A team has a customer service agent ready. It will serve customers in Spain, Brazil and Argentina. At the last meeting before launch, someone asks whether the agent has to say it is an AI. Someone else answers that the European law was postponed and that Latin America has no AI law. The meeting moves on.
Are they right? In part. The part they get wrong is already in force. And it is the part that applies to an agent that talks to customers.
This chapter uses official texts where they could be read, and press coverage where they could not. Statuses are as of November 2026 and they change quickly. None of this replaces legal advice.
The argument that there is still time has a basis. The calendar of the European Union’s AI Act changed. The European Commission explains it on its official page (exhibit 143). The Digital Omnibus package of changes postponed the rules for high-risk systems.
The Annex III rules moved from August 2, 2026 to December 2, 2027. Annex III covers biometrics, critical infrastructure, education, employment and migration. High-risk rules for regulated products, under Annex I, moved to August 2, 2028.
December 2, 2027
Annex III high-risk rules (previously August 2, 2026). Exhibit 143
The EU AI Act was postponed, so there is time before designing to comply.
Verdict: Only part of it was postponed. The high-risk rules in Annex III moved to December 2, 2027. That annex covers employment, education, biometrics and critical infrastructure. But the prohibited practices and AI literacy have applied since February 2, 2025. Article 50 transparency has applied since August 2, 2026.
Rule
Classify every agent against the law's categories today. If it touches employment, education, biometrics or critical infrastructure, design for the earliest date. The calendar has already changed once and may change again.
Why medium confidence. The source is the European Commission's official page, updated on October 19, 2026, and the text of Article 50 agrees. Confidence is medium because the calendar changed with the Digital Omnibus package and may change again.
Show the 2 exhibits in detailHide the detail1 for, 1 against, 0 qualify
According to the European Commission's official page, the Digital Omnibus package changed the AI Act timeline. The Annex III high-risk rules (biometrics, critical infrastructure, education, employment, migration) moved to December 2, 2027. High risk within regulated products (Annex I) moved to August 2, 2028. Prohibited practices, AI literacy, rules for general-purpose models and Article 50 transparency already apply.
Numbers
February 2, 2025Prohibited practices and the AI literacy obligation apply.
August 2, 2025Governance and obligations for general-purpose models apply.
August 2, 2026General application and Article 50 transparency rules.
December 2, 2027Annex III high-risk rules (previously August 2, 2026).
August 2, 2028High risk built into regulated products (Annex I).
Limitation Page updated on October 19, 2026; the timeline may change again. According to the Commission and its Transition Pathways portal, the Omnibus amendments apply from July 27, 2026. They came after the political agreement of May 7. A regulation number and a watermarking deadline cited by compliance blogs were not verified in the Official Journal and are not used.
EU Artificial Intelligence Act (artificialintelligenceact.eu)
This is the text of Article 50 of the EU AI Act. A provider of a system that interacts with people must say it is AI by the first interaction, unless it is obvious. Synthetic content (made by AI) must carry a machine-readable mark, and deepfakes must be disclosed. The rule applies to customer service agents.
Numbers
August 2, 2026Date from which it applies.
Limitation It was read on artificialintelligenceact.eu, a reference site that is not the EU Official Journal. Applying it to a specific case requires legal advice.
Read at the source. Checked on Nov 1, 2026.
In Latin America, the phrase “there is no AI law” is true for all five countries in the base. The base is the set of sources on this site.
In Brazil, the Senate passed the AI bill on December 10, 2024 (exhibit 145). It is now in the Chamber of Deputies, in a special committee. It is waiting for the rapporteur’s report, which is the report from the legislator in charge of the bill. Its latest moves, on October 26 and 27, 2026, only attach other bills to it.
In Chile, the Executive’s bill was filed on May 7, 2024 (exhibit 129). The Chamber sent it to the Senate on October 13, 2025. The latest steps, through September 2026, only renew its urgency.
In Colombia, the Government’s bill died on June 19, 2026 (exhibit 137). Its time ran out, without debate. In Mexico, in April 2026, the Senate was discussing a committee report that is not yet law (exhibit 85). Argentina has no enacted law, and the Executive prefers not to regulate the use of AI (exhibit 134).
Technical frameworks do not bind anyone either. For agents, they are still being prepared. NIST is the US standards institute. In January 2026, it opened a public consultation on agent security (exhibit 39). It is a question, not a rule. In February, NIST launched an agent standards initiative with 3 pillars (exhibit 59). It promises guidance, but the announcement contains none and gives no deadline.
On the most sensitive point, the law even took a step back. The original text of the LGPD, Brazil’s data protection law, required a person to review automated decisions. A provisional measure removed that requirement. A provisional measure is an Executive rule with the force of law for a limited time. Later, the bill that became Law 13.853 restored the requirement in a new paragraph.
In 2019, when signing that law, the Executive vetoed the paragraph (exhibits 1 and 2). It gave this reason: requiring human review of every automated decision would make the business models of many companies unviable. It named startups in particular.
The European postponement left out exactly what applies to the agent in the meeting. Article 50 of the AI Act has applied since August 2, 2026 (exhibit 3). Anyone who offers a system that talks to people must disclose that it is an AI, unless that is obvious. They must do so at the latest at the first interaction.
The same article asks for two more things. Synthetic content, the content AI generates, must be marked in a format a machine can read. And deepfakes, fake images, audio or video that look real, must be disclosed.
That is not all that already applies. Prohibited practices have applied since February 2, 2025 (exhibit 143). AI literacy has applied since that date too: the duty to train the people who use these systems. And the European Commission updated its page on October 19, 2026. A calendar that has already changed once can change again.
A customer-facing agent does not need to say it is an AI.
Verdict: In the European Union it does. Article 50 of the AI Act has applied since August 2, 2026. It requires telling people no later than the first interaction, unless it is obvious. It also requires marking synthetic content in a machine-readable format. Outside the EU, the sources in the evidence base show no obligation in force. In Mexico, for example, the Senate bill is not law yet.
Rule
Have the agent say it is an AI in the first message of every conversation, in every market. It costs little and in the EU it is already mandatory. Mark the content it generates in the metadata.
Why high confidence. The text of Article 50 and the European Commission's official page agree on the obligation and the date. The text was read on a reference site, not in the Official Journal. Applying it to a specific case requires legal advice. No source supports the claim. The only nuance is that outside the EU the sources in the evidence base show no obligation in force.
Show the 3 exhibits in detailHide the detail0 for, 2 against, 1 qualifies
EU Artificial Intelligence Act (artificialintelligenceact.eu)
This is the text of Article 50 of the EU AI Act. A provider of a system that interacts with people must say it is AI by the first interaction, unless it is obvious. Synthetic content (made by AI) must carry a machine-readable mark, and deepfakes must be disclosed. The rule applies to customer service agents.
Numbers
August 2, 2026Date from which it applies.
Limitation It was read on artificialintelligenceact.eu, a reference site that is not the EU Official Journal. Applying it to a specific case requires legal advice.
According to the European Commission's official page, the Digital Omnibus package changed the AI Act timeline. The Annex III high-risk rules (biometrics, critical infrastructure, education, employment, migration) moved to December 2, 2027. High risk within regulated products (Annex I) moved to August 2, 2028. Prohibited practices, AI literacy, rules for general-purpose models and Article 50 transparency already apply.
Numbers
February 2, 2025Prohibited practices and the AI literacy obligation apply.
August 2, 2025Governance and obligations for general-purpose models apply.
August 2, 2026General application and Article 50 transparency rules.
December 2, 2027Annex III high-risk rules (previously August 2, 2026).
August 2, 2028High risk built into regulated products (Annex I).
Limitation Page updated on October 19, 2026; the timeline may change again. According to the Commission and its Transition Pathways portal, the Omnibus amendments apply from July 27, 2026. They came after the political agreement of May 7. A regulation number and a watermarking deadline cited by compliance blogs were not verified in the Official Journal and are not used.
Mexico has no general AI law. A draft under discussion in the Senate proposes a horizontal (cross-sector), risk-based approach. It includes a national system registry, explainability (explaining how a system decides) and human review of automated decisions. It also gives the power to suspend systems facing a serious and imminent risk. The opposition warns of a censorship risk.
Limitation It is press coverage of a bill under discussion, not a law. Two points were only seen in secondary sources and are not used. They are whether a constitutional reform is needed and the details of the new 2025 personal data law.
Read at the source. Checked on Nov 1, 2026.
In Latin America, having no AI law does not mean having no rules. Data protection laws already apply to AI, and their authorities act. The clearest case is in Brazil (exhibit 5).
The ANPD is Brazil’s data authority. On July 2, 2024, it issued a precautionary stop on a use of data. Precautionary means temporary, while the case is reviewed. Meta was using that data to train its AI. On August 30, the ANPD lifted the suspension with conditions. It asked for notice by email and in the app, and a simple way to opt out. It also asked for a 30-day wait and no data from anyone under 18.
It is a model training case, not an agent case. But it shows how a data authority reasons when there is no AI law: it applies the law it has.
Chile adds a date that matters. Its new data protection law is Law 21.719 (exhibit 130). It was due to take effect on December 1, 2026. It creates a data protection agency. The Executive asked to postpone it to December 1, 2027. At the date of the article, the bill had been filed, not approved.
Latin America has no AI law, so an agent has no rules to follow.
Verdict: It is true that there is no AI law in force. As of November 2026, Argentina, Brazil, Chile, Colombia and Mexico do not have one. There are bills in progress and one that died. But personal data laws already apply to AI. Brazil's data protection authority suspended the use of personal data to train Meta's AI. Later it allowed it again with conditions. It required notice, a simple way to opt out and a 30-day wait. It also required no data from anyone under 18.
Rule
Treat every agent that handles personal data like any other data processing. Give notice, offer a simple way to opt out, and do not use minors' data without a separate review. Check when each data law takes effect. Chile's may move from December 1, 2026 to December 1, 2027.
Why medium confidence. The Brazilian and Chilean bill records and the authority's decision are official. The status in Argentina, Colombia and Mexico comes from the press and a tools website. The Mexican article is from April 2026, and all of it changes fast. The Brazilian authority's case is about model training, not agents.
Show the 7 exhibits in detailHide the detail4 for, 2 against, 1 qualifies
The official record of Brazil's AI legal framework bill, risk-based and inspired by the EU law. The Senate approved it on December 10, 2024, but it is not law yet. In the Chamber of Deputies, a special committee awaits the rapporteur's report. The latest moves, on October 26 and 27, 2026, only attach other bills.
Numbers
December 10, 2024Approval in the Senate.
May 20, 2025Appointment of the rapporteur in the Chamber of Deputies.
Limitation The status changes: check the record before citing it. A private monitor claims the rapporteur postponed the vote until after the October 2026 elections; that was not verified and is not used.
This is the official legislative record of Chile's AI bill. The Executive introduced it on May 7, 2024. The Chamber of Deputies sent it to the Senate on October 13, 2025. It is in its second constitutional stage, awaiting the first report of the Committee on Challenges of the Future, Science, Technology and Innovation. The latest steps, up to September 1, 2026, only renew its urgency. It is not law.
Numbers
May 7, 2024Bill introduced.
October 13, 2025The Chamber's dispatch to the Senate.
September 1, 2026Latest recorded step (simple urgency).
Limitation The Senate's public legislative-tracking web service was read. The human-facing page (senado.cl, bulletin 16821-19) did not show the bulletin when opened automatically, and the Chamber's page returned HTTP 403. The status changes: check before citing.
Colombia has no AI law. The government's bill, sent with an urgency message, died without debate on June 19, 2026 because its deadline expired. There are two new bills, one in the House and one in the Senate.
Numbers
June 19, 2026Date the government's bill died.
Limitation Analytical press. The guidelines of the Superintendence of Industry and Commerce on personal data in AI systems are in External Circular 002 of 2024. They could not be opened and are not used.
Mexico has no general AI law. A draft under discussion in the Senate proposes a horizontal (cross-sector), risk-based approach. It includes a national system registry, explainability (explaining how a system decides) and human review of automated decisions. It also gives the power to suspend systems facing a serious and imminent risk. The opposition warns of a censorship risk.
Limitation It is press coverage of a bill under discussion, not a law. Two points were only seen in secondary sources and are not used. They are whether a constitutional reform is needed and the details of the new 2025 personal data law.
On July 2, 2024, Brazil's personal data authority provisionally stopped Meta from training its AI with personal data. On August 30 it lifted the suspension with a compliance plan. The plan includes notice by email and in the app, and a simple form to object. It also sets a 30-day wait and no data from people under 18. Without an AI law, the data authority already acts on AI.
Numbers
30 daysWait between the notice and the use of the data.
under 18People whose data was excluded.
Limitation It is a model training case, not an agent case. Other actions by the authority (an AI regulatory sandbox and AI as an enforcement priority) were only seen in secondary sources and are not used.
Chile's Law 21.719 on personal data was published on December 13, 2024. It was due to take effect on December 1, 2026, with a new Data Protection Agency. The Executive sent the Senate a bill, with top urgency, to postpone it to December 1, 2027. At the date of the article the bill was filed but not approved, even though the headline says the law is postponed.
Numbers
December 1, 2026Effective date set in the law.
December 1, 2027New date proposed by the postponement bill.
Limitation Secondary source. The effective date changes the compliance plan of any agent that processes data about people in Chile: confirm it in the Official Gazette. The status of the AI bill is in its own record, the Senate of Chile page on Bill 16821-19.
Argentina has no enacted AI law, but there are bills with a national registry and risk classification. One bill from the Executive covers automated companies, but it is a corporate law reform, not AI regulation. The Executive prefers not to regulate AI use. The personal data law, Law 25.326, applies. The data authority issued a resolution and a provision on AI that do not bind private companies.
Limitation A secondary source from a tools website, not an official publication. The status of the bills changes. Article 20 of Law 25.326 limits judicial or administrative decisions based only on automated processing. It is not a general right like Article 22 of the EU data regulation.
Read at the source. Checked on Nov 1, 2026.
What about human review? Here it pays to read the text of the law, because people say far more than it contains. Automated processing means that a system processes the data without a person. In Brazil, Article 20 of the LGPD grants a right (exhibit 1). A person can ask for review of a decision made solely through automated processing. But after the veto, the law does not require that review to be done by a person.
In Argentina, Article 20 of Law 25.326 covers a narrow case (exhibit 134). It covers judicial or administrative decisions based solely on automated processing. It is not a general right. In Mexico, the Senate report proposes human review of automated decisions, but it is a bill.
The law requires a person to review every decision an agent makes.
Verdict: Not in general, and it depends on the country. In Argentina, Article 20 of Law 25.326 limits only one kind of decision. These are judicial or administrative decisions based solely on automated processing. In Brazil, Article 20 of the LGPD gives a right to ask for review of an automated decision. But it does not require a person to do it. The paragraph that required this was vetoed when Law 13.853 of 2019 was enacted. In Mexico, the Senate bill proposes human review, but it is not law yet.
Rule
Give the affected person a way to ask for review of an agent's decision. Decide in writing who reviews it and how fast. If the decision has a significant effect on someone, have a person review it even if the law does not require it.
Why medium confidence. The Brazilian part comes from the official text of the LGPD and of the veto. Both were read on the Chamber of Deputies site, because the Presidency's site blocks reading. The Argentine part comes from a secondary source. The Mexican part is press coverage of a bill.
Show the 4 exhibits in detailHide the detail1 for, 3 against, 0 qualify
Mexico has no general AI law. A draft under discussion in the Senate proposes a horizontal (cross-sector), risk-based approach. It includes a national system registry, explainability (explaining how a system decides) and human review of automated decisions. It also gives the power to suspend systems facing a serious and imminent risk. The opposition warns of a censorship risk.
Limitation It is press coverage of a bill under discussion, not a law. Two points were only seen in secondary sources and are not used. They are whether a constitutional reform is needed and the details of the new 2025 personal data law.
Argentina has no enacted AI law, but there are bills with a national registry and risk classification. One bill from the Executive covers automated companies, but it is a corporate law reform, not AI regulation. The Executive prefers not to regulate AI use. The personal data law, Law 25.326, applies. The data authority issued a resolution and a provision on AI that do not bind private companies.
Limitation A secondary source from a tools website, not an official publication. The status of the bills changes. Article 20 of Law 25.326 limits judicial or administrative decisions based only on automated processing. It is not a general right like Article 22 of the EU data regulation.
Câmara dos Deputados (Brazil), Legislação Informatizada
Official text of Article 20 of Brazil's data protection law (LGPD), as worded by Law 13.853 of 2019. People have the right to ask for review of decisions made only by automated processing of their data that affect their interests. This includes consumer or credit profiling. Paragraph 3, which required the review to be done by a person, is shown as vetoed.
Numbers
Law 13.853Law of July 8, 2019 that gave Article 20 its current wording (Diário Oficial da União, July 9, 2019).
§ 3 (vetoed)Paragraph that required review by a person.
Limitation planalto.gov.br blocked automated reading (Acesso Negado), so the Chamber's official copy was read; it states that it does not replace the Diário Oficial. The original 2018 text (Lei 13.709) required review “por pessoa natural” in the caput. MP 869/2018 removed it and the paragraph that restored it was vetoed. Applying it to a specific case requires legal advice.
“tem direito a solicitar a revisão de decisões tomadas unicamente com base em tratamento automatizado”
Presidência da República, Mensagem nº 288 (read on the Câmara dos Deputados site)
Partial veto message on the bill enacted as Law 13.853. It vetoed paragraph 3 of LGPD Article 20, which required a person to review automated decisions. According to the veto, requiring human review of every automated decision would make many companies' business models unworkable, especially startups. It would also affect credit risk analysis.
Numbers
Mensagem nº 288, July 8, 2019Veto message.
Limitation The reasons are given by the Executive that issued the veto. According to the press, Congress upheld the veto in October 2019; that was not verified and is not used.
“A revisão de que trata o caput deste artigo deverá ser realizada por pessoa natural”
Quote checked against the original
Read at the source. Checked on Nov 2, 2026.
That leaves the objection to voluntary frameworks: if they do not bind anyone, what are they for? They bring order to design while the law catches up. OWASP is a software security community. In December 2025 it published its first consensus list specific to agents (exhibit 33). It names 10 risks, from ASI01 to ASI10.
Three weigh most on an agent. Goal hijacking (ASI01) is outside text changing what the agent pursues. Tool misuse (ASI02) and identity and privilege abuse (ASI03) have the same cause. The agent has more permissions or autonomy than it needs. Each risk comes with concrete mitigations. Two examples are human approval for destructive actions and cost caps with an automatic cutoff.
NIST also has something more usable than the consultation. The NCCoE, its cybersecurity center, published a draft (exhibit 50). It proposes treating the agent as an identity of its own. The agent authenticates and is authorized through policies. It is tied to the person who delegates to it, and it logs its actions. It is a draft for comment, with questions rather than controls.
Voluntary frameworks like NIST's or OWASP's are useless because they bind no one.
Verdict: They are not binding, but they bring order to design while the law catches up. OWASP's list for agents names ten risks with concrete mitigations. They include goal hijacking (ASI01) and tool misuse (ASI02). Identity and privilege abuse (ASI03) is also on the list. In early 2026, NIST was still consulting on and announcing agent standards, with no guidance published.
Rule
Use OWASP's list for agents as a design review for every agent. Write down which control covers each risk, from ASI01 to ASI10, and which one stays open.
Why medium confidence. All four sources are primary and agree that these frameworks are not binding. That they are useful is a design judgment, not a measurement. OWASP is an expert consensus list. NIST's work consists of a public consultation, an announcement and a US draft. The draft already proposes something usable: treating the agent as its own identity, tied to the person who delegates to it.
Show the 4 exhibits in detailHide the detail2 for, 2 against, 0 qualify
NIST launches a standards initiative for agents with three pillars. They are support for industry-led standards, open protocols for agents, and research on agent security and identity. It promises guidance in the coming months, but the announcement includes no guidance and no concrete deadline.
Numbers
3 pillarsWorkstreams of the initiative.
Limitation Government announcement. Everything is voluntary and there are no deliverables yet.
NIST, Center for AI Standards and Innovation (CAISI)
NIST asks for public input on the security of agents that can make persistent changes outside themselves. It says its own research showed agent hijacking risks. As an example attack, it gives indirect prompt injection (hidden instructions in content the agent reads). It asks whether standard cybersecurity practices are enough for agents. It is an open consultation, not guidance: in early 2026 the US government still had no guidance for agents.
Numbers
March 9, 2026Deadline for submitting responses.
Limitation Federal Register entry (vol. 91, no. 5). The official copy on govinfo.gov was read. It is a question, not a rule; nothing is mandatory.
It is OWASP's first consensus list built for agents. The ten risks run from agent goal hijacking (ASI01) to rogue agents (ASI10). They include tool misuse (ASI02) and identity and privilege abuse (ASI03). They also include the agentic supply chain (ASI04) and cascading failures (ASI08). It adds the principle of least agency: do not grant autonomy that is not needed. Its mitigations include least-privilege profiles per tool, human approval for destructive actions and controls outside the model. It also asks for cost, rate or token caps with automatic cutoff.
Numbers
10 risks (ASI01 to ASI10)Size of the list. ASI02 and ASI03 extend excessive agency (LLM06) from the 2025 list to multi-step flows.
Limitation It is a voluntary consensus framework: it binds no one and measures nothing. In the agent list, unbounded consumption is no longer an item of its own: it appears as a scenario and mitigation inside ASI02. The full PDF (57 pages) was read.
“deploying agentic behavior where it is not needed expands the attack surface”
Booth, H., Fisher, B., Galluzzo, R., Roberts, J. NIST, NCCoE
It is a NIST draft for a demonstration project that applies identity standards to agents. It proposes to identify the agent as distinct from a person, and to authenticate and authorize it with OAuth 2.0 and policies. It also proposes to bind each agent to the person who delegates to it, and to log its actions. It asks how to know the sensitivity of the data the agent gathers, and whether the user may see the combined answer. It also asks how to set least privilege (only the permissions needed) when the agent's actions are not predictable.
Numbers
6 groups of questionsGeneral, identification, authentication, authorization, auditing and prompt injection. Comments from February 5 to April 2, 2026.
Limitation It is a draft for consultation, with questions rather than controls. It leaves RAG out of scope. The PDF (11 pages) was read; the project page returned 403.
“How do we establish "least privilege" for an agent”
The person in the meeting was right on one point.The heaviest part of the European law was postponed. And none of the five Latin American countries in the base has an AI law in force.
They were wrong about what matters for their agent. In the European Union, disclosing that it is an AI is already mandatory. In Latin America, data protection laws already apply. One authority has already used them to halt AI training on people’s data.
No law in force in the base requires, as a general rule, that a person review every decision. Brazil grants a right to ask for review, without requiring a person. Argentina limits a narrow case. Mexico proposes it in a bill.
For agents, the most useful voluntary framework today is the OWASP list. NIST’s work is still consultations and drafts. The claim about disclosure carries high confidence, because the text of the article and the Commission’s page agree. The rest carry medium confidence, because the calendar changes and several Latin American sources are press coverage.
Design for the strictest rule already in force, not for the one that has not arrived yet. Treat every agent that touches personal data as one more processing of personal data.
Now:
Make the agent say it is an AI in the first message of every conversation, in every market. Mark the content it generates in its metadata.
Classify each agent under the categories of the European law. If it touches employment, education, biometrics or critical infrastructure, design it for the earliest date.
If the agent uses personal data, give notice and offer a simple way to opt out. Do not use data from minors without a separate analysis.
Leave a path for the affected person to ask for review of an agent’s decision. Write down who reviews it and within what time. If the decision has a significant effect, have a person review it even if the law does not require it.
Use the OWASP list for agents as a design review. Note which control covers each risk, from ASI01 to ASI10, and which one stays open.
Write down the date you read each rule. Read it again before every launch.
The design review already includes disclosure, a table of personal data rules by country and decision review. The threat map covers goal hijacking through prompt injection: instructions hidden in a text. The action policy puts limits on tools and privileges. The technical controls behind those risks are in Security by design. The last chapter brings every rule together into a single decision framework.