What should enterprises do to manage risks arising from employees’ use of AI?

The digital era that forces businesses to accelerate innovation to maintain a competitive advantage is creating a major polarization within organizations. On the one hand, sales and product development teams are under pressure to apply AI to optimize performance; The other side is the risk management teams (technical and legal teams) that strive to prevent incidents that cause loss of trust or ensure compliance with the law. This interference leads to a common but dangerous phenomenon: "Shadow AI" – the arbitrary use of AI tools by employees without the approval or supervision of the IT department[1].

Source: USAII

1. The Nature of Shadow AI and Risk Amplification

Shadow AI emerges when employees use user-level AI applications such as ChatGPT, coding aids, or design aids to handle specialized work without going through a formal procurement or moderation process. In this scenario, businesses have almost no visibility into data management, lack of policy commitments from suppliers, and employees are often unaware of the security risks that come with it.

AI is not just a new tool; It serves as an amplifier. If a business already has gaps in data governance or privacy, the use of AI will exacerbate the impact of those gaps many times over. The risks in this context are enormous.

2. Core data privacy and security risks

When employees feed business data into public AI models, they are inadvertently creating serious legal and technical challenges in the safe and sustainable operation of the business. In many respects, the risks can be:

Risks in the intended use of data

General-purpose AI models often use input data in ways that the original consumer or data owner doesn't agree with. Employee uploading customer data to AI for analysis may violate the principle of using the data only for the intended purpose, leading to legal compliance risks according to the respective legal framework in each country. Especially in Vietnam, with the new legal framework to protect privacy from the 2024 Data Law, the 2025 Artificial Intelligence Law and the 2025 Data Protection Law, it has set strict and closed compliance requirements. These regulations force businesses to use customer and user data for the purpose of initial consent. Otherwise, there is a great risk that businesses will be subject to sanctions for these violations.

The right to be forgotten

AI is a technology that requires data, and machine learning models, once trained, are difficult to "forget" or delete a specific piece of information without greatly affecting the performance of the entire system. If sensitive data is leaked into a third-party vendor's training process through Shadow AI, the business almost loses the ability to enforce the customer's right to delete data. Obviously, this is the downside of this technology that is difficult for humans to intervene. At this time, the right to be forgotten on technology platforms is almost impossible to enforce and mechanical intervention by humans.

Transparency and the "Black Box" effect

Modern AI models, especially agentic systems, operate in a non-linear manner[2], making it extremely difficult to explain why a particular decision is made. This lack of transparency creates skepticism from customers and makes it difficult for businesses to correct errors or explain to regulators when there is an incident of data leakage to the outside.

3. Autonomous magnetic risk and false accumulation

The shift from traditional AI to Agentic AI[3] – systems capable of independently inferring, planning, and executing action steps independently via APIs – is increasing the level of risk.

Real-time skewed cloning: AI models are often based on historical data. If this data has biases such as gender or ethnic biases, those biases multiply rapidly as the system processes thousands of transactions per minute in a real-time environment. An employee who uses AI to score credit or screen applicants can inadvertently create a large-scale ethics and reputational crisis in a short period of time.

Loss of action control: Unlike fixed-rule-based automation flows, agent AI is adaptive and can make unexpected human decisions. Without "human-in-the-loop" controls, these systems can take actions that harm an organization's assets or reputation.

Source: LexisNexis Canada

4. Risk of suffering strict sanctions from regulatory agencies

One of the biggest legal risks that global businesses may face is sanctions called "model disgorgement" or "algorithm destruction".[4]

When a business is found to be using data illegally or without consent (often as a result of employees arbitrarily putting the data into AI training), regulatory agencies such as the U.S. Federal Trade Commission (FTC) have the right to require the business to not only delete the infringing data, but also to destroy the entire model or algorithms built from that data.

Having to destroy a model that has cost millions of dollars and years to train is a huge financial and strategic loss. Precedents such as the Everalbum or Ring[5] cases  have shown that the FTC is increasingly aggressive in applying this measure as a "stick" to punish data abuses.

Although in Vietnam, the newly promulgated Law on Artificial Intelligence 2025 does not provide for this sanction, the guiding documents for implementation in the near future can completely choose this measure to be implemented in order to bring effectiveness to the law in practice and also limit the great potential risks brought by this technology.

5. Risk management strategy and action framework for businesses

To reconcile the need for innovation and safety requirements, organizations need to shift from a static control model to flexible governance:

Businesses need to proactively and actively respond to data risks

Moving from a privacy impact assessment (PIA) to a Data Enablement Plan (DEP) is a necessity for businesses at this time. Instead of passive compliance checklists, businesses should adopt a "Data Activation Plan" to proactively respond. DEP integrates privacy, ethics, security, and AI risk into a single process, focusing on "how to enable data securely" rather than single, discrete plans.

Establishing an ethical and secure architecture from the start

Businesses need to build automated technical barriers to prevent data leaks. Technical solutions businesses should consider as follows:

• Data encryption and masking: Automatically filter out personally identifiable information before it is fed into AI processing streams.
• Technical departments need to warn and prevent risks: Teams of technical experts should simulate adversarial attacks to find vulnerabilities in AI systems before deploying them widely. This is a necessary technical measure as a "test" to secure data against risks in the actual use of personnel.
• Develop and operate effective AI governance policies: Use AI governance tools to track and enforce privacy policies in real-time. This is a technical solution often used by global businesses to manage risks in the process of business operations in the new context.

Improve digital competence and sense of responsibility

Ultimately, the risk from Shadow AI can only be thoroughly addressed through improving the level of understanding of AI and data for the entire workforce. Employees need to be trained to understand that responsible use of AI is not just a matter of compliance, but also a core principle for maintaining long-term business value. The development and operation of clear and transparent internal policies on AI by enterprises to guide personnel to use AI safely and effectively is an effective way for businesses to manage the risks of data disclosure and leakage from the use of AI by employees.

The use of AI by employees is an irreversible trend that offers tremendous potential for productivity. However, without a strict governance framework, "Shadow AI" will become a legal slow-detonating bomb. Businesses that succeed in the AI era will be those that know how to turn privacy and data governance into the foundation of trust, thereby creating a launching pad for innovative and sustainable innovation in the new era of humanity – the dominant technology era.