There’s an AI for that. And that’s exactly the problem

Almost every operational challenge seems to invite the same response: there’s artificial intelligence (AI) for that. 

Healthcare is no different – whether facing challenges with patient engagement or waiting times, teams often think of AI as the fix-all. 

But according to A/Prof Adam Chee, Director of the AI Development Office at Singapore General Hospital (SGH), this enthusiasm risks putting the solution before the problem. 

Speaking at HealthTechX Asia 2026, Prof Chee urged healthcare organisations to resist the temptation to adopt AI before understanding the underlying challenges they are trying to solve.  

“If you do not fix the root cause, the technology that comes next, be it AI or not, is not going to solve the root cause of your problem. In fact, it is going to accelerate it.” 

When successful AI pilots fail in practice 

One of the most common pitfalls is assuming that a promising AI pilot will translate into real-world success. 

In a controlled environment, teams can define a specific use case, establish success criteria and test a solution against predictable scenarios.  

But patients and workflows in real life do not always follow standard pathways, and unusual cases can expose weaknesses that were not apparent during testing. 

When a solution fails after deployment, the AI model often takes the blame.  

Yet Prof Chee argued that the failure may have begun much earlier, when the problem was poorly defined or the solution was not designed for the realities of clinical practice. 

Edge cases – rare or unusual conditions that fall outside the normal operating parameters the model was trained on – are particularly challenging.  

As more processes become automated, staff may have fewer opportunities to intervene. 

Without a clear way to handle exceptions, automation can make recovery more difficult rather than easier. 

Model performance alone is also an inadequate measure of clinical value.  

An AI tool may demonstrate strong predictive accuracy but still fail to support clinical decisions if it is introduced at the wrong point in the care pathway. 

Similarly, launching an application or recording high usage figures does not necessarily mean it has changed clinical behaviour or improved patient outcomes. 

Differentiating between AI model, product, and solution 

Prof Chee drew a clear distinction between an AI model, a product and a complete solution. 

A well-performing product may not necessarily make a good fit for a hospital. 

The right solution must be integrated into existing clinical information systems and workflows, adapted to the organisation’s needs, and supported after deployment.  

That includes planning for technical failures, maintaining the system, managing licences and allocating people to oversee its operation. 

These requirements carry ongoing costs that can be overlooked when organisations focus primarily on the initial development or purchase of an AI tool. 

Building a bespoke solution is not always the best option either, said Prof Chee.  

In some cases, an existing product may be more economical, particularly when its development and maintenance costs can be spread across multiple customers. 

Ask the right questions first 

To help organisations avoid solution-first thinking, Prof Chee recommended returning to a few fundamental questions whenever an AI proposal arises:  

"Who is the solution for? What unmet need is it addressing? What decision, behaviour or workflow should change? And how will success be defined?” 

Vague answers should prompt further investigation.  

For instance, a proposal to improve turnaround time needs to establish what turnaround time means in that context, what currently prevents the team from meeting its target, and how the proposed intervention would remove that barrier. 

He also cautioned against copying solutions that have worked elsewhere.  

Different institutions have different patient populations, workflows and resources, meaning an approach that works in one setting may not transfer directly to another. 

Optimising limited AI development resources 

Prof Chee’s message was that organisations should be more deliberate about where to use AI and why they use it. 

For a lean team, taking on a new project means another may have to wait.  

In balancing competing priorities, leaders need to be prepared to direct limited resources towards initiatives with clearly defined needs and credible benefits first. 

“What I tell people is that they don’t need AI – they need to fix their workflow first,” he noted. 

AI can be powerful, but its value depends on whether it addresses a genuine need, fits into clinical practice and can be sustained over time.  

For healthcare executives, the goal is not to find an AI for every problem, but to solve the right problem in the first place. 

The article was originally published in Hospital Management Asia here, and edited. 

---------------------------------------------------

Cindy Peh is the Content and Community Manager at Hospital Management Asia, which showcases trends and best practices in healthcare management via in-person and digital events, as well as an online publication.