On May 29, 2026, the Argentine government introduced a revised General Corporation Law to Congress. According to President Javier Milei, companies labeled as "automated" could conduct business through autonomous algorithms or AI agents without employing human workers. However, the firm would be liable for any harm caused by these systems.
Following extensive discussions, the concept became more grounded. On August 19, Patricia Bullrich, the head of the ruling bloc in the Senate, announced that automated corporations would have a designated responsible party.
This article explores whether people are ready to let AI manage finances and how various regions are regulating this new agent economy, ranging from bold to conservative approaches.
Is the Law Cracking?
As long as advisory programs are in use, the legal system operates smoothly: chatbots generate text while humans make decisions. However, when agents operate independently, questions arise that lawmakers still need to address: who is involved in the transaction, who owns the funds spent by the agent, and who can be held liable for errors?
On October 13, 2025, California Governor Gavin Newsom signed AB 316, the only regulation currently addressing this critical loophole. The law took effect on January 1, 2026, stating that a defendant who developed, modified, or deployed AI cannot claim in court that harm was caused autonomously by the system. However, AB 316 does not establish a presumption of guilt, leaving causality, foreseeability, and the comparative contribution of third parties untouched.
In November 2025, Amazon filed a lawsuit against Perplexity regarding its Comet browser, which featured an integrated AI assistant. The complaint asserted that the assistant acted without additional clicks after users logged into their accounts. Amazon contended that Comet did not utilize a distinct browser identifier and presented requests with the same user-agent as Chrome.
On August 4, the Ninth Circuit Court ruled that the user interacts directly with the servers, while Comet serves merely as a tool in their hands.
A phrase from IBM’s internal materials from a 1979 seminar states:
Source: IBM seminar participant's account on X from 1979."A computer cannot bear responsibility, thus it should never make management decisions."
Over the past fifty years, this logic has flipped: machines now make decisions, while accountability falls on the individuals who created, modified, or deployed them.
Freedom for AI Agents
On May 29, 2026, the Argentine government under President Javier Milei submitted a new draft of the General Corporation Law to Congress, proposing nearly 300 articles to replace the existing law from 1972. This new legislation introduced two previously unused constructs: "automated company" and "operational DAO."
The structure of the first allows a legal entity to operate through autonomous algorithms or AI agents without any human employees for daily operations. This does not constitute a separate type of company; any corporation can adopt this status by including "automated" in its charter and name. In exchange, it gains full legal personality and limited liability. For damages caused by its systems, the firm is liable with its assets, rather than the program itself.
On June 3, 2026, Milei and Deregulation Minister Federico Sturzenegger published an op-ed in the Financial Times titled "Argentina Invites AI to Free Itself."
The proposal was based on three main points:
- not to regulate AI to allow for free development;
- to create a separate legal category for it;
- to offer a competitive tax environment.
The president likened the idea of a separate legal status to the historical establishment of the Dutch East India Company in 1602, which he argued introduced limited liability to the world.
This view was contested by Israeli historian and futurist Yuval Noah Harari, who, three days after the publication of the op-ed, addressed readers of the Financial Times. He warned that countries granting AI a separate legal status risk creating not a company-state but an unprecedented entity in history—an AI state governed by non-human corporations. In response to Milei's historical analogy, Harari suggested that Buenos Aires might not become a new Amsterdam but rather a new Batavia, a colonial port managed by a corporation instead of a government.
Batavia, or "Queen of the East," as Jakarta was known until 1942, combined a rigid racial hierarchy with tremendous ethnic diversity. In this city, where Europeans sat atop the societal ladder, lived Chinese, Malays, Arabs, Javanese, and thousands of slaves brought from various corners of the Indian Ocean.
Dear @harari_yuval, thank you for engaging in this fascinating and transcendental debate. We are at the dawn of a new age, which places us, I believe, in a position not that different from the one you described so well in Homo Sapiens and your other books: that time… https://t.co/xdMkXHxjYw
— Javier Milei (@JMilei) June 8, 2026
Milei thanked Harari for participating in the discussion and promised to respond in detail.
In an official government communiqué No. 149, he described the developments as "not the launch of Judgment Day, as in 'Terminator,'" but rather a way to facilitate the regulation of AI structures, asserting that this does not exempt a company from the law but, conversely, subjects it to it.
On June 5, Santiago Siri, founder of The DemocracyOS Foundation, commented to Chequeado that granting AI legal personhood is unprecedented, effectively making the intellectual system an independent legal entity.
In the Senate, discussions quickly shifted to the practical matter of accountability. On June 24, the minister defended the project in the General Legislation Commission. Martin Soria of Union por la Patria compared the construct to the SAS form once used by drug cartels and reminded the committee of the FATF requirements.
On August 19, Milei's allies retreated. Bullrich announced a modification: automated companies would have a responsible individual—either a physical or legal person. According to her, accountability needed to be made stricter, so that in the event of serious harm, it would be clear who could be held accountable. The motivation was straightforward: without this amendment, coalition members refused to sign the committee's conclusions.
The requirement for a human representative was included in the DAO construct from the outset, in Article 260.
In mid-June 2026, researchers from Princeton University presented a paper titled "CEO-Bench: Can Agents Play the Long Game?" In the experiment, 13 leading AI models were tasked with managing a virtual company with an initial capital of $1 million over 500 simulated days. Most models ultimately went bankrupt.
Three models significantly increased the starting capital: Claude Fable 5 finished the experiment with $47.1 million, Claude Opus 4.8 with $27.8 million, and GPT-5.5 with $21.3 million. A simple script written by a human outperformed nearly all participants, raising the virtual capital to $15.8 million.
CEO-Bench results after changes to the simulator in July 2026. Source: CEO-Bench.Milei proposed establishing companies managed by AI agents upon discovering they could not maintain a virtual balance even in a simulated market environment.
The Global Map in Search of Measures for Human Presence
No global standard for accountability regarding the actions of AI agents currently exists. Numerous jurisdictions are testing their solutions—from strict limits on agent authority to simple public reprimands of violators—but nearly all boil down to one aspect: the level of human presence and responsibility.
China
On May 8, 2026, the Cyberspace Administration of China, together with the National Development and Reform Commission and the Ministry of Industry, issued recommendations for the orderly application of AI agents. The regulatory framework took effect on July 15. Article 6 of the document requires that before deployment, a decision level for each agent's decision be established:
- the decision is made solely by a human;
- the decision requires human confirmation;
- the agent acts entirely independently.
Users have the right to be informed of the decision and to approve it last, and the agent's actions must not exceed the mandate given to it. For the highest level, preliminary checks, real-time monitoring, and quarterly reporting are required.
China chose authority over status, becoming the first in the world to implement such a scheme at the administrative level.
European Union
On July 24, 2026, the European Union published regulation 2026/1744, known as the Digital Omnibus. The preparation for its adoption took about eight months, and at the last moment, obligations for high-risk systems were postponed from August 2, 2026, to December 2, 2027, and for those embedded in regulated goods from Annex I to August 2, 2028. The bans, transparency requirements, and obligations for general-purpose model providers remained unchanged. In other words, the EU opted for a risk classification approach, which requires significantly more time.
South Korea
On January 22, 2026, South Korea's framework law on AI came into effect—the second comprehensive act in the world after the European one. Article 36 mandates foreign suppliers who exceed certain revenue or audience thresholds to appoint a representative with an address in the republic; this individual bears legal responsibility for compliance with the requirements by their principal. Here, human accountability at the end of the chain is explicitly established in the norm, rather than inferred from general principles, as in most other jurisdictions. The penalties, which were postponed for a year, remain symbolic, with a ceiling of about 30 million won.