operational infrastructure.
In these cases, the assessment
must consider
what information employees enter,
how they use the results,
and which security and
authorization conditions apply
.
For example, a consultation
involving public information
does not present the same
implications as entering
confidential documents,
customer records,
or personal data
into an external platform.
Similarly, using a personal
account for work-related
activities does not automatically
make business information
the employee's property.
Account ownership,
rights over information,
and authority to access
account contents are
separate matters,
subject to the applicable
contractual and legal conditions.
The
level of technological
exposure and risk
depends primarily on
the functions performed,
the sensitivity of information,
authorized connections,
operational autonomy,
and the potential consequences
of the outputs.
A tool used to prepare
drafts subject to human review
may require different controls
from those needed for a system
that accesses financial records,
modifies files,
transmits information
to third parties,
or participates in decisions
affecting other individuals.
The assessment must also consider
potential errors,
operational limitations,
and the specific obligations
applicable to the
company's activities.
Based on these findings,
the organization must examine
whether its existing
business administration
and control mechanisms
are sufficient
.
Organizations may already
have policies governing
confidentiality,
information security,
data protection,
access authorization,
supplier contracting,
and work supervision.
Such mechanisms may be sufficient
for certain uses of AI
or may require specific adjustments,
such as establishing
authorized tools,
restricting what information
may be shared,
verifying outputs,
and defining responsibilities.
The use of artificial intelligence
does not necessarily require
the creation of an independent
administrative structure.
Conversely, when a company
incorporates AI systems
with significant operational
functions, access to sensitive
information, execution autonomy,
or participation in decisions
with substantial consequences,
it may be necessary
to establish specialized
evaluation, oversight,
and control mechanisms
.
Their scope must be determined
according to the characteristics
of each system,
the identified risks,
the effectiveness of
existing controls,
and applicable legal obligations.
The need to adopt
a formal AI governance structure
must be justified through
this assessment,
without assuming that
all organizations require
the same administrative model.
The procedure should conclude
with a
technical and operational report
proportionate to the
organization's complexity
,
identifying its needs,
existing technology,
alternatives examined,
AI systems currently used
or proposed,
their capabilities,
limitations, risks,
and available controls.
Where appropriate,
it should be supplemented
by a legal analysis
of the activities
and technologies involved.
This documentation will enable
company management to make
informed decisions regarding
whether to incorporate AI,
retain existing systems,
modify their controls,
or establish additional
oversight mechanisms.
This assessment establishes
a fundamental premise
for analyzing artificial
intelligence governance:
business modernization
does not necessarily require
incorporating AI,
nor does the use of AI tools
automatically justify
a specialized governance structure.
Decisions must be based on
demonstrable needs,
technological characteristics,
actual risks,
the sufficiency of
existing controls,
and applicable obligations.
Building on these foundations,
the next step is to examine
the nature of AI governance
within the corporate environment,
its differences from ordinary
corporate governance,
and the circumstances
that may justify its adoption
as a specific mechanism
of administration
and oversight.