Part of the AI Regulation News hub.
Invima set AI rules for its own staff and built in a clause that pushes them onto counterparties by contract
Colombia's food and drug regulator wrote an artificial intelligence policy for its own desks. The part worth reading is the sentence that lets the same policy travel out to suppliers through a contract clause.
Bottom line: Binding, and already signed. Numeral 2 makes the guidelines of mandatory compliance for Invima dependencies, public servants, contractors and other collaborators. For third parties holding any legal relationship or link with the Institute and having access to its information systems, technology resources or digital assets, the text says the guidelines are exigible only to the extent they have been incorporated through clauses, obligations, annexes or express provisions of the relevant contract, agreement or legal instrument. No transition period and no compliance deadline appear anywhere in the document.
Who this affects: Regulatory affairs and quality officers at pharmaceutical and medical device sponsors that hold or seek Colombian sanitary registrations, the local regulatory agents and consultancies that file on their behalf, Colombian counsel advising on market access, physicians and laboratory staff whose submissions pass through Invima systems, and Invima personnel themselves.
Issue date: Decision date 6 August 2026, the date printed on the circular and on its electronic signature registry. Published in the Diario Oficial of 10 August 2026. The draft was posted for public comment from 23 to 27 July 2026. No deadline is set.
What changed: Invima now has a written artificial intelligence directive where it previously had none of its own: a three-tier risk classification for AI systems it acquires, develops or uses, an algorithmic impact assessment that may be run before a high-risk deployment, six categories of prohibited use, separate rules for generative tools, an algorithmic transparency publication duty, and a table assigning responsibilities to named offices including third parties bound by contract.
Analysis: The interesting move is not the content of the controls, which tracks CONPES 4144 and the joint directive it cites. It is the delivery mechanism. An internal administrative instruction becomes a supplier obligation the moment a procurement officer drops it into an annex, with no new statute and no rulemaking. That characterisation is ours; the circular does not describe itself in those terms.
Primary sources: Circular Externa 1000-018-2026 (PDF, Invima) · Direct PDF download (Invima library) · Gazetted text record (vLex, Diario Oficial 10 August 2026)
- Instrument (EN)
- External Circular No. 1000-018-2026, Guidelines for the use of artificial intelligence
- Authority
- Instituto Nacional de Vigilancia de Medicamentos y Alimentos (Invima), Direccion General
- Jurisdiction
- Colombia
- Status
- Signed and issued. Ten of ten electronic signatures recorded, signed by the Director General
- Bindingness
- Of mandatory compliance for Invima dependencies, public servants, contractors and other collaborators. Exigible against third parties only to the extent incorporated by clause, obligation, annex or express provision of the governing contract or agreement
- Issue date / next deadline
- 6 August 2026. Gazetted 10 August 2026. No deadline, transition period or effective date is stated in the text
- Signed by
- Francisco A. G. Rossi Buenaventura, Director General
- What it is not
- Not a marketing authorisation pathway, not a medical device classification rule, and not a general-market AI statute
- Primary source
- https://www.invima.gov.co/biblioteca/circular-externa-ndeg-1000-018-2026pdf
What the circular sets out to do
Numeral 1 states the objective in ordinary administrative language: establish guidelines for the use and application of artificial intelligence that strengthen and drive the activities of the Institute while preserving information security, personal data protection and copyright protection, in its digital and physical environment, through adequate management of use, risks and associated services.
It is issued in the framework of CONPES 4144 of 2025 and Joint Directive 007 of 2025 of the Procuraduria General and the Defensoria del Pueblo. Neither of those is new. What is new is that a sanitary authority has now written its own operating version of them.
The normative reference list in numeral 3 runs to sixteen items, from the 1991 Constitution and Law 23 of 1982 on copyright through Law 1581 of 2012 on personal data to Superintendence of Industry and Commerce External Circular 002 of 2024. Invima is not claiming new legal authority here. It is compiling authority it already sits under and applying it to a technology.
The four control areas, checked against the text
Risk management is present and it is structural, not decorative. Numeral 5 requires the Institute to classify the AI systems it acquires, develops or uses into high, medium and low risk. High risk covers systems that intervene in decisions with significant legal or material effects on people's rights or on sanitary safety, and the text gives examples: systems supporting surveillance and control actions, prioritisation of filings that affects access to products, and processing of sensitive data. Those require reinforced human supervision and periodic review. Before deploying a high-risk system the responsible unit, with the Information Technology Office and the Information Security Officer, may carry out an algorithmic impact assessment. The modal there is podra, not debera. It is a permission, not a mandate, and a reader who reports it as a required assessment has overstated the text.
Information security is present twice. Numeral 4.2 sets availability, confidentiality, integrity and authenticity as principles. Numeral 6 then converts them into use rules, including a duty to adopt administrative and technological security measures over information generated, consulted or applied through AI, and it cross-references numerals VIII and IX of the Superintendence of Industry and Commerce circular.
Personal data is present and specific. Numeral 6 forbids the use of sensitive data absent the data subject's authorisation and reproduces the article 6 and article 5 categories from Law 1581 of 2012, health data, sexual life and biometric data among them. Numeral 10(e) adds that where an algorithmic system involves processing of personal data, the Institute will guarantee compliance with Law 1581 and its implementing rules.
Copyright is present and, unusually for a document of this kind, it reaches the vendor. Numeral 6 requires measures to prevent AI tools from producing outputs that reproduce protected works, verification of content before use, and demanding from the provider the declarations and warranties on the licensing of the tool and of the data used. Numeral 7(f) then prohibits using tools based on protected works without authorisation. All four reported control areas confirm.
Human decision, social scoring, mass surveillance
The sentence the scout flagged sits in numeral 4.1, under the proportionality and harmlessness principle, and it reads in the source: Las decisiones finales con impacto irreversible o de vida o muerte deben ser tomadas por un ser humano. La IA no debe usarse para calificacion social ni para vigilancia masiva. We matched it character for character against both the Invima PDF and the gazetted text at vLex.
Two notes on how to read it. First, the verb is deben, an obligation, but it appears in the principles section, which in Colombian administrative drafting carries interpretive weight rather than a standalone sanction. Second, the same idea is restated in operative form elsewhere, which is what actually gives it teeth. Numeral 7(a) prohibits adopting final administrative decisions with legal effects in an automated way without human intervention. Numeral 7(b) prohibits substituting the motivation of an administrative act. Numeral 10(b) requires that any definitive administrative decision with significant legal or material effects be adopted or validated by a competent public servant who assumes responsibility for it.
The social scoring and mass surveillance prohibition likewise appears twice: as a principle in 4.1 and as a flat prohibited use in numeral 7(d), which also bars processing that generates discrimination.
Generative tools get their own numeral. Every output must be reviewed and validated by the responsible servant before use; normative, case law, technical and quantitative references must be verified specifically because the tool may produce inexact or non-existent information, a risk the circular names in quotation marks as alucinaciones; nothing unverified may enter an administrative act, legal opinion, petition response or official communication; and where a generative tool is used in drafting those, a record of that use is left in the document.
The mechanism a US sponsor should care about
Read numeral 2 and the Paragrafo to numeral 7 together and the shape becomes clear. Numeral 2 extends the guidelines to third parties with any legal relationship or link to the Institute who have access to its information systems, technology resources or digital assets, and then conditions that extension: the guidelines are exigible in the measure in which they have been incorporated through clauses, obligations, annexes or express provisions of the contract, agreement or legal instrument governing the relationship. The Paragrafo to numeral 7 repeats the condition for the prohibited uses specifically, adding that the restrictions bind linked third parties only when expressly incorporated, without prejudice to legal obligations already applicable to them.
The responsibilities table in numeral 12 says the same thing a third time, in a row headed terceros vinculados mediante contrato o convenio: comply with the AI, data protection, confidentiality, information security and intellectual property obligations agreed in the instrument signed with the Institute, and attend to the AI controls Invima has incorporated contractually.
So the regulator has drafted a policy it can hand to procurement as ready-made contract text. Nothing forces the clause into any given contract. But once it is in, a US sponsor's Colombian agent is holding an obligation set written by the agency that reviews its filings, enforceable through the contract rather than through sanitary law. We call that an internal directive made contractually exportable. The circular does not use that framing and does not present itself as a supply chain instrument; the reading is ours.
The practical consequence is narrow and checkable. Someone should read the current Invima-facing agreements and any renewal drafts for an AI annex or an incorporation-by-reference clause. If one is there, the sponsor has inherited the verification duty, the sensitive-data restriction, the public-tool restriction on non-public information, and the vendor warranty requirement on licensing. If one is not there, the circular does not reach the sponsor at all.
Why this is not the SIC circular or the telemedicine resolution
Three Colombian AI instruments now sit close together and they are easy to blur. The Superintendence of Industry and Commerce circular addresses personal data processing and is issued by the data protection authority. Resolucion 1644 addresses telemedicine and human supervision in clinical practice. This one is issued by the sanitary authority, is addressed in the first instance to its own personnel, and governs how AI is used inside and around Invima operations.
Different authority, different addressees, different enforcement route. Invima's circular cites the SIC circular as a reference document in numeral 3, which is a citation, not an incorporation. Reading any of the three as a substitute for the others will produce the wrong compliance answer.
One more boundary worth stating plainly, because the subject matter invites the error. Nothing in this circular approves, classifies, registers or clears an AI-enabled product. It says nothing about software as a medical device. A sponsor looking for a Colombian AI device pathway will not find it here.
What we did not verify
We opened and read the full Spanish text of Circular Externa 1000-018-2026 from Invima's own library, extracted through a PDF parser, covering numerals 1 through 14, the responsibilities table, the signature block and the electronic signature registry. We separately opened the gazetted record at vLex, which states a publication date of 10 August 2026 and a decision date of 6 August 2026, and read the portion of the gazetted text available without a subscription, running through numeral 4.1. The quoted sentence was matched character for character against both extractions.
We did not open the Diario Oficial issue of 10 August 2026 directly, so the gazette date rests on the vLex record rather than on the gazette itself. We did not open Invima's normatividad index page and therefore did not independently confirm how the circular is listed there. We did not open CONPES 4144 of 2025, Joint Directive 007 of 2025, Circular Invima 1000-009-2026, Decreto 385 de 2026, or the referenced ethics guide. We did not read the tecnico annex that numeral 13 says will complement the guidelines, because it does not appear to exist yet. We did not review any actual Invima contract or annex, so we have no evidence about how often the incorporation clause is in fact used.
We refuse to claim that this circular regulates Colombian health companies generally, that it binds any vendor or agent absent express contractual incorporation, that it creates an approval or clearance route for AI products, or that the algorithmic impact assessment is mandatory. We also decline to say how many contracts are affected. We have no denominator for that and did not look for one.
Treat this as contract-review work, not regulatory-monitoring work. The circular binds Invima and its own people directly, and it reaches your Colombian agent only where somebody wrote it into the contract, the annex or an express provision. Ask for the current agreement and search it for an AI clause. If the clause is there, the verification duty, the sensitive-data limits and the vendor licensing warranty are already yours.
Source File
https://www.invima.gov.co/biblioteca/circular-externa-ndeg-1000-018-2026pdf
Open the Invima PDF and confirm three things: numeral 2 (Alcance) contains the phrase serian exigibles en la medida en que hayan sido incorporados mediante clausulas, obligaciones, anexos o disposiciones expresas; numeral 4.1 contains the quoted human-decision and social-scoring sentence; and the Paragrafo to numeral 7 repeats the incorporation condition for the prohibited uses. Then open the vLex record to confirm the decision date of 6 August 2026 and the Diario Oficial publication date of 10 August 2026.
Las decisiones finales con impacto irreversible o de vida o muerte deben ser tomadas por un ser humano. La IA no debe usarse para calificacion social ni para vigilancia masiva. ยท Circular Externa 1000-018-2026, Invima, numeral 4.1, 6 August 2026
FAQ
Does this circular regulate every health company operating in Colombia?
No. It is addressed to Invima dependencies, public servants, contractors and other collaborators of the entity, and to third parties holding a legal relationship with the Institute who have access to its information systems, technology resources or digital assets. Even for that third group, numeral 2 says the guidelines are exigible only to the extent they have been incorporated through clauses, obligations, annexes or express provisions of the governing contract or agreement. A company with no such instrument and no incorporated clause is outside its reach.
Our Colombian regulatory agent files on our behalf. Are they bound?
Only if the incorporation happened. The circular reaches third parties where a clause, obligation, annex or express provision of the contract, agreement or other legal instrument brings it in, and the Paragrafo to numeral 7 repeats that condition for the prohibited uses. The check is a document check, not a legal-interpretation question: read the agreement and look for the clause or an incorporation by reference.
Is this an approval route or a classification rule for AI-enabled medical devices?
No. The circular governs how artificial intelligence is used in Invima's own processes and, where contractually incorporated, by counterparties with access to its systems. It does not create, modify or describe any marketing authorisation, sanitary registration or device classification pathway, and it says nothing about software as a medical device.
Does it ban artificial intelligence from regulatory decision-making?
It does not ban the tools. It reserves the decision. Numeral 7(a) prohibits adopting final administrative decisions with legal effects automatically and without human intervention, numeral 7(b) prohibits letting AI substitute the reasoning of an administrative act, and numeral 10(b) requires a competent public servant to adopt or validate any definitive decision with significant legal or material effect and to carry responsibility for it. Assisted drafting stays available, subject to the verification and traceability rules in numeral 8.
Related briefings
Sponsored Training
Practical AI training for regulated professionals, built around verification, documentation and a defensible process. See the courses.