Legal risk management of AI in business activities: not simple

Perplexity, a U.S. artificial intelligence company that has developed an "AI agent" feature called Comet that can shop on its own and "infiltrate" the world's largest shopping platform Amazon, presenting major legal, technical, and administrative challenges in the context of booming AI commerce. So what do businesses need to do to use and manage AI safely and effectively in the context of today's technology boom?

There needs to be a clear process for AI leakage or abuse scenarios.

The Amazon incident – Perplexity

The event began when Amazon discovered that Perplexity's AI agent Comet had automatically placed an order on the Amazon platform without permission. According to Amazon's warning letter dated October 31, 2025, Comet used the "Buy with Pro" feature to book customers from November 2024, and Perplexity at that time agreed to suspend operations. In early August 2025, Amazon discovered a new version of Comet anonymously in the form of a Chrome browser to circumvent the rules, forcing Amazon to block access. Just 24 hours later, Perplexity updated Comet to bypass Amazon's control. At the end of October 2025, Amazon officially asked Perplexity to comply with its Terms of Use and immediately stop this behavior.

When Perplexity did not stop, on November 4, 2025, Amazon filed a lawsuit in the federal court for the Northern District of California, accusing Perplexity of "stealing" user account access and faking human behavior to automate transactions. In the lawsuit, Amazon asserts that the Comet system poses a potential risk to customer data security and that Perplexity has repeatedly ignored requests for termination. On 09/03/2026, the Judge issued a preliminary injunction prohibiting Perplexity from allowing Comet to access Amazon, noting that there is clear evidence of illegal intrusion[1].

Legal issues

According to Amazon's argument in the petition as well as related documents, the company argues that Perplexity violates U.S. law despite being warned repeatedly about unauthorized behavior.

Violation of terms of service: Amazon alleges that Perplexity intentionally violated Amazon's Terms of Use by allowing Comet to operate non-transparently. According to the warning letter, Amazon stated that Perplexity's behavior had "secretly intruded" into the system and "violated its Terms of Use". Amazon also cited a provision in the clause that requires all AI agents to operate transparently on its online shopping site and not impersonate a browser. This breach led Amazon to determine that Comet's activity was not authorized.

Unauthorized computer intrusion: Amazon relies on the American Computer Fraud and Abuse Act (CFAA) and similar regulations of the State of California to sue Perplexity. Amazon claims that Perplexity illegally hacked into Amazon systems and user accounts. Amazon's legal partner in the case emphasized that when access is revoked and the third party still resists circumventing the law, it could constitute a crime under the State of California regulations.

Liability and compensation: The concept of AI liability is increasingly promoted. In this case, the focus is on whether the AI agent, specifically Comet, is equally responsible for the user or a third party intrusion. The California preliminary court appears to view Perplexity as an automated third party, not the average user.  Because in this case, the Comet impersonates the user's behavior. This will affect the level of compensation if damage is incurred. Businesses need to carefully review compensation terms, limit liability in contracts with AI partners, and ensure compliance with governing laws.

Risk management from the adoption of AI

Managing AI operations safely and effectively when AI is increasingly able to handle complex and diverse tasks is always a difficult problem for businesses. The greater the function and capacity, the higher the legal risk. Therefore, businesses cannot apply each risk management plan individually, but must combine many measures to achieve the highest efficiency.

First, develop appropriate internal policies and processes

Businesses need to develop a clear AI policy, including regulations on data security, legal compliance, and intellectual property rights. For example, an effective AI policy[2] can require all AI systems to undergo a risk assessment before deployment. The content of the policy includes limiting the purpose of use, transparency for customers to know when AI intervention occurs, as well as countermeasures when problems arise.

Secondly, supervision and control must take place regularly and synchronously

An AI oversight team or committee should be established to review critical AI systems. They will review the AI's output, detect unexpected issues, and check the integrity of the transaction. Applying the right technical measures can help detect signs of intrusion or fraud. Another recommendation is that businesses need to set up a tool to trace all AI queries and responses, for incident investigation and reporting to the authorities if necessary. Accordingly, businesses deploying AI need to set up full logging: when the agent accesses, with what account information, what behavior has been performed and what the results. This log also needs to be well secured so that it cannot be modified.

Retail or e-commerce companies should consider the benefits of AI shopping agents but put safety first.

Third, humans are involved in AI tasks

For sensitive tasks such as financial transactions, automated ordering of great value, AI should not be fully automated without human supervision. Each time the AI proposes a decision action, there needs to be a final confirmation from the user or employee of the business. This not only reduces errors caused by AI "delusional" prices or products, and helps businesses comply with the law and prevent legal liability if AI makes decisions that cause damage.

Fourth, incident prevention and response plans

There needs to be a clear process for AI leakage or abuse scenarios: who is the first to receive the notification, quarantine steps, information for senior management, notification to users/partners, and regulators. For example, when suspecting that a partner's AI agent has illegally infiltrated, businesses can temporarily close the service, contact the partner for handling, or conduct an internal investigation into the possibility of malware. The availability of AI response scenarios will help minimize damage and protect business reputation when controversy occurs like this case.

Recommendations for Vietnamese businesses

Businesses need to clearly define the goals of using AI and assess the corresponding risks. Retail or e-commerce companies should consider the benefits of AI shopping agents but put safety first. If applying AI, it is recommended to prioritize tested solutions and prioritize autonomous development, rather than risking relying on unverified third parties.

Before signing a contract to purchase AI services from partners, Vietnamese businesses need to check the terms of use and legal compliance. Priority should be given to partners with a reputation, transparency in data sources, and privacy policies. In the contract, it is necessary to have an intellectual property clause that clearly defines the ownership of the input data as well as the output intellectual property rights, security and privacy, as well as the applicable law clause, compensation for violations. When signing an "agentic" AI service, it is necessary to clearly commit to the scope of permitted activities and limit the provider's liability for customer data.

Vietnam's AI Law has come into effect on March 1, 2026, requiring businesses to classify AI systems, assess risks, and conduct due diligence before deployment. At the same time, it is important to note the current laws on copyright and cybersecurity regulations. In internal communication and with customers, businesses should be transparent about the use of AI to avoid legal trouble or negative communication.

When allowing AI to access internal systems, businesses need to secure their accounts. Measures such as multi-factor authentication (MFA), permission restriction, and anomalous traffic monitoring should be adopted. In addition, it is necessary to assess the security impact before AI accesses sensitive data.

AI is changing the way the world and businesses operate. Along with that, there are unpredictable legal, financial, and business risks of the enterprise, even greatly affecting the reputation of the business that has been built. Therefore, the safe and effective management of AI is a mandatory requirement for businesses to operate safely in the context of a world that is changing day by day.