Vietnam’s hospital AI story is no longer only about digitizing paper records. Once software reads a electronic medical record and advises a clinician, the question shifts from data storage to clinical judgment.
The stakes change with that shift. A record system preserves what happened. An AI system interprets what happened and may influence what a doctor does next.
The record and the advisor are governed separately
The Law on Digital Technology Industry, passed 14 June 2025 and effective 1 January 2026, gave Vietnam its first statutory frame for the sector. The later Law on Artificial Intelligence, passed 10 December 2025 and effective 1 March 2026, carries the operating rules for AI systems used in care.
Circular 13 from the Ministry of Health pushed hospitals toward the digital record. It is about the electronic medical record, the chart itself. The newer AI layer is about software that reads that chart and offers advice on top of it.
That split matters inside a hospital EMR. A hospital can digitize charts and still face a separate set of duties for AI tools attached to those charts. One rule answers the question, “where is the record?” The other asks, “what is the software doing with it?”
This is why summaries that merge the two legal stories can mislead. The Law on Digital Technology Industry opened the statutory door by naming AI as a regulated category. Baker McKenzie describes its AI frame as principle-based, and says the general AI provisions have since been replaced by the standalone AI Law. One Asia Lawyers and Baker McKenzie both read the first law that way.
So the first law matters as the starting point. The operating questions for bedside AI now sit in the dedicated AI law: risk classification, labeling, and the duties attached to AI output.
Clinical AI belongs in the sensitive part of the risk model
The dedicated AI law sorts systems into three risk levels (high, medium, low). Baker McKenzie reports that the tier depends on impact on human rights, safety and security, field of use, scope of users, and scale of impact.
Healthcare is one of the named higher-impact fields. That is the right place for it. A prescription safety check is not a grammar suggestion. A chart summary is not a harmless autocomplete. If an output can change a treatment decision, it deserves closer scrutiny.
The grace period makes the point less abstract. Baker McKenzie reads the law as giving providers and deployers an 18-month compliance window for health, education, and finance systems, measured from 1 March 2026. Other systems receive 12 months.
That longer runway is not a free pass. It signals that health systems need more time because the regulator treats medical AI as a more sensitive case. That is the expected result for software whose output can change a treatment decision.
This is the useful way to read the risk model. It is not asking whether software uses AI in the abstract. It is asking where the output lands, who reads it, and what that person may do next.
A bedside label is only useful when the source is visible
Baker McKenzie says the Law on Artificial Intelligence requires AI-generated audio, image, and video content to be marked in a machine-readable format under government regulation. Deployers must also flag content that could be mistaken for a real event or person.
Clinical text is not the same object as audio or video. Still, the principle carries into care. The reader should know when content came from an AI system, not from a human author.
At the bedside, that is only the first half. A doctor reading an AI-produced summary or safety alert should know it came from AI. The doctor should also be able to trace the statement back to the note, rule, or reference behind it.
A label says software produced the claim. A source says what the software relied on. Without the source, the label leaves the clinician with a new kind of guesswork.
That difference sounds small until the doctor has to act. “AI-generated” tells a clinician to be cautious. “AI-generated from this note, this medication rule, and this reference” gives the clinician something to check.
Meddies treats traceability as a clinical design rule
Meddies is being designed so each answer carries the source behind it. A chart summary should point back to the underlying note. A prescription screen should point to the rule or reference it applied. A decision-support reply should show the document it drew from.
We made that choice for clinical reasons first. Doctors do not need a confident paragraph floating beside the chart. They need a claim they can verify quickly, especially when the claim could alter a decision.
That design direction also fits the intent now visible in Vietnamese AI regulation: machine output should be marked, traceable, and accountable. The wording matters here. This is design intent that points in the same direction as the law, not a compliance certification.
Meddies does not claim approval or certification it has not earned. Final compliance depends on implementing decrees and ministerial circulars still being issued, and on each hospital’s own deployment.
That distinction protects everyone. Hospitals should not buy AI tools because a vendor hints at legal safety. They should test whether the tool gives clinicians enough information to review, reject, or trust an output.
The next hospital software test is traceable judgment
The sequence is becoming clear. First, hospitals move from paper charts to a electronic medical record under Circular 13. Then the law turns to the intelligence layer that reads the record and suggests what to do.
Those are not the same project. A digital record is the filing cabinet. Clinical AI is the advisor standing beside it. A hospital can finish the filing cabinet and still need to judge the advisor.
For hospitals choosing tools now, the practical test is short. The doctor should be able to see that an answer came from AI, and the doctor should be able to see what it was based on.
That test will age well because it is clinical before it is legal. If the system cannot show its sources, it is asking the doctor to trust software at the exact moment the doctor needs to verify it.
Vietnam’s AI rules are moving toward that standard. The product work should move there first.