Vietnamese LLM for Banking and Finance: Uses and Risks

01/10/2026

Picture a credit officer receiving a loan file of two hundred scanned pages: statements, employment contracts, property papers, an internal proposal. The first thing they do is not underwriting. It is working out what the file actually says. That is where a Vietnamese LLM starts to help in banking. It is also, if done carelessly, where the biggest risks appear. This piece covers both sides, from the point of view of someone who builds products rather than sells them.

Three kinds of work an LLM can do

Answering customers

The most visible use. Customers ask about fee schedules, card opening procedures, how disputed transactions are handled. A chatbot built on an LLM and connected to the bank's official documents can answer in natural Vietnamese and cite its sources. The word "cite" matters. An answer about an interest rate that cannot point to the original document should not be sent.

The chatbot also has to know when to stop. A customer asking "should I borrow more to invest?" is a question the system must politely decline and hand to a person, because that is personal financial advice, not information lookup.

Summarizing files and internal documents

Back to the loan file. An LLM can extract key fields, summarize the proposal, and flag contradictions between documents: company names spelled differently, income on the employment contract that does not match the statement. It also helps with internal reports, meeting notes, and regulations that keep changing. Here the LLM is a fast reader, not a decision maker.

Internal support

New employees often spend a long time finding which procedure lives in which document. An assistant that searches internal rules in natural language shortens that. The risk is lower than customer-facing use, so many organizations begin here.

Risks you cannot wave away

A bank is not an online shop. A wrong answer can become a complaint, a lawsuit, or a penalty from a regulator. Here is what needs controlling, in the order I find them most painful.

• Hallucination: an LLM can confidently state a fee or condition that does not exist. Mitigation is to restrict the model to supplied documents, require citations, and refuse when there is no basis. This reduces the problem, it does not eliminate it.
• Customer data leakage: financial data is among the most sensitive there is. Decide clearly where the model runs, whether data leaves the bank's infrastructure, and how conversation logs are stored.
• Prompt injection: a customer or a malicious document can embed instructions that make the model ignore its rules. For any system that reads outside documents, this is a practical risk, not a theoretical one.
• Bias and fairness: if a model takes part in credit assessment, you must test whether it treats groups of customers differently. Personally, I think credit decisions should not be handed to an LLM. Use it to prepare information for the person who decides.
• Explainability: when a regulator or a customer asks why an answer was given, you need to trace it to the documents it was based on.

Why Vietnamese, and why finetune

Banking language has its own texture. Terms such as outstanding principal, collateral and overdraft limit have precise meanings, and regulatory documents are written in a dense administrative style. Large multilingual models handle everyday Vietnamese reasonably, but for domain style and terminology, finetuning an LLM on Vietnamese data from the field usually gives a better fit. That is what AIVISION does: we train and finetune LLMs for Vietnamese and for specific domains, on our own NVIDIA H200 and B300 clusters, and we have released L1.0, an LLM for Vietnamese. For finance I am not claiming any specific result. I am saying this is the direction we consider sensible.

A cautious path

If I were advising a bank, I would go from an internal assistant, to document summarization with human review, and only then to a customer-facing chatbot with a narrow scope. Each step gets its own test set, including trap questions that deliberately fall outside scope. At every step the system is a support tool only: it is not investment advice and it does not replace a specialist's judgment.

You can read more about AIVISION's approach to language models for Vietnamese on our website.