Clinical AI as Organizational Change in Viet Nam: How Healthcare Organizations Should Assess and Redesign Workflows, Roles, Escalation and Accountability for Safe Implementation
IMPLEMENTATION INSIGHTS
Clinical AI as Organizational Change in Viet Nam: How Healthcare Organizations Should Assess and Redesign Workflows, Roles, Escalation and Accountability for Safe Implementation
IMPLEMENTATION INSIGHT
Last reviewed: 6 October 2026
Healthcare organizations sometimes talk about clinical AI implementation as though it were primarily a technology deployment: select a system, validate it, integrate it with existing information systems, train users and go live. In practice, introducing AI into care can change much more than software. It can change who sees information, when decisions are made, which tasks require verification, where exceptions are handled, how clinicians communicate, who carries additional workload and what happens when people disagree with the system.
Clinical AI should therefore be understood as an organizational change intervention, not simply an IT implementation.
A deterioration model illustrates the point. Before AI, a nurse may recognize a change in a patient's condition, contact a physician and initiate an established escalation pathway. After AI is introduced, an alert may appear earlier, reach a different person, create a new expectation for response time or generate additional documentation. The organization has not merely added a prediction; it has changed the flow of information, attention and responsibility.
An ambient AI scribe creates a different kind of change. It may reduce typing but introduce new work involving audio capture, patient communication, review of generated documentation, correction of errors and decisions about what happens when the system is unavailable. An imaging prioritization system may improve the order in which studies are reviewed while also changing queues, handoffs and expectations elsewhere in the service.
International approaches increasingly reflect this organizational perspective. WHO/Europe's 2026 report on implementation insights for AI in health care focuses on translating principles into practice in real healthcare environments [7]. Joint Commission's voluntary Responsible Use of AI in Healthcare certification similarly recognizes that AI is becoming embedded across clinical, operational, administrative and care-support workflows and frames responsible AI as a patient-safety, quality, governance, privacy and trust issue rather than a technology issue alone [8]. CHAI's 2026 AI Governance Playbooks are likewise intended to help healthcare organizations integrate responsible-AI controls into existing organizational processes and adapt them to local context [9].
For healthcare organizations in Viet Nam, the implementation question should therefore move beyond “How do we integrate the AI into our systems?” toward “What will this AI change about how care is actually delivered, and how must the organization adapt so that the resulting system remains safe, accountable and workable?”
Assess the workflow before redesigning it
AI implementation should begin by understanding the current workflow rather than immediately designing a future workflow around the technology. Formal process maps are useful, but actual clinical work often contains variation that policies and procedures do not fully capture. Different shifts may handle escalation differently. Experienced staff may rely on informal checks that have never been documented. A specialty clinic may depend heavily on one coordinator to manage exceptions. Two hospitals in the same group may use different information-system configurations or have different downstream capacity.
Before introducing AI, organizations should therefore examine work as it is actually performed. Who initiates the process? What information is available at each point? Where do delays occur? Which handoffs matter? What workarounds already exist? Which tasks depend heavily on individual experience? What happens at night, on weekends or during staffing shortages? Which existing controls currently prevent error?
This matters because AI can unintentionally bypass, weaken or overload an existing safeguard. A technology designed to automate one step may simply shift work elsewhere. A faster diagnostic output may deliver little benefit if the next clinical step remains constrained. An alert intended to accelerate escalation may instead create alert fatigue when added to several existing notification systems.
The purpose of assessment is not to preserve current practice simply because it already exists. It is to understand what should be preserved, what genuinely needs to change and what new risks the proposed change could introduce.
Define exactly what role the AI is intended to play
Safe organizational redesign becomes difficult when the AI's role is vague. Healthcare organizations should be able to describe what the system is expected to do in operational terms. Is it prioritizing patients, drafting documentation, predicting deterioration, recommending an action, summarizing information, identifying cases for further review or communicating directly with patients?
The organization should also define what the AI is not intended to do. A system validated to prioritize imaging studies should not gradually become a diagnostic substitute because users find its outputs persuasive. A generative assistant introduced for drafting may begin influencing clinical reasoning if staff routinely accept its interpretation without independent review. A prediction model may effectively become a treatment trigger if local practice evolves around it even though that was never the approved or validated use.
This distinction between the intended role of the technology and the role it acquires in routine practice deserves explicit monitoring. The technical system may remain unchanged while the organization gradually expands what it expects the AI to do.
For each material use case, the organization should therefore be able to answer a practical sequence: What information enters the system? What does the AI produce? Who receives the output? What decision or action can follow? Who reviews or verifies it? What happens if the output is uncertain, delayed, wrong or absent?
Those questions frequently reveal that the central implementation challenge is not the algorithm itself but the care system surrounding it.
Map the new decision points and handoffs
AI often creates new decision points even when it appears to automate work. Consider a system that identifies patients at increased risk of deterioration. A score may be generated automatically, but somebody still needs to determine what that score means operationally. Does every high-risk alert require review? Who reviews it? Within five minutes, thirty minutes or an hour? Can a nurse initiate escalation directly, or must a physician first confirm the concern? What happens when clinical judgment conflicts with the AI? What happens when several alerts arrive simultaneously?
Each of these questions represents a possible handoff, delay or failure point. Healthcare organizations should therefore map the future workflow from end to end, including both normal operation and exception pathways. A useful implementation map is rarely just data → AI → clinician. It is more often:
trigger → AI output → human interpretation → decision → action → handoff → outcome → feedback
Every transition can fail. Information may not reach the intended person. A user may assume that someone else has acted. Escalation may stop at a departmental boundary. An AI recommendation may increase tests, referrals or reviews beyond the capacity of downstream services.
Making these dependencies visible before go-live is one of the most practical patient-safety functions of workflow redesign.
Redesign roles, not only interfaces
Organizations sometimes redesign an AI interface while leaving the surrounding responsibilities implicit. That is risky because AI can blur existing professional and operational boundaries without formally changing anyone's job description.
If AI identifies a potential medication problem, is the physician expected to review it, the pharmacist or both? If a generative tool drafts clinical documentation, who is responsible for verifying factual accuracy before the record becomes final? If an imaging system changes prioritization, who manages a case that appears inconsistent with the patient's clinical condition? If an automated patient communication requires clinical interpretation, who owns the follow-up?
The organization should distinguish who uses the system, who owns the clinical pathway, who operates the technology, who manages relevant data and privacy risks, who monitors safety, who investigates incidents, who communicates with the provider and who has authority to restrict or suspend use. A RACI-type matrix or similar responsibility map can be useful for operational clarity, but it should not be mistaken for a legal allocation of liability. Internal role assignments cannot override duties established by law or applicable professional obligations.
This distinction is particularly important under Viet Nam's current AI framework. Law No. 134/2025/QH15 has been effective since 1 March 2026 and provides for tighter risk management of AI used in sectors directly affecting life, health and lawful rights and interests. In health care specifically, the Law refers to patient safety, reliability under real-world conditions of use and protection of health data [1].
For high-risk AI systems, Article 14 of the Law places defined responsibilities on both providers and deployers. Providers must, among other duties, design systems so that human supervision and intervention remain possible. Deployers have responsibilities relating to safe operation, monitoring, risk control, data protection, human intervention and coordination with providers within the scope established by law [1]. Decree No. 142/2026/NĐ-CP, effective since 1 May 2026, further operationalizes lifecycle risk-management responsibilities for these actors [2].
AI therefore does not create an organizational vacuum in which responsibility simply moves from healthcare professionals or healthcare organizations to an algorithm.
Human oversight has to work inside the workflow
Human oversight is often described too simply: “The clinician remains in control.” That statement may be formally true while being operationally meaningless.
A person cannot meaningfully oversee an AI system if the output appears after the relevant decision has already been made, if the user lacks sufficient information to question it, if workload makes independent review unrealistic, if overriding the AI requires excessive effort or if organizational culture treats disagreement with the system as a user error.
Meaningful oversight therefore requires practical conditions. The person needs sufficient information to understand the output, enough competence to interpret its limitations, realistic time and opportunity to intervene, and authority to change, reject or escalate the resulting action.
For high-risk AI systems, these principles are reflected directly in Viet Nam's regulatory framework. The AI Law requires provider-side design that preserves the possibility of human supervision and intervention and places corresponding operational responsibilities on deployers [1]. Decree No. 142 further embeds human supervision and intervention within lifecycle risk-management arrangements for high-risk systems [2].
Healthcare organizations should therefore test human oversight under realistic conditions rather than assume it exists because a policy says that a human is “in the loop.” Can users recognize when AI is operating outside its intended conditions? Can they override it? Do they know what happens after an override? Can an unresolved concern be escalated? Can safe care continue without the AI?
The necessary solution may involve workflow, interface design, staffing, training or escalation rather than changes to the model itself.
Escalation pathways should be designed before something goes wrong
Clinical AI introduces exceptions that existing escalation processes may not handle clearly. A system may generate an implausible output. A clinician may disagree with an AI recommendation but remain uncertain whether the disagreement reflects a model problem or an unusual patient. Several staff members may independently notice the same pattern. A system may generate substantially more alerts after a vendor update. One site may experience a problem that other sites have not yet detected.
These situations require an escalation architecture. The first level may be ordinary clinical handling, in which the user disregards or overrides an output and proceeds with appropriate care. Repeated concerns may need review by a local clinical or operational owner. A suspected systemic problem may require involvement of quality and patient safety, AI governance, IT, privacy, cybersecurity or the provider. Some circumstances may justify immediate restriction or suspension.
An escalation pathway should therefore answer more than who to contact. It should define what triggers escalation, what information should be captured, who is authorized to make the next decision and how quickly a response is expected.
Formal regulatory reporting operates at a different threshold. The AI Law defines the concept of a serious AI incident, while Article 19 of Decree No. 142 establishes the reporting and handling pathway for events that meet that statutory definition [1,2]. Internal clinical escalation should operate well below that threshold. A mature healthcare organization should be able to investigate repeated errors, unusual outputs, emerging patterns and near misses before an event becomes a serious reportable incident.
AI may reduce one workload while creating another
Efficiency claims require careful interpretation because AI may reduce one form of work while creating another elsewhere in the system. An ambient scribe may reduce typing while creating review and correction work. A prediction model may reduce manual screening while generating alerts that require assessment. Automated patient communication may reduce drafting time while creating more clarification requests. AI triage may shift work from physicians toward nurses or administrative teams.
The useful question is therefore not simply “How much time did the AI save?” but “How did the distribution, complexity and timing of work change across the healthcare system?”
Recent implementation literature reinforces this sociotechnical perspective. A 2026 narrative review by Wenderott, Krups and Weigl included 26 studies and identified 55 dimensions of facilitators and barriers to healthcare AI implementation. Among the recurring factors were efficiency, compatibility with local IT infrastructure, stakeholder involvement, transparency and clinician trust, with many implementation issues concentrated in organizational and technology-related aspects of the work system [10].
Evidence from real-world generative AI remains much earlier. A 2025 systematic review by Artsi and colleagues identified only four eligible empirical studies of LLM deployment in real clinical workflows. The reviewed applications showed promising effects on measures such as operational efficiency, user satisfaction and workload, but effects remained context-dependent, and the authors highlighted limitations involving generalizability, regulatory barriers, interoperability and insufficient post-deployment monitoring [11].
These findings support treating workflow design, organizational capability and human factors as core parts of AI implementation rather than issues to be addressed only after technical deployment.
Healthcare organizations should therefore monitor hidden work: verification time, documentation corrections, alert handling, additional handoffs, explanations to patients, exception management, troubleshooting and workload transferred to teams that were not expected to be major users of the technology. A system that saves time for one role while creating substantial fragmented work for several others may not represent a true organizational improvement.
Training should follow changed responsibilities
AI training should not be treated as a generic digital-literacy exercise. Different roles require different competencies because AI changes different parts of their work.
A clinician using AI-supported decision-making may need to understand intended use, major limitations, uncertainty and override procedures. A nurse receiving alerts may need to know when and how to escalate them. Quality personnel need to recognize AI-related patient-safety signals. IT teams need to understand dependencies and version changes. Managers need to understand which evidence should trigger review or restriction. Procurement teams need to recognize which provider commitments are operationally important.
Training should therefore be designed around what each person is expected to do differently because AI is present.
Joint Commission's RUAIH certification includes education and training within its organizational standards, alongside governance, data management, risk and bias reduction, and monitoring, evaluation and validation [8]. CHAI similarly places education, training and feedback within a wider governance architecture that also includes organizational structures, resources, lifecycle management and third-party management [9].
Competency should also be revisited when a system changes, when new employees begin using it or when the technology is introduced into a new clinical environment. One training session before launch is not a substitute for sustained implementation support.
Safe fallback should be part of the redesigned workflow
An AI-enabled workflow should remain safe when AI does not function as expected. A system may become unavailable because of cloud failure, network disruption, cybersecurity incidents, software problems, provider downtime or planned maintenance. It may remain technically available while producing degraded outputs because input data have changed or an upstream integration is failing.
Healthcare organizations should therefore define what happens when AI is unavailable or unreliable. Does the pathway revert to the previous process? Is that process still resourced and understood? Who decides that the AI should no longer be relied upon? How are users informed? Can accumulated work be managed safely when service resumes?
Fallback becomes particularly important when staffing or workflow has been redesigned in reliance on AI. A manual process that exists only on paper but is no longer practiced or adequately resourced is not a reliable fallback.
Responsible implementation therefore requires enough organizational resilience to maintain safe care without assuming continuous AI availability.
Test the redesigned workflow before routine use
Testing an algorithm and testing a clinical workflow are not the same thing. Broader deployment should therefore include testing under realistic organizational conditions, which may involve simulation, shadow deployment, controlled pilots, structured observation or staged implementation depending on the use case and level of risk.
Testing should examine routine operation but also foreseeable difficult conditions: competing clinical priorities, high workload, missing data, uncertain outputs, delayed responses, disagreement between clinician and AI, system downtime, atypical patients and escalation that crosses departmental boundaries.
The objective is not merely to determine whether users can operate the software. It is to determine whether the organization can safely absorb what the software changes.
A useful pilot should therefore produce organizational evidence in addition to technical performance evidence. Did responsibilities remain clear? Did the system create unexpected workload? Could users override it safely? Did escalation work as expected? Did downstream services have enough capacity? Did nights, weekends or different sites behave differently? What workarounds appeared? Which assumptions proved wrong?
Moving from pilot to routine use should therefore be a governance decision, not an automatic next step after technical success.
Workflow redesign can create regulatory change-control questions
Organizational change and regulatory change control are connected, but the responsibilities of providers and deployers need to remain distinct.
Under Viet Nam's AI framework, a deployer may rely on the provider's original risk classification. However, where the deployer modifies, integrates or changes the functions or purpose of use of the system in a manner that creates a new risk or higher risk level, the deployer must coordinate with the provider to review and reclassify the system [1,2].
For high-risk AI systems, Article 13 of Decree No. 142 separately requires the provider to reassess conformity when a significant change affects the basis of the original conformity assessment. Relevant changes can include material changes to key functions, intended use or application scope; system architecture, AI model or major technical configuration; principal data sources or processing methods; or integration into another system or operating environment that creates new risks or changes previously assessed operating conditions [2].
This distinction matters for healthcare organizations. A hospital changing a workflow does not automatically become responsible for the provider's conformity reassessment, and not every local workflow adjustment triggers statutory reassessment. However, a material change in use, integration, functionality or operating environment may require the organization to review whether previous classification, validation and safety assumptions remain valid and whether coordination with the provider is necessary.
A practical governance question after any important change is therefore: Which assumptions supported the previous approval, and are those assumptions still true?
Workflow redesign also changes data flows
Organizational change often creates new data relationships. A redesigned AI pathway may provide access to additional staff, transfer information to a cloud service, introduce audio capture, generate inferred data, involve new third parties or combine datasets that were previously separate.
Viet Nam's Law No. 91/2025/QH15 on Personal Data Protection and Decree No. 356/2025/NĐ-CP have both been effective since 1 January 2026 [3,4]. Healthcare organizations should therefore reassess relevant data flows and applicable legal requirements when AI materially changes how personal data are collected, accessed, processed, stored, shared or transferred.
This is particularly important when an apparently small workflow change modifies who receives patient information, whether data leave the organization's direct environment or whether an external AI platform becomes part of the care pathway.
Workflow redesign should therefore not be separated from data governance, privacy and cybersecurity review.
AI implementation should connect with existing quality and patient-safety systems
Healthcare organizations do not need to create a second quality-management system exclusively for AI.
Circular No. 19/2013/TT-BYT remains recorded as in force and provides an existing framework for hospital quality management in Viet Nam, including organizational responsibilities, quality indicators and continuous improvement activities [5]. Circular No. 43/2018/TT-BYT provides a framework for reporting, analyzing, responding to and preventing medical incidents and expressly recognizes near misses. The national legal database currently records Circular No. 43 as partially expired, so individual provisions should be applied according to their current legal status and together with later applicable law [6].
AI-related workflow problems should enter these existing organizational systems where relevant. If an alert repeatedly fails to reach the right clinician, that may become a patient-safety concern. If staff create an unsafe workaround because an AI pathway is too burdensome, that is more than an adoption problem. If AI-generated documentation repeatedly leads to medication clarification, the pattern may represent a quality signal.
The goal should be to integrate AI into routine organizational quality, safety and improvement, rather than leaving AI governance isolated inside IT or an innovation team.
Measure whether the new workflow actually improves care
High adoption is not the same as successful implementation. A healthcare organization may achieve very high use of an AI system while creating unnecessary investigations, additional workload or new safety risks. Conversely, selective use may be appropriate when clinicians understand when a technology adds value and when it should not influence care.
Implementation evaluation should therefore connect several levels: AI behavior, human response, workflow performance, staff burden, patient experience, clinical outcomes and unintended consequences.
For a deterioration system, relevant questions may include whether alerts are acknowledged, how quickly patients are reviewed, whether escalation occurs appropriately and whether meaningful clinical outcomes improve. For an ambient scribe, organizations may examine correction burden, documentation quality, clinician time and patient experience. For imaging prioritization, the relevant question may be whether urgent patients actually receive faster care without creating unacceptable delays elsewhere.
The objective is not simply to prove that an AI tool works. It is to determine whether the healthcare system works better with the AI embedded within it.
Multi-site deployment requires both standardization and local adaptation
Organizational change becomes more complex when the same AI system is deployed across multiple hospitals or clinics. Central governance can establish common requirements for evidence, human oversight, monitoring, escalation, incident management and provider or vendor management, but local workflows may differ substantially because of staffing, patient populations, equipment, information systems, referral pathways and operational capacity.
A centrally designed workflow should therefore not be assumed to function identically everywhere. Before scaling, organizations should identify which controls must remain standardized and which parts of implementation require local adaptation. Minimum patient-safety controls should not disappear because a local team prefers another workflow, but central governance should also avoid imposing an operational design that does not reflect how care actually works at a particular site.
The strongest multi-site model combines common governance with visible local variation.
A practical Clinical AI Organizational Change Cycle
Healthcare organizations do not need to adopt a complicated change-management methodology before they can begin. A practical implementation sequence is:
understand current work → define the intended AI role → map the changed workflow → assign roles and decision rights → design oversight, escalation and fallback → test under realistic conditions → implement with active support → monitor real-world effects → redesign, restrict or scale
Understanding current work prevents the organization from designing around assumptions that are not true. Defining the AI's role establishes boundaries. Mapping the changed workflow makes new handoffs and dependencies visible. Assigning roles clarifies operational responsibility. Oversight, escalation and fallback prepare the organization for disagreement and failure rather than treating them as unexpected events.
Testing then evaluates the organizational system, not only the technology. Go-live becomes an active implementation period rather than the end of the project. Real-world monitoring shows whether technical performance, workflow, safety and outcomes remain acceptable. What the organization learns then informs the decision to redesign, expand, restrict, suspend or stop the use of AI.
This Clinical AI Organizational Change Cycle is proposed here as a practical healthcare implementation framework. It is not a statutory framework under Vietnamese law, a regulatory conformity-assessment method or a validated organizational change model.
Sometimes the safest redesign is not to deploy the AI
Organizational redesign should not become an exercise in forcing a healthcare system to accommodate a technology that does not fit.
If meaningful human oversight cannot be achieved, downstream capacity is inadequate, workflow burden becomes excessive, accountability remains unclear or safe fallback cannot be maintained, the appropriate governance decision may be to delay deployment, narrow the use case or choose another solution.
The same principle applies after go-live. If real-world evidence repeatedly shows that expected benefits do not materialize or that new risks outweigh those benefits, the organization should be willing to restrict, suspend or withdraw the system.
Previous investment in money, time or reputation should not determine patient-safety decisions. Responsible implementation means being willing to change the organization around AI when doing so improves care, but also being willing not to change care around a technology that cannot be integrated safely.
What this means for healthcare organizations in Viet Nam
Clinical AI implementation should not begin and end with installation, integration and training. It should be treated as a controlled change to the healthcare system.
Before go-live, leaders should understand what the AI will change about information flow, workload, decision-making and escalation. Responsibilities should be explicit. Human oversight should be operationally realistic. Exception and fallback pathways should already exist. Users should be prepared for the responsibilities they actually hold, not simply taught how to operate the software.
After go-live, the organization should examine whether work is occurring as expected. Are staff developing workarounds? Has workload shifted unexpectedly? Can clinicians challenge the AI when necessary? Are escalations reaching the right people? Have delays emerged elsewhere in the care pathway? Are quality and patient-safety signals changing? Are expected benefits becoming visible in outcomes that matter?
These are organizational-change questions, but they are also healthcare quality and patient-safety questions. The practical transition is therefore from installing AI → to understanding how AI changes work → to deliberately redesigning the care system → to monitoring whether that redesign actually improves care.
That is the difference between deploying an AI product and implementing clinical AI responsibly.
Professional education notice
This article is intended for professional education and knowledge exchange. It does not constitute individualized legal, regulatory, clinical, human-factors, cybersecurity, data-protection, accreditation or quality-management advice.
The workflow, role, escalation, accountability and organizational-change approaches described in this article are practical healthcare implementation concepts. They are not presented as universal statutory requirements under Vietnamese law.
Operational assignment of responsibility within a healthcare organization does not itself determine legal liability or alter responsibilities established by applicable law. Requirements relating to high-risk AI systems, risk classification, conformity assessment, significant changes, human oversight, serious incidents, personal-data protection, medical devices and other regulated activities should be assessed according to the characteristics and deployment context of the particular system.
The Clinical AI Organizational Change Cycle proposed in this article is a practical synthesis for organizational implementation and improvement. It is not an official Vietnamese regulatory framework, a WHO standard, a Joint Commission requirement or a CHAI framework.
References
National Assembly of Viet Nam. Law No. 134/2025/QH15 on Artificial Intelligence. Adopted 10 December 2025; effective 1 March 2026.
Government of Viet Nam. Decree No. 142/2026/NĐ-CP Detailing a Number of Articles and Measures for Implementation of the Law on Artificial Intelligence. Issued 30 April 2026; effective 1 May 2026.
National Assembly of Viet Nam. Law No. 91/2025/QH15 on Personal Data Protection. Adopted 26 June 2025; effective 1 January 2026.
Government of Viet Nam. Decree No. 356/2025/NĐ-CP Detailing a Number of Articles and Measures for Implementation of the Law on Personal Data Protection. Issued 31 December 2025; effective 1 January 2026.
Ministry of Health of Viet Nam. Circular No. 19/2013/TT-BYT Providing Guidance on Management of Quality of Medical Examination and Treatment Services in Hospitals. Issued 12 July 2013; effective 15 September 2013; currently recorded as in force.
Ministry of Health of Viet Nam. Circular No. 43/2018/TT-BYT Providing Guidance on Prevention of Medical Incidents in Medical Examination and Treatment Establishments. Issued 26 December 2018; effective 1 March 2019; currently recorded in the national legal database as partially expired.
World Health Organization Regional Office for Europe. Bridging Theory and Practice – Implementation Insights on Artificial Intelligence in Health Care: Report of Working Group 1 of the Strategic Partners’ Initiative for Data and Digital Health. Published 26 June 2026.
Joint Commission. Responsible Use of AI in Healthcare Certification. Launched 1 June 2026.
Coalition for Health AI. AI Governance Playbooks. Released 27 May 2026.
Wenderott K, Krups J, Weigl M. Comprehensive recommendations for the implementation of artificial intelligence in healthcare: a narrative review on facilitators and barriers. BMJ Open Quality. 2026;15:e003639.
Artsi Y, Sorin V, Glicksberg BS, Korfiatis P, Nadkarni GN, Klang E. Large language models in real-world clinical workflows: a systematic review of applications and implementation. Frontiers in Digital Health. 2025;7:1659134.