XAI: Why Your AI System Is Rejecting Transactions

10/09/2026

Why do regulators mandate that AI rejections be explainable?

Many business owners ask us this when preparing to deploy credit scoring or order filtering systems into production. The short answer is: legal liability cannot be attributed to a black box. When a transaction is rejected, customers have the right to know why. Regulators, particularly in the financial and insurance sectors, require evidence that decisions are not based on discriminatory factors such as gender or region.

In Vietnam, regulations regarding information security and personal data protection are becoming increasingly stringent. You cannot tell an inspector that "the computer just decided that." You need a clear logical path from input data to output results. This is where Explainable AI (XAI) becomes a mandatory requirement rather than a redundant feature. We have seen many projects stall at the acceptance stage simply because they could not generate explanations for each rejected case. Algorithmic transparency is not just about passing inspections; it is the foundation for building long-term trust with partners and customers.

Before writing a line of code, what do we need to agree on?

This phase is often overlooked but is the most important. Many technical teams jump straight into selecting the most complex neural network model without asking: "Who will read this explanation?" If the end user is a customer support agent, they need a concise answer like "Due to more than three late payments in the past six months." But if the reader is an auditor, they need to see the weight of each variable.

The trade-off here is: the more detailed the explanation, the more the model tends to be simplified. You cannot have the 99.9% accuracy of an extremely complex model while also clearly explaining every step. At AIVISION, when working with partners in the retail and manufacturing sectors, we typically start by defining the risk acceptance threshold. For example, with an order fraud detection system, we accept slightly lower accuracy in exchange for the ability to clearly indicate that an order was blocked because the IP address matched a known fraud case, rather than due to some vague correlation. This significantly reduces the workload for the exception handling team later on.

During deployment, how does XAI work in practice?

Once the appropriate model is selected, applying XAI does not mean adding a superficial layer. It must be embedded in the core of the development process. We typically use techniques such as SHAP (SHapley Additive exPlanations) or LIME for analysis. It sounds technical, but what is the practical significance? Imagine you are an operations manager for a supply chain. The AI system forecasts raw material demand and proposes a 20% reduction in inventory for the Southern warehouse.

• Without XAI: The system simply returns the number "Reduce by 20%."
• With XAI: The system displays that "The main reason is that the average temperature this week is 5 degrees higher than forecast, leading to reduced consumption of hot items, combined with a competitor recently cutting prices by 10% in this region."

This explanation is generated automatically and linked to each system decision. It is not only displayed for humans but also logged for traceability. In a recent project on manufacturing data analysis, we helped the technical team discover that a temperature sensor was reporting incorrect data, causing the AI to issue false alarms. Without the ability to explain, they might have irrationally changed their entire operational process. XAI acts as a truth filter here, helping humans trust the computer more, but also helping humans catch the computer when it is wrong.

When the system is live, what happens that surprises you?

Once operational, the biggest issue is no longer technical, but human. Initially, employees tend to ignore explanations if they disagree with the AI's outcome. They still rely on intuition to make decisions. This is a cultural barrier that many businesses in Vietnam and neighboring markets like Thailand and the Philippines are facing when deploying AI at scale.

However, after about 3 to 6 months, when they see that cases rejected by the AI are frequently identified as actual fraud, or that AI alerts lead to significant cost savings, trust begins to shift. At this point, the explanation becomes a training tool. Instead of having to verbally explain to new employees, the system automatically provides the reason why a transaction was accepted or rejected. This creates operational consistency that was previously nearly impossible to achieve. We have noticed that multinational companies operating in Southeast Asia have more frequent and detailed explainability requirements than purely domestic enterprises, due to the need to comply with international standards. This forces the system to be designed from the outset to be audit-ready at any time.

What remains unresolved and what do we still need?

Although XAI has advanced rapidly, it is not yet perfect. Currently, we can only explain "why" based on existing data. We cannot explain "why this data is the way it is" if the data source was biased from the start. For example, if historical data contains invisible bias, the AI model will learn and replicate that bias, and XAI will faithfully explain that bias as a "valid reason" because it is present in the data. This is an ethical paradox that current technology has not fully resolved.

In addition, the computational cost of running explanation algorithms for millions of transactions daily remains a significant burden. We are working with partners like Masan and other food and beverage companies to find a balance between processing speed and the depth of explanation. In the near future, we may need clearer industry standards regarding the minimum level of explanation required. Currently, we still recommend that businesses start with high-risk cases, where the consequences of a wrong decision are most severe, and gradually expand the scope. The journey toward algorithmic transparency has no end; it is a continuous improvement process that requires close coordination between technology teams and business units. If you are preparing for upcoming inspections, building a system with clear explainability is not only about protecting your business but also a competitive advantage when partners demand higher transparency in the shared supply chain.

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.