LLM healthcare in Vietnam: what to delegate, where to stop

01/10/2026

It is three in the morning on the internal medicine ward of a provincial hospital. The doctor on call has just finished the fourth admission of the night. The patient is stable, but a pile of paperwork is still waiting: a case summary, a transfer form, an explanation for the insurer. Nobody went to medical school for this, yet it eats a large share of the day. This is where LLM healthcare tools can help, and also where expectations run furthest ahead of reality.

At AIVISION we build Vietnamese language models, and when we talk to people in the medical field the first question is nearly always the same: what can it do, and what must it never do? This article answers both halves plainly.

What a language model can support in a hospital

Most of the value sits in dull work, not in flashy demos of a machine diagnosing patients. A language model is best at reading, writing and organising text, and medical administration is full of text.

Administrative drafting

Draft letters, meeting minutes, replies to patients about procedures, appointment reminders built on an approved template. These documents have a fixed shape, a small mistake is easy to fix, and the person who signs is still a hospital employee. The model only produces a draft for a human to approve.

Summarising records

A patient who stays for two weeks can leave dozens of pages of notes written by many hands. A doctor taking over needs the story quickly. A model can condense the notes into a timeline and point back to where each item came from in the original record. That last part is non-negotiable. A smooth summary with no traceable source is more dangerous than no summary at all.

Searching internal documents

Technical procedures, infection control guidelines, billing rules. They live in scattered PDFs and shared folders, and a new nurse often does not know whom to ask. A retrieval system built on the hospital's own documents, answering with the original passage attached, saves many repeated questions. To me it is the safest place to start.

• Drafting administrative documents from approved templates.
• Summarising long notes with sources you can check.
• Finding answers quickly in internal procedures and guidelines.
• Normalising spelling and keeping terminology consistent.

The limits a model should not cross

This part matters more than the last, so I will be more careful with it.

A language model does not diagnose. It does not prescribe. It does not suggest doses. It does not tell a patient whether or how to be treated. The reason is not that the model is always wrong. It is that when it is wrong, the answer still sounds confident. A fluent sentence is not evidence of a correct one, and in medicine the cost of that confusion is a real person.

I often use an analogy with people outside engineering. Picture a very diligent intern who has read an enormous amount and writes very fast, but has never met a real patient. You would ask that intern to draft a letter, summarise a document, find a regulation. You would not let them decide a treatment plan. A language model sits somewhere near that position.

So the design principles we follow are fairly simple:

• The doctor, pharmacist or nurse always makes the final decision.
• Any output used in clinical work is reviewed by a person before it enters the record.
• The system declines or hands off when a question touches diagnosis, prescribing or dosing.
• Answers based on documents carry their sources, and when there is no source the system says so.

Why Vietnamese makes it harder

Vietnamese medical records are a mixture: Vietnamese prose, Latin and English terms, abbreviations specific to each department, sometimes retyped handwriting. A general model reads these and easily gets them wrong, especially abbreviations whose meaning changes with the ward. That is why we focus on Vietnamese LLM work and on fine-tuning for specific domains rather than hoping one general model does everything well. On the infrastructure side, AIVISION trains on a cluster of 24 NVIDIA H200 GPUs and 8 NVIDIA B300 GPUs, and has released the L1.0 language model for Vietnamese along with the E1.0 speech to text model. How well any of this performs on a particular medical task has to be measured on the deploying organisation's own data, and I will not quote a number before such a measurement exists.

Rolling it out without pain

Teams that have done this before tend to say the same thing: do not start with the big problem. Pick a workflow that is heavy on text and low on clinical risk, such as drafting transfer forms or searching internal procedures. Let a small group use it, log every time the model gets something wrong, and only then widen the scope.

Two things are routinely underestimated. One is whether clinicians actually have time to review. If the tool produces a draft but checking it takes longer than writing from scratch, you have only added work. The other is data. Medical records are sensitive, so where the system runs, who can see what, and whether anything is stored must be settled before anyone discusses model quality. That deserves its own article, and we will write it.

One more point that gets skipped: someone has to own the system after launch. Models do not degrade by themselves, but procedures change, forms change, departments pick up new vocabulary. If nobody looks again, quality drifts and nobody notices.

A practical way to think about it

If I had to put it in a single sentence: use LLMs to give time back to the people doing the clinical work, not to replace their judgment. The doctor on call at the start of this article does not need a machine that can diagnose. She needs someone to write the transfer form so she can sit down for a sip of water. It is a modest goal, but reaching it is genuinely useful, and it falls within what the technology can do responsibly today.