A clinician in Bogotá reviews a diagnostic image supported by software that identifies features requiring closer examination. In another part of Colombia, a rural health service considers whether an automated system could help organize referrals to specialists. Meanwhile, a residential care organization explores whether artificial intelligence could reduce the administrative time employees spend preparing records and reports.
These situations illustrate different possibilities for artificial intelligence in Colombian health and care services. They also raise different questions about professional responsibility, data protection, technical reliability and the consequences of incorrect recommendations.
Colombia's wider institutional context is examined through the Colombia Aging, Long-Term Care and Community Support Knowledge Hub, which considers the relationship between national policy, territorial delivery, workforce capacity, quality governance and the needs of older people and persons with disabilities.
Artificial intelligence offers opportunities to support this system, but its application is not uniform. Some technologies are being researched or evaluated in clinical settings, while others are commercially available for administrative tasks. Neither situation establishes that AI has been adopted throughout Colombia's health system or that its effectiveness has been demonstrated across all populations.
The central policy challenge is to distinguish useful innovation from technology that creates additional risk, expense or complexity.
For Colombia, this requires particular attention to the responsibilities of Entidades Promotoras de Salud (EPS), Instituciones Prestadoras de Servicios de Salud (IPS), national and territorial authorities, technology suppliers and professionals delivering care.
It also requires recognition that residential services, community support organizations and unpaid family caregivers operate within different arrangements from regulated clinical institutions.
The strongest opportunity lies in using AI to improve decisions and service coordination while preserving human accountability, individual rights and equitable access.
Colombia's health system and the institutional context for AI
Colombia's Sistema General de Seguridad Social en Salud (SGSSS) provides the principal institutional framework through which much of the country's health care is organized.
EPS undertake responsibilities relating to the management of health coverage and access within the applicable system arrangements. IPS deliver health services through hospitals, clinics, outpatient facilities and other authorized service settings.
The Ministerio de Salud y Protección Social establishes national health policy and regulatory requirements. The Superintendencia Nacional de Salud exercises oversight functions within its statutory competence, while departmental, district and municipal health authorities undertake responsibilities assigned to their respective levels.
Artificial intelligence may affect each institution differently.
An IPS might consider an AI-assisted diagnostic application or software supporting clinical documentation. An EPS might examine technology intended to identify administrative inconsistencies or improve the organization of service requests. A territorial health authority could consider analytical tools for understanding population needs, subject to its legal responsibilities and information governance arrangements.
However, the availability of these technologies does not automatically establish their suitability, authorization or effectiveness.
Different applications involve different risks. A system that helps categorize administrative correspondence is not equivalent to software that influences a clinical diagnosis or treatment recommendation.
The distinction is important because the consequences of an error differ substantially.
Beyond the SGSSS, Colombia's care environment includes residential institutions for older people, Centros Vida, disability-related community programs, privately arranged support and extensive unpaid family caregiving.
These arrangements do not constitute one integrated national long-term care insurance system, and they do not all share the same regulatory requirements or funding structures.
An AI application designed for a hospital cannot therefore be assumed appropriate for a residential support organization simply because both work with older people.
Organizations need to examine the intended use, the people affected, the information involved and the actual responsibilities of the institution considering adoption.
This connects Colombia's emerging experience with wider questions about AI and automation in care services.
Where artificial intelligence could provide practical value
Artificial intelligence is not one technology. The term encompasses systems capable of identifying patterns, classifying information, generating content, supporting predictions or assisting decisions through computational methods.
In health and care services, the potential applications range from relatively contained administrative functions to complex clinical systems.
Administrative uses may include organizing documents, identifying incomplete information, preparing draft summaries or assisting with scheduling. These functions can reduce repetitive work where outputs are reliable and appropriate human checks are maintained.
Clinical applications may include image analysis, support for interpreting diagnostic information, identification of potential deterioration or assistance with clinical documentation.
Such applications require substantially greater scrutiny because their outputs may influence diagnosis, treatment or patient safety.
Population-level analysis offers another possibility. Appropriately governed analytical systems may help institutions examine service demand, identify patterns of delayed access or understand where additional capacity could be required.
However, predictive outputs are influenced by the data used to develop them. A model trained primarily on urban populations may perform differently when applied in remote communities or among people whose circumstances are poorly represented in the original data.
The operational question is therefore not whether an application contains AI, but whether it performs a clearly defined function better than an appropriate alternative.
Before adoption, an institution should understand the task being addressed, the evidence supporting the technology and the consequences of incorrect output.
It should also consider whether existing organizational problems could be resolved through simpler changes to procedures, staffing, communication or information systems.
AI may help employees work more effectively, but it cannot compensate for absent clinical capacity, unclear responsibilities or incomplete service pathways.
Clinical decision support and the limits of automated recommendations
Clinical decision support is among the most consequential potential uses of AI.
Systems may assist professionals by identifying patterns in diagnostic images, highlighting potentially relevant information or supporting the assessment of clinical risk.
These functions can be valuable when the technology is appropriately validated, integrated into professional workflows and used within its intended purpose.
However, AI-generated recommendations should not be treated as independent clinical authority.
Clinical responsibility remains governed by the applicable professional, institutional and regulatory framework. The use of software does not remove the need for qualified professionals to exercise judgment.
In Colombia, a clinical institution considering such technology needs to establish whether the application falls within the regulatory arrangements applicable to medical devices or other regulated health technologies.
Depending on the product's intended purpose and classification, the Instituto Nacional de Vigilancia de Medicamentos y Alimentos (INVIMA) may have a relevant role under the applicable medical device framework.
Not every AI application is automatically a regulated medical device, and not every clinical software product follows identical authorization requirements. The classification depends on its characteristics, claims and intended use.
Organizations should therefore obtain appropriate regulatory clarification rather than assuming that a supplier's description of a product as innovative or clinically validated establishes authorization for use in Colombia.
Validation also needs to address the population in which the technology will operate.
A diagnostic model may demonstrate strong performance in a controlled study but encounter different image quality, disease prevalence or patient characteristics in routine practice.
Clinical teams need to understand the model's limitations, including circumstances in which its recommendations should not be relied upon.
For higher-risk applications, governance should consider:
- The clinical purpose and intended patient population.
- The evidence supporting safety, accuracy and effectiveness.
- The applicable regulatory status and institutional authorization.
- The professional responsible for interpreting and acting upon outputs.
- The arrangements for identifying errors, adverse events and performance deterioration.
- The ability to suspend use when unacceptable risks emerge.
These considerations reflect the wider principles of clinical governance and professional accountability.
AI should strengthen the quality of professional decisions rather than obscure how those decisions were reached.
Illustrative scenario: AI-assisted diagnostic review in Bogotá
Consider a hypothetical hospital in Bogotá evaluating software intended to assist radiologists with the review of diagnostic images.
The hospital's leadership believes the application may help identify images requiring closer attention, particularly during periods of high demand.
Before implementation, the clinical team examines the supplier's evidence, the technology's intended purpose and any applicable regulatory requirements.
The hospital also considers whether the software has been evaluated using imaging equipment and patient populations sufficiently comparable to those encountered locally.
During a controlled evaluation, radiologists continue to interpret images independently while the institution assesses the software's performance.
The team discovers that some images are incorrectly prioritized. Although these errors do not necessarily make the technology unsuitable, they demonstrate why its outputs cannot replace professional interpretation.
The hospital establishes that radiologists remain responsible for clinical reporting and that the system's recommendations must not be presented as definitive diagnoses.
Patients' clinical information is processed according to the applicable confidentiality and data protection requirements. The institution also examines where information is stored and whether the supplier has access to identifiable records.
Following the evaluation, leadership reviews whether the application improves reporting times without creating unacceptable risks or additional workload.
The decision is based on observed performance rather than the supplier's promotional claims.
If the hospital proceeds, it establishes arrangements for monitoring errors, reviewing unexpected outputs and suspending the application where necessary.
The scenario illustrates that responsible AI adoption is a clinical governance decision supported by technical evaluation, not simply an information technology purchase.
For patients, the relevant outcome is more reliable and timely diagnostic care, with qualified professionals retaining responsibility for interpretation and treatment decisions.
Colombia's AI policy direction and the distinction between policy and law
Colombia has developed national policy initiatives addressing artificial intelligence, digital transformation and the responsible use of emerging technologies. These initiatives provide an important context for health and care organizations, but their legal status must be understood accurately.
CONPES 3975 of 2019 established a national policy direction concerning digital transformation and artificial intelligence. CONPES 4144 of 2025 subsequently set out Colombia's National Artificial Intelligence Policy, including objectives relating to governance, capabilities, infrastructure, research, innovation and the responsible adoption of AI.
CONPES documents are national policy instruments. They should not be described as comprehensive, directly applicable legislation governing every AI application used by an IPS, EPS or community organization.
Specific legal obligations arise from the relevant laws, regulations and professional requirements. AI-related legislative proposals and policy initiatives must also be distinguished from provisions that have entered into force.
For health institutions, this creates a layered governance environment. An application may be relevant to national AI policy while also engaging existing rules concerning health service quality, medical devices, clinical records, professional responsibility and personal data protection.
The Ministerio de Tecnologías de la Información y las Comunicaciones (MinTIC) has responsibilities relating to Colombia's digital development and technology policy. The Ministerio de Ciencia, Tecnología e Innovación also contributes to the wider research and innovation environment.
However, these institutional roles do not displace the responsibilities of health authorities, professional practitioners or the organizations actually deploying technology.
The practical requirement is to identify which obligations apply to the particular AI application rather than treating a general national policy statement as sufficient authority for implementation.
Organizations examining their preparedness can use the Digital Transformation, AI and Cybersecurity Readiness Assessment to structure an internal examination of capabilities, risks and implementation conditions. This is an organizational planning resource, not an official Colombian regulatory assessment or certification.
Policy ambition can encourage innovation. Operational assurance determines whether that innovation is appropriate for the people and services concerned.
Personal data protection, clinical confidentiality and algorithmic accountability
AI applications depend on information. In health services, that information may include diagnoses, medication histories, laboratory results, diagnostic images, disability-related information and other sensitive personal data.
Colombia's Constitution protects fundamental rights relating to privacy and personal information. Law 1581 of 2012 establishes the general personal data protection regime, including particular protections for sensitive data.
Clinical histories are also subject to specific confidentiality and custody requirements, including those established through Resolution 1995 of 1999 and subsequent applicable provisions.
Law 2015 of 2020 provides the framework for interoperable electronic clinical histories. Its purpose should not be confused with granting AI suppliers unrestricted access to health records.
The Superintendencia de Industria y Comercio exercises personal data protection functions within its statutory competence, while health-sector authorities have distinct responsibilities concerning health services and clinical requirements.
An IPS considering an AI application must therefore understand how the technology obtains, processes, retains and potentially transfers information.
Questions concerning lawful processing are especially important when suppliers propose using patient records to train, improve or evaluate algorithms.
Information collected for clinical care cannot automatically be assumed available for unrelated commercial model development. The applicable legal basis, purpose, safeguards and rights of individuals require separate consideration.
Organizations should also distinguish between information used temporarily to generate an output and information retained by a supplier for subsequent purposes.
De-identification may reduce privacy risks, but it is not a guarantee of anonymity. The possibility of re-identification, particularly where datasets contain unusual combinations of characteristics, requires careful evaluation.
Where an AI system generates clinical summaries, staff must verify that the output accurately represents the source information. Generative systems can produce plausible but incorrect statements, omit important qualifications or confuse information relating to different individuals.
These risks make privacy-by-design and proportionate risk mitigation important from the beginning of procurement and development.
Privacy cannot be treated as an issue to resolve after a technology has been connected to operational systems.
Illustrative scenario: automated referral support in Antioquia
Consider a hypothetical EPS operating within the applicable health system arrangements in Antioquia that is evaluating an AI-assisted application to help organize incoming specialist referral requests.
The proposed system would extract information from referral documents, identify missing administrative details and suggest categories for review by authorized personnel.
Its intended function is administrative support rather than determining whether an individual is clinically entitled to treatment or replacing professional assessment.
During testing, the EPS discovers that the system handles standardized electronic referrals more consistently than scanned handwritten documents from smaller facilities.
Some requests containing incomplete information are also assigned inappropriate categories.
The EPS examines whether these differences could disadvantage patients referred from institutions with less developed digital infrastructure.
It decides that the application must not automatically reject requests or make final decisions about access. Employees remain responsible for checking information, obtaining clarification and applying the relevant procedures.
The organization also reviews whether the supplier's handling of personal information complies with the applicable legal requirements.
After a limited evaluation, management compares the time taken to process referrals, the frequency of corrections and the experiences of patients whose requests require additional attention.
Particular attention is given to whether rural referrals experience longer delays than urban referrals.
The review finds that the software reduces some repetitive administrative work but requires additional safeguards for documents that cannot be interpreted reliably.
Rather than extending automation indiscriminately, the EPS retains manual review arrangements and works with participating institutions to improve the quality of referral information.
The scenario demonstrates that AI can improve administrative processes without being permitted to determine individual access through unverified or opaque decisions.
Its success depends on preserving appropriate human responsibility and identifying where the technology performs less reliably.
Algorithmic bias and unequal access across Colombia's territories
Colombia's geographic and social diversity creates an important challenge for the development and evaluation of AI.
Data generated by large urban hospitals may not adequately represent the circumstances of people receiving care in dispersed rural communities.
Differences in disease prevalence, access to diagnostic equipment, environmental conditions, language, ethnicity, disability and socioeconomic circumstances may affect how well a model performs across populations.
Algorithmic bias does not necessarily arise because developers intentionally discriminate. It can result from incomplete datasets, historical patterns of unequal access or the use of variables that indirectly reflect disadvantage.
For example, an algorithm predicting future health service use may assign lower apparent need to a population that historically received fewer services because of geographic barriers.
If previous utilization is treated as a straightforward measure of clinical need, existing inequalities may be reproduced rather than corrected.
Similarly, an application trained on clinical documentation from well-resourced institutions may struggle with records created in facilities using different terminology or less consistent electronic systems.
This makes data-led equity planning an essential component of responsible AI evaluation.
Organizations should examine whose information was used to develop a model, whose circumstances may be underrepresented and whether performance differs across relevant groups.
In departments such as Chocó, La Guajira, Amazonas and Vaupés, implementation may be shaped by transport, connectivity, workforce distribution and the availability of specialist services.
These territories are not interchangeable. Each contains different populations, institutions and local conditions.
Engagement with Indigenous and Afro-Colombian communities should recognize local participation, cultural identity and appropriate communication arrangements without treating communities as homogeneous.
A model that performs adequately in one setting should not be assumed suitable everywhere simply because it has been developed or tested within Colombia.
The stronger approach is to evaluate technology against the population and service environment in which it will actually operate.
Illustrative scenario: specialist access and AI-supported triage in rural Chocó
Consider a hypothetical rural IPS in Chocó exploring whether an AI-assisted tool could help clinicians organize information before requesting specialist advice.
The service supports patients who may face long journeys and substantial practical difficulties accessing specialist appointments.
Its clinical team hopes that structured information gathering could make referrals clearer and reduce avoidable requests for additional documentation.
However, the proposed application has primarily been evaluated in larger institutions with reliable connectivity and standardized electronic records.
Before considering routine use, the IPS examines whether its staff can obtain the information the application requires and whether the technology performs appropriately with the available equipment.
During a small evaluation, clinicians identify cases in which incomplete clinical information produces misleading recommendations.
The service decides that the application may assist with organizing documentation but should not independently determine clinical urgency or replace professional assessment.
Where connectivity is unavailable, employees continue using established clinical and referral procedures.
The IPS also considers whether the tool is accessible to employees with different levels of digital experience and whether its use would increase consultation times.
Patients are informed about relevant aspects of the service process, and personal information is handled through authorized arrangements.
Management reviews the quality of referrals, the number requiring clarification and whether patients experience improvements in access to appropriate specialist advice.
The evaluation reveals that clearer referral templates and better communication with receiving institutions deliver some benefits independently of AI.
The service therefore prioritizes those improvements while continuing to assess whether the technology adds sufficient value to justify its costs and risks.
This example demonstrates why rural innovation should begin with the actual service pathway rather than with the assumption that an advanced application will overcome infrastructure and workforce constraints.
Workforce implications: supporting professionals without displacing responsibility
Artificial intelligence may change the work undertaken by clinicians, administrative employees, care workers and managers.
It could reduce repetitive documentation, help employees locate relevant information and support the organization of service activity.
However, these benefits are not automatic. AI can also introduce additional verification work, increase dependence on suppliers and create uncertainty about professional responsibilities.
In Colombia, the workforce implications need to be understood within the existing differences between regulated health professions, administrative roles and nonclinical care work.
A physician using clinical decision support requires different competencies from an employee using software to organize appointments.
Residential support workers may benefit from tools that assist with preparing routine documentation, but they should not be encouraged to interpret AI-generated clinical recommendations beyond their responsibilities or competence.
Training should address the limitations of AI, the importance of verifying outputs and the circumstances requiring escalation to a qualified professional.
Employees also need to understand that fluent or confident language generated by software is not evidence of accuracy.
For managers, the challenge is to establish clear expectations about permitted uses and the information that may be entered into AI systems.
Unapproved use of publicly accessible generative AI applications can create confidentiality and information security risks, particularly where employees enter identifiable clinical or personal information.
Workforce consultation should therefore form part of implementation.
Employees are often well placed to identify where technology creates duplication, interrupts professional relationships or produces recommendations that do not reflect practical service conditions.
This connects AI adoption with the broader principles of workforce innovation and role redesign.
The objective should be to improve the capacity and effectiveness of human services, not to treat professional judgment, personal relationships or family support as functions that can simply be automated.
AI in residential care, disability support and community services
Artificial intelligence may also have applications outside Colombia's formal clinical service environment, but these require careful distinctions.
Residential institutions for older people, Centros Vida and community organizations may undertake activities involving daily support, social participation, nutrition, mobility, communication and assistance with practical needs.
Law 1315 of 2009 establishes minimum conditions for specified residential institutions serving older people. Law 1276 of 2009 addresses Centros Vida, while Law 1618 of 2013 provides an important statutory framework concerning the rights of persons with disabilities.
These arrangements should not be interpreted as placing every community organization within the same clinical regulatory framework as an IPS.
For nonclinical organizations, potential AI applications might include preparing draft activity reports, organizing staff schedules, identifying gaps in routine documentation or analyzing anonymized service feedback.
Some organizations may also consider assistive technologies incorporating AI, such as communication applications or systems that support accessibility.
However, a technology that assists communication is not equivalent to one that makes decisions on behalf of a person.
Colombia's Law 1996 of 2019 recognizes the legal capacity of adults with disabilities and establishes arrangements concerning supports and safeguards for exercising that capacity.
AI should not be used to displace individual preferences or assume that disability justifies automated restrictions on choice.
For example, a residential organization might use software to identify patterns in recorded falls or mobility difficulties. Such information could prompt professional review, but it should not independently determine that a resident must be prevented from participating in activities.
Similarly, a communication tool may assist someone in expressing preferences, but staff must remain attentive to whether the output accurately reflects the individual's intentions.
These issues connect emerging technology with supported decision-making and individual autonomy.
Family caregivers may also encounter consumer AI applications offering health advice or assistance with care planning. Such tools can provide general information, but they should not be treated as substitutes for qualified clinical assessment or reliable service arrangements.
The opportunity for community services lies in reducing unnecessary administrative burdens and improving accessibility while maintaining human relationships, professional boundaries and personal dignity.
Illustrative scenario: AI-assisted administration in a Cali residential institution
Consider a hypothetical residential institution in Cali supporting older people with varying levels of mobility and personal assistance needs.
The institution's coordinator is considering a generative AI application to help prepare monthly management summaries from existing operational information.
The proposed system would organize information about activities, staffing, incidents and service improvements, potentially reducing the time spent preparing repetitive reports.
However, the institution holds sensitive information about residents, including health-related observations, family circumstances and personal support requirements.
Before proceeding, management examines whether the software can be used without unnecessarily disclosing identifiable information to an external supplier.
The coordinator establishes that employees should not enter residents' names, clinical details or other sensitive information into an unapproved public AI service.
The institution first tests the application using fictional information and non-identifiable operational examples.
During the evaluation, the software produces a summary suggesting that an incident had been resolved, although the underlying example indicates that follow-up remained outstanding.
This error demonstrates the importance of verification.
Management decides that any AI-generated report must be checked against original records and approved by an employee with appropriate responsibility.
The institution also considers whether the time saved justifies the cost of the technology and the additional checking required.
Following the evaluation, the coordinator identifies a simpler improvement: standardizing the information employees record throughout the month.
The institution introduces clearer reporting arrangements before deciding whether to adopt AI more widely.
The example illustrates that responsible innovation may lead an organization to postpone implementation.
The appropriate outcome is not necessarily the purchase of new software. It is a more reliable management process that protects residents' information and produces accurate evidence of service quality.
Funding, procurement and the sustainability of AI investment
AI adoption creates financial commitments extending beyond the initial purchase of software.
Organizations may need to fund technical integration, cybersecurity, employee training, data preparation, supplier support, performance monitoring and future system changes.
For an IPS, these costs must be considered within its operating arrangements and the applicable relationships through which health services are financed and paid for.
EPS may also examine administrative technology expenditure within their legal and financial responsibilities. However, investment decisions cannot be separated from the rules governing the use of health system resources.
Public institutions considering AI procurement must comply with the applicable public purchasing and contractual requirements. Private organizations operate under their own relevant contractual and regulatory obligations.
For residential and community services, financial capacity may be particularly variable. Some organizations depend on municipal programs, contractual arrangements, private payments, charitable contributions or combinations of funding sources.
A technology that appears affordable during a short demonstration may become difficult to sustain once recurring charges and support requirements are included.
Procurement decisions should therefore examine the full operational cost and the consequences of supplier dependence.
An organization needs to know whether it can retrieve its information in a usable format, change suppliers without unacceptable disruption and continue essential activities during an outage.
Contracts should address the intended use of data, service availability, security responsibilities, changes to the model and arrangements for investigating errors.
These questions are particularly important where a supplier can update an AI system after implementation. A material change in how the model operates may affect the validity of an earlier evaluation.
Leaders can use the Governance Maturity Assessment to examine whether responsibilities, oversight and assurance arrangements are sufficiently developed to support complex technology decisions. The resource provides general organizational guidance and does not replace Colombian legal or procurement requirements.
Financial sustainability should be assessed through demonstrable service value rather than anticipated technological prestige.
Evaluating AI performance, incidents and service outcomes
Responsible AI governance requires continuing evaluation after implementation.
A system may perform adequately during a limited trial but encounter different information, workflows or populations when deployed more widely.
Organizations should therefore establish what successful performance means before introducing the technology.
For administrative applications, relevant measures may include processing time, correction rates, employee workload and delays experienced by people using services.
For clinical applications, evaluation may need to consider diagnostic accuracy, false-positive and false-negative results, patient safety, professional interpretation and outcomes relevant to the intended clinical purpose.
For community support applications, the most meaningful evidence may concern accessibility, continuity, employee time and the experiences of people receiving assistance.
Different measures are appropriate because the applications do not perform identical functions.
Organizations should also identify the circumstances in which the technology may no longer be suitable.
Examples include changes in the population served, substantial modifications to software, deterioration in data quality or evidence that particular groups experience less reliable outcomes.
Where incidents occur, the institution should distinguish between technical failure, incorrect data, inappropriate employee reliance and weaknesses in the underlying service process.
This is important because replacing the software may not resolve a problem caused by unclear professional responsibilities.
Conversely, additional employee training cannot correct an algorithm that consistently produces unreliable results for a particular population.
The wider principles of pilot evaluation and structured learning provide a useful foundation for examining these questions.
The Quality Improvement Action Plan Builder can help organizations translate evaluation findings into defined actions, responsibilities and review arrangements. It is a general improvement resource, not an official Colombian AI regulatory mechanism.
Evaluation should also include the people affected by the technology.
Patients may identify confusing explanations, delays or difficulties challenging decisions. Employees may recognize excessive administrative demands or misleading outputs that are not visible in summary performance data.
Such evidence can help leaders distinguish between technology that performs well in isolation and technology that genuinely improves the service pathway.
Building a proportionate AI governance framework for Colombian institutions
The strongest governance approach begins by distinguishing applications according to their intended purpose, potential consequences and institutional context.
An administrative tool that drafts routine correspondence may require relatively straightforward controls, although confidentiality and accuracy remain important.
A clinical decision-support system influencing diagnosis or treatment requires more extensive evaluation, professional oversight and consideration of applicable regulatory requirements.
An application affecting access to services, prioritization or individual rights may require particularly careful scrutiny because errors can create unequal treatment or significant practical harm.
Organizations should identify who has authority to approve deployment, who is responsible for monitoring performance and who can suspend use if concerns emerge.
These responsibilities should be reflected in ordinary management arrangements rather than confined to a technical team.
Clinical leaders need visibility of patient safety implications. Information governance personnel need to understand data processing and security risks. Operational managers need to know how technology changes employee responsibilities and service pathways.
Where relevant, legal and procurement expertise should inform decisions concerning contractual obligations, intellectual property, regulatory classification and personal information.
Governance should also recognize the distinction between an AI supplier's claims and the adopting institution's own responsibilities.
A supplier may provide evidence of technical performance, but the institution still needs to determine whether the application is appropriate for its actual services and users.
For Colombia, national policy can support the development of common principles and institutional capacity. However, territorial differences mean that implementation needs to be realistic about available infrastructure, workforce capability and local service arrangements.
Organizations should avoid creating unnecessarily elaborate governance processes for low-risk activities while ensuring that higher-risk uses receive proportionate scrutiny.
The essential objective is clear accountability: decisions about AI should remain understandable, reviewable and capable of correction.
Future directions and lessons for international care systems
Colombia's AI policy direction creates opportunities to develop technical capability, research partnerships and more effective uses of information across health and related services.
Potential developments include improved diagnostic support, better organization of administrative activity, enhanced population analysis and more accessible communication technologies.
However, the extent and pace of adoption will depend on investment, institutional readiness, professional capability, public trust and the evidence generated by actual implementation.
Colombia's experience is particularly relevant internationally because digital innovation is being considered within a health system characterized by multiple institutional actors and significant territorial diversity.
The challenge is not unique to Colombia. Many countries are seeking to use AI while managing differences in infrastructure, workforce capacity and access to services.
Nevertheless, Colombia's EPS and IPS arrangements, territorial administration, health financing structures and legal framework cannot be treated as directly equivalent to those of other countries.
The transferable lesson lies in the relationship between national ambition and local capability.
AI policy can establish strategic priorities, but safe adoption requires clear responsibilities, appropriate data, workforce preparation and meaningful evaluation.
A second lesson concerns equity. Systems developed using information from better-resourced institutions may not perform consistently in settings where service access, documentation and infrastructure differ.
A third concerns human accountability. Technology can support clinical and operational decisions, but organizations need to preserve the ability to understand, question and correct its outputs.
Finally, sustainable innovation depends on recognizing when simpler improvements may produce greater value than advanced automation.
For Colombia, the next stage should be judged not by the number of AI applications introduced, but by the quality of evidence demonstrating that they improve services without weakening rights, safety or public accountability.
Conclusion
Artificial intelligence offers Colombia opportunities to improve clinical support, administrative coordination and the use of information across health and care services. Its potential is significant, but its practical value depends on the conditions under which it is developed, purchased and used.
Colombia's national AI policy direction provides an important strategic foundation. Existing legal and professional frameworks concerning health services, clinical records, personal data protection and individual rights remain essential to responsible implementation.
For EPS and IPS, AI decisions must reflect the institution's actual responsibilities, the intended application and the potential consequences for patients. For residential and community organizations, opportunities should be assessed against different service functions, workforce capabilities and legal obligations.
Territorial variation is equally important. Technology evaluated in well-resourced urban institutions may not perform in the same way in rural or remote settings. Responsible implementation therefore requires attention to infrastructure, data quality, cultural accessibility and the risk of reproducing existing inequalities.
The strongest forward direction combines proportionate governance with practical evaluation. Professionals should understand the limitations of AI, organizations should retain accountability for decisions and people receiving care should remain central to assessments of value and safety.
Colombia does not need to treat every technological possibility as an immediate operational priority. It needs to identify where AI can make a demonstrable contribution to reliable, equitable and person-centered services.
Ultimately, the success of AI in Colombian health and care will be measured by better decisions, stronger continuity and improved human outcomes—not by the sophistication of the technology itself.