Moving healthcare AI from a successful pilot into routine hospital operations requires reliable data, integration with existing systems, clear accountability and staff prepared to assess its outputs.

In this interview, Whee Teck Ong, Head of Kyndryl Consult, ASEAN & Korea, discusses how hospitals can prepare for AI deployment, preserve clinical judgement and determine which applications are ready to scale.

What most often prevents a successful healthcare AI pilot from becoming part of everyday hospital operations, and how should institutions address those barriers?

One of the biggest obstacles is the gap between how an AI tool performs during a pilot and what it takes to operate at scale across a hospital. A pilot may work well within one department, using a limited dataset, but wider deployment introduces dependencies on existing applications, patient records and clinical workflows.

Kyndryl’s healthcare readiness findings published in March 2026 illustrate the scale of this challenge: 76% of healthcare organisations surveyed reported having more AI pilots than they could scale, while 31% identified regulatory and compliance concerns as a key barrier.

Hospitals should assess these dependencies before expanding a pilot, identifying which systems the AI must connect with, whether the underlying data is reliable and how the application performs across departments. Testing should reflect everyday conditions, including inconsistent data, unexpected inputs and system disruptions.

The decision to scale should depend on whether the tool can operate reliably within existing hospital workflows, rather than how well it performed in a controlled pilot.

As Singapore develops locally contextualised healthcare AI through initiatives such as SIMFONI, what should hospitals prepare before integrating these tools into existing electronic medical records and clinical workflows?

Singapore’s SIMFONI initiative aims to develop AI models that reflect Singapore’s population, clinical guidelines and healthcare practices. As these models progress towards deployment, hospitals will need to determine how AI will access and use information across existing clinical systems.

Hospitals should first map where relevant patient information resides, how it moves between applications and whether existing systems can provide the data AI needs to generate reliable insights. This includes assessing data quality, interoperability and access permissions before connecting AI to electronic medical records.

The next consideration is how AI fits into the clinician’s workflow. Recommendations should surface where clinicians already work and at the point they are needed. If staff need to switch between applications, manually retrieve information or repeatedly verify outputs, AI risks adding work rather than reducing it.

Before deployment, hospitals should also define which tasks AI will support and where independent clinical review is required. The test is whether AI can fit into existing care processes without compromising clinical judgement or creating new operational friction.

How should hospitals divide responsibility for monitoring AI performance, managing patient data access and responding when an AI tool produces unreliable outputs?

Hospitals should agree on responsibilities before using AI in clinical care. The hospital remains accountable for how the system is used, including who can access patient data, how performance is monitored and how unreliable outputs are handled. It should also document who is responsible for each decision and how incidents are reviewed.

Within the hospital, clinical leaders should decide where AI can inform care and when a clinician must make an independent assessment. Technology and data teams should manage system performance, data quality and access controls. Risk and compliance teams should oversee regulatory obligations and the hospital’s governance arrangements.

Hospitals can use policy-as-code to implement specific technical rules, such as which users or systems can access patient data. Those rules can support consistent controls, but they do not replace accountable owners or clinical judgement.

Technology providers should address technical issues and explain their systems’ performance and limitations. Clinicians remain responsible for how AI-generated recommendations are applied to individual patients.

Hospitals should decide what triggers an incident review, who can pause an AI system and how care will continue while it is unavailable.

What does meaningful workforce readiness look like in healthcare, beyond training staff to use a new tool, and how should hospitals preserve human judgement within AI-supported workflows?

Kyndryl’s 2026 People Readiness Report found that only 16% of healthcare respondents say their workforce is ready to successfully leverage AI. The findings point to a broader challenge: organisations need to prepare employees for changes in how work is performed, rather than focus solely on training them to operate new tools.

Take clinical documentation as an example. AI may help generate notes or summarise patient information, but hospitals must determine who reviews the output, how inaccuracies are corrected and whether the process genuinely saves clinicians time.

Training should include realistic scenarios involving incomplete information, incorrect recommendations and situations requiring escalation. Healthcare professionals must be able to recognise unreliable outputs, understand when further assessment is required and retain the authority to challenge AI recommendations before the technology becomes part of routine workflows.

How should healthcare leaders decide which AI applications are ready to scale, and what evidence would justify expanding, revising or stopping a deployment?

Healthcare leaders should assess AI applications against their business and care goals. They should also consider whether the application improves the intended process, performs reliably under operational conditions and keeps risks manageable as its use expands.

The evidence should reflect the application. For administrative AI, hospitals could track processing time, manual corrections, staff workload and operating costs. For AI supporting clinical decisions, they would also need evidence of accuracy, reliability and clinical safety. Clinicians should be able to review the recommendations and understand the basis for them, with safeguards in place to protect patient data.

Results should be compared with the existing process and tested across representative settings. An AI tool that reduces documentation time but requires substantial additional verification, for example, may deliver little overall benefit. Performance should also be monitored after deployment as the data, systems and workflows change.

Leaders should scale an application when it meets its operational and safety targets and supports the organisation’s goals. They should revise it when it shows value but integration or workflow problems remain, and stop it if unreliable performance, unacceptable risk or limited value persists.