AI Ethics in Business: Who Is Responsible When Systems Fail?

07/09/2026

A Fatal Misconception: AI as a Self-Running Black Box

Many operations directors I have met still believe that once AI is deployed, the system will handle everything on its own. They assume AI ethics is solely the responsibility of the technical team or the legal department. This is an extremely dangerous misconception in real-world deployment. In reality, AI has no consciousness. It merely reflects input data and programmed logic. Without a clear code of conduct, when AI makes a wrong decision, no one takes responsibility. Everyone blames each other. Leadership blames IT for incorrect configuration. IT blames operations for entering poor-quality data. Operations blames leadership for not approving the process. The result is project stagnation, diminished internal trust, and increased legal risk.

Leadership Perspective: Shaping AI Governance Culture

The greatest responsibility lies at the executive level. Leadership does not need to know how to write code, but they must understand the core limitations of the system. When implementing an AI usage policy, leaders must directly answer three questions. First, what data is permitted to be fed into the model? Second, when AI provides a recommendation, do humans have the right to veto it? Third, who bears the ultimate legal responsibility?

I once advised a multinational corporation with a supply chain spanning from Vietnam to Thailand and the Philippines. They faced a major issue when using AI to forecast inventory demand. The system continuously proposed reducing purchases for fast-moving items in major urban areas, while suggesting increased purchases for slow-moving items in rural areas. The cause was noisy historical data due to short-term promotions that had not been filtered out. If leadership views AI only as a cost-saving tool without reviewing input data, they will lose billions of dong due to forecasting errors. Here, AI ethics is not just theory. It requires leaders to proactively control inputs and verify outputs before taking action.

The Role of IT: Algorithm Transparency and Data Security

Information technology teams are often under pressure to deploy quickly. But speed does not mean recklessness. In AI governance, IT must ensure algorithmic transparency. This means they must be able to explain why the AI made a specific decision. If the system rejects a loan application or eliminates a job candidate, IT must be able to trace the factors influencing that decision. If it cannot be explained, the system should not be allowed to operate autonomously.

Another challenge is handling employee data. Many businesses use AI to analyze work performance based on emails, internal messages, or online time. This is a gray area in AI ethics. Employees' personal data should not be excessively monitored without clear consent. I always advise CIOs to separate technical data (such as system logs and server performance) from human behavioral data (such as chat content and security camera footage). Mixing these two types of data into a single AI model often leads to privacy violations and loss of trust. In projects where we partner with food and retail businesses like Masan or Meat Deli, separating this data is the first step to ensure the system does not infringe on employee privacy while still optimizing operational efficiency.

Operations Perspective: Humans in the Loop

Operations teams interact with AI directly on a daily basis. They need to understand that AI is an assistant, not a boss. The most basic code of conduct is: there must always be a human behind important decisions. This principle is also known as Human-in-the-loop.

Consider the application of computer vision in product quality control. At an instant noodle factory, an AI camera system is used to detect packaging defects. The initial accuracy rate was quite high. However, when a new packaging design was introduced, the AI began missing minor defects. If the operations team fully trusts the machine's "Pass" signal without manually inspecting random batches, defects will reach the market. The consequences are not just customer complaints but also damage to brand reputation. Therefore, the AI usage policy must clearly specify the frequency of cross-checks. Approximately one-fifth of batches should be manually inspected to recalibrate the model. This is a necessary cost to ensure reliability.

Lessons from the Multinational Market

The challenges for Vietnamese businesses do not stop at home. When expanding to Mexico or the Philippines, regulations regarding data security and AI ethics vary significantly. The GDPR in Europe sets a high standard, but markets in Latin America and Southeast Asia are gradually tightening their rules. A unified AI usage policy for the entire system may violate local laws if not adjusted. For example, storing employee biometric data in Mexico has specific requirements regarding storage location and data retention periods. If a business focuses only on technology deployment while ignoring the layer of legal and ethical governance, it is creating massive risk. Building an AI code of conduct requires the involvement of legal experts well-versed in multinational law, not just the internal IT team.

Frequently Asked Questions

Who signs off on decisions made by AI?

The ultimate responsibility always lies with the human who has authority over that process. AI only provides supporting data. It is not permitted to fully automate decisions with significant financial or personnel impact without human approval.

Should employee data be shared with third parties for AI training?

The principle is no, unless the data has been fully anonymized and there is clear consent from the employees. The top priority is building models on cleaned internal data to ensure security and compliance with personal data protection laws.

How do you measure the effectiveness of an AI ethics code?

Measure it through metrics such as the rejection rate of decisions, the number of data security incidents, and employee satisfaction when working with the system. If the rejection rate spikes, you need to review the reliability of the algorithm or the control processes.

AIVISION helps enterprises turn AI into working systems. Explore our enterprise AI solutions, read more on the AIVISION blog, or talk to our team about your own use case.