As hospitals adopt more AI, there is a risk that technology intended to reduce workload simply adds another system to an already fragmented environment. We need to pay more attention to what happens around the AI, writes Niru Rajakumar.

Hospitals are being offered an extraordinary number of AI tools. Clinical documentation, pre admission screening, decision support, coding and administrative processes are all attracting new applications. 

Many of these tools can demonstrate a useful capability. The problem is that hospitals do not operate as a collection of individual tasks. 

A clinician’s work crosses departments, systems and stages of the patient journey. Information created in one place needs to be available somewhere else. Decisions made by one team affect what another team does next. Introducing a tool that makes one task faster does not necessarily make the hospital workflow more efficient. 

As hospitals move from experimenting with AI to deploying it more widely, I think we need to become much more interested in what happens on either side of the model. 

Saving time on a task is not the same as improving the workflow 

Ambient clinical documentation is a useful example. 

The capability is compelling. AI can listen to a clinical interaction and prepare documentation for the clinician to review, potentially reducing the time spent producing notes. 

But the value to the hospital depends on what happens next. 

If the clinician needs to move into another application, review the output and then transfer information into the patient record, we have automated part of the documentation task while leaving much of the surrounding workflow untouched. 

This is why productivity claims around individual AI tools need to be considered in context. A few minutes saved at one point in the process can be valuable, but we also need to understand whether implementation creates new steps, handovers or work elsewhere. 

The same principle applies well beyond documentation. 

Hospitals need to evaluate the journey around the AI 

Consider pre-admission screening. 

AI can assess available information and identify patients who may need additional attention before admission. From a technical perspective, producing the risk assessment may be the sophisticated part of the process. 

Operationally, it is only the beginning.

The assessment needs to draw on the right patient information. It needs to reach the appropriate clinician early enough to be useful. That person needs to understand why the patient has been flagged and be able to take the next step without searching across other systems. 

This is where evaluating an AI application in isolation becomes problematic. Hospitals need to look at the journey around it: what information the AI requires, where that information already lives, where the output goes and how it affects the people and processes that follow. 

The unit of evaluation should increasingly be the workflow, rather than the AI task. 

Another application can mean another integration problem 

Hospitals have spent years trying to connect fragmented digital environments. There are clinical systems, departmental systems, diagnostic systems and administrative platforms, often accumulated over many years and connected to varying degrees. 

AI could help simplify that environment. It could also make it more complicated. 

If every new capability arrives with its own interface, identity requirements, data connections and workflow, hospitals risk building another layer of fragmentation on top of the one they already have. 

This does not mean every AI capability has to come from the hospital’s core system. Hospitals will continue to use specialist technologies, and innovation will come from many different places. 

What matters is how those capabilities become part of the hospital environment. 

Can they use information that already exists without unnecessary duplication? Can their output appear where someone is already working? Can the information become part of the patient record when appropriate? Can identity, permissions and audit requirements be handled without creating parallel processes? 

These may appear to be integration questions, but they have a direct effect on whether the promised benefits of AI are realised in practice. 

Governance is part of the workflow too 

As AI moves closer to clinical work, integration cannot be separated from governance. 

Hospitals need to know what patient information an AI capability can access, who has authorised that access and whether permissions reflect the person’s role and context. They need an audit trail and clarity about where information is processed and stored. 

There also needs to be a clear point at which a person reviews, accepts or acts on AI-generated information. 

That human involvement should not be treated simply as a safeguard added after an AI capability has been designed. It affects where the technology belongs in the workflow and what information needs to be presented to the person making the decision. 

Good implementation therefore requires more than connecting an API or deploying a model. It requires understanding the clinical and operational context into which that model is being introduced.

Start with the hospital workflow 

None of this argues for moving more slowly on AI. There are areas of hospital work where the technology has obvious potential to reduce administrative burden and make useful information available earlier. 

But the starting point matters. 

Rather than asking where a new AI tool could be deployed, hospitals can start by examining workflows where staff are already spending time finding information, entering the same data more than once, moving between systems or waiting for information before they can act. 

Then AI becomes one possible way of addressing a known problem rather than another technology that has to find a place in the hospital after it has been purchased. 

As AI capabilities become easier to acquire, this distinction will become more important. Hospitals are unlikely to suffer from a shortage of AI applications. 

The challenge will be making sure each one leaves the hospital with less fragmentation than it found.

Viewpoint articles are the author’s opinion, produced without payment or sponsorship.