Agentic AI is quickly becoming one of the most discussed developments in enterprise technology. Unlike conventional generative AI systems that mainly respond to prompts, agentic AI can potentially make decisions, carry out multi-step tasks and act with a much higher level of autonomy.
That makes the technology exciting, but it also raises a much more serious question: How much decision-making should organisations safely hand over to AI?
For healthcare leaders, that question is especially important. A poorly written AI-generated summary may be inconvenient, but an autonomous system making the wrong decision about medication, patient prioritisation or clinical workflow could have much more serious consequences.
Interestingly, healthcare may be able to learn something from another industry where reliability and safety are also critical: aviation.
At HIMSS26 APAC, Joe Chiu, a consultant at Singapore's Changi Airport Group, shared how the airport organisation is approaching generative and agentic AI. While airports and hospitals obviously operate in very different environments, the underlying governance challenge is remarkably similar.
Both deal with complex operations, huge amounts of data, real-time decision-making and situations where technology cannot simply be allowed to fail silently.
Agentic AI Raises the Stakes
Generative AI already requires careful governance, but agentic AI introduces another layer of risk because the system may not merely provide an answer.
It may take action.
An agent could potentially retrieve information, analyse a situation, make a decision, trigger another system and continue working through several steps without requiring constant human intervention.
That means organisations need a higher level of confidence before allowing these systems into operational environments.
Chiu acknowledged that Changi Airport Group has experimented with AI initiatives that ultimately never reached production because the organisation could not obtain sufficient assurance that the systems were safe or reliable enough.
That is an important lesson for healthcare.
Not every AI proof-of-concept needs to become a deployed system.
Sometimes the right outcome of an experiment is simply:
This technology is not ready for this particular use case yet.
In sectors where mistakes can affect safety, that is not a failed project. It is responsible governance.
Build AI Agents Around Specific Problems
One of the most important parts of Changi Airport Group's approach is its focus on purpose-built agents.
Instead of assuming that a powerful general-purpose AI model can be connected to an organisation and immediately perform reliably, CAG is developing and training agents around clearly defined use cases.
That distinction matters.
General-purpose AI models are designed to handle a wide variety of questions and tasks. Enterprise environments, however, usually require much more controlled behaviour.
A hospital may not want an AI system that can do hundreds of things reasonably well.
It may instead want an agent that does one specific task extremely reliably, within carefully defined boundaries.
For example:
These narrower applications can be easier to test, monitor and govern.
The more autonomy an AI system receives, the more important it becomes to define precisely what the agent is allowed to do.
Agentic AI Is Still a Work in Progress
Chiu was also clear that CAG's agentic AI journey is not finished.
That is another useful reminder for healthcare organisations.
There is enormous enthusiasm around autonomous AI agents, but the technology is still evolving rapidly. Deployment models, governance standards, security controls and monitoring practices are all developing alongside the underlying AI models themselves.
Healthcare organisations therefore need to avoid treating agentic AI as a mature, ready-made replacement for existing processes.
A better approach is controlled experimentation.
Start small.
Measure carefully.
Expand only when the system consistently performs within acceptable boundaries.
CAG Is Building an AI Layer Into Its Existing Architecture
Rather than creating isolated AI applications, Changi Airport Group is adding AI capabilities to a technology architecture that it has already been developing for several years.
That includes components such as:
This architectural approach is particularly relevant to healthcare.
Hospitals already operate dozens, sometimes hundreds, of interconnected systems including:
If every AI project creates its own separate integration architecture, the environment can quickly become difficult to maintain.
A shared AI services layer could potentially allow governance, monitoring, authentication and integration controls to be implemented once and reused across multiple applications.
That could make future AI deployments both faster and safer.
Generative AI and Agentic AI Should Not Be Treated as Completely Separate Journeys
Another interesting element of CAG's strategy is that the organisation is preparing for generative AI and agentic AI at the same time.
It is easy to imagine these technologies as sequential stages:
First implement generative AI.
Then, several years later, move into agentic AI.
In reality, the line is becoming increasingly blurred.
A generative AI assistant may begin by drafting content, then gain the ability to retrieve data, interact with systems and execute actions.
At that point, the system is moving toward agentic behaviour.
Healthcare organisations implementing AI today should therefore consider not only what a system can do now, but also what governance capabilities will be needed when that system becomes more autonomous.
Autonomous Systems Require a Different Operational Model
As AI becomes more agentic, organisations will need to think differently about system operations.
CAG is considering capabilities such as:
Those capabilities sound attractive from an efficiency perspective.
But every reduction in human involvement also increases the importance of controls.
For healthcare, questions would include:
Those questions need answers before an AI agent is given meaningful authority inside a clinical or operational workflow.
Start With the User, Not the Technology
Perhaps the most transferable lesson from Changi Airport Group is surprisingly simple.
Do not begin with AI.
Begin with the user.
Chiu explained that CAG works backwards from the customer journey rather than starting with whichever technology happens to be fashionable.
The organisation uses techniques such as:
Only after understanding the problem does the organisation select the technology.
Chiu summarised the principle as putting the customer ahead of the product.
Healthcare organisations could apply exactly the same thinking.
Instead of asking:
"Where can we use AI?"
Ask:
"What problem are our patients, clinicians or staff struggling with?"
The difference is significant.
The first question encourages technology-driven projects.
The second encourages outcome-driven projects.
Healthcare AI Should Work Backwards From the Patient and Clinician Journey
Consider outpatient appointments.
A technology-first approach might say:
"We want to deploy an AI chatbot."
A user-centred approach would first examine the actual experience.
Perhaps patients are frustrated because they cannot easily:
Once those problems are understood, the organisation can decide whether AI is actually the best solution.
Sometimes it will be.
Sometimes a better website, clearer instructions or improved system integration may solve the problem more effectively.
That is why problem definition needs to come before technology selection.
More Than Half of Experiments May Never Reach Production
CAG's willingness to abandon experiments is another useful cultural lesson.
Chiu estimated that more than 60% of projects do not survive the organisation's experimentation process.
At first glance, that sounds inefficient.
In reality, it may indicate exactly the opposite.
Testing a concept cheaply and abandoning it early is much better than spending months implementing technology that users do not actually need.
Healthcare organisations sometimes struggle with this because technology projects can quickly acquire political or organisational momentum.
Once a project receives a budget, vendor contract and executive sponsor, there can be pressure to deliver something even if the original assumptions turn out to be wrong.
AI experimentation should allow organisations to say:
"We tested it. It did not provide enough value. We stopped."
That needs to be considered a valid outcome.
Build Reusable Technology Instead of Starting From Zero Every Time
CAG describes its modular technology components as something similar to Lego blocks.
The idea is straightforward.
Instead of building every new digital solution from the ground up, the organisation creates reusable components that can be assembled differently depending on the business need.
Healthcare could benefit enormously from the same approach.
An AI platform might contain reusable components for:
A new healthcare AI application would then reuse those existing controls rather than recreating them.
This could significantly reduce implementation time while also creating more consistent governance across the organisation.
Internal Expertise Matters
Changi Airport Group has also invested in its own data engineers, data scientists and software developers instead of depending entirely on external vendors.
That strategy provides two important benefits.
First, the organisation retains more of the intellectual property and technical knowledge created during implementation.
Second, internal teams gradually develop an understanding of how AI actually behaves inside their own operational environment.
That second advantage is particularly important.
AI systems are not something organisations can simply purchase and forget.
Models change.
Data changes.
User behaviour changes.
Security threats change.
Regulations change.
An organisation that outsources every aspect of AI may eventually find itself unable to understand or govern its own systems without external assistance.
Healthcare organisations do not necessarily need to build every AI model internally, but they should maintain enough internal capability to understand:
Do Not Give Away All of the Knowledge to Vendors
Vendor partnerships will remain important, especially because healthcare organisations rarely have the resources to build every platform themselves.
But CAG's experience highlights the importance of internal participation.
When internal engineers and data teams work directly alongside vendors, the organisation retains institutional knowledge.
That makes future improvements easier.
It also reduces the risk of becoming completely dependent on one technology supplier.
For hospitals exploring AI, this may become a major strategic consideration.
The goal should not necessarily be to own every piece of technology.
The goal should be to understand and control the critical parts of the ecosystem.
Good AI Still Depends on Good Data
Another major part of CAG's digital strategy is its Customer 360 platform, which brings together information from multiple touchpoints to provide a more complete view of the customer.
This enables more personalised services.
But Chiu described building that capability as a difficult process.
That should sound familiar to almost anyone working in healthcare IT.
Healthcare organisations often have information spread across multiple systems:
AI becomes far more useful when it can access relevant information from across those systems.
But connecting everything is complicated.
Data formats differ.
Patient identities need to be matched correctly.
Consent needs to be respected.
Access permissions need to be enforced.
Older systems may have limited integration capabilities.
This means the hardest part of an AI programme may not be the AI model at all.
It may be the data architecture underneath it.
Consent Becomes Even More Important as AI Uses More Data
CAG's experience also highlights the importance of customer consent.
The same principle becomes even more critical in healthcare, where the underlying information is often highly sensitive.
An AI system capable of seeing more information can potentially provide better assistance.
But organisations need clear rules governing:
Greater AI capability cannot come at the expense of privacy.
Healthcare Can Learn From Safety-Critical Industries
Healthcare often looks within its own industry when developing technology governance.
But there is considerable value in studying other safety-sensitive sectors.
Aviation, financial services, energy and industrial manufacturing have all spent decades developing systems where automation must operate inside carefully controlled environments.
The exact governance models may differ, but some principles are widely transferable:
Those principles become particularly important with agentic AI.
Agentic AI Needs a Different Definition of Success
One of the biggest misconceptions about autonomous AI may be that success means eliminating human involvement.
That is probably the wrong target, especially in healthcare.
A more useful definition of success might be:
How much safe, repetitive work can the AI handle while ensuring that consequential decisions still receive appropriate human oversight?
An agent that completes 80% of a workflow and reliably escalates the remaining 20% may be far more valuable than one attempting to automate 100% but occasionally making dangerous decisions.
The best agentic systems may therefore be designed not only to know what they can do, but also to recognise when they should stop and ask a human.
Final Thoughts
Changi Airport Group's experience offers healthcare organisations an important reminder: successful AI transformation is not primarily about chasing the newest model.
It is about building the organisational capability to use AI safely.
That means creating modular technology foundations, developing internal expertise, establishing guardrails, improving data integration and being willing to abandon projects that fail to provide sufficient value or assurance.
Most importantly, it means starting with people rather than technology.
Hospitals should work backwards from the needs of patients, clinicians and operational teams, identify genuine problems and only then determine whether AI is the appropriate solution.
Agentic AI will almost certainly become more capable over the coming years. Systems will perform increasingly complex tasks, interact with multiple applications and operate with greater independence.
But greater autonomy should also bring greater governance.
For healthcare, the goal should not simply be to create AI that can act on its own.
The goal should be to create AI that knows what it is allowed to do, operates within clearly defined boundaries, explains what it has done and reliably hands control back to humans when the situation requires it.
That may ultimately be the difference between impressive AI experimentation and AI that healthcare organisations can genuinely trust.
Comments 0