Tuesday, 15 September 2026
Artificial intelligence is moving deeper into healthcare, but hospitals are discovering that a successful algorithm is only a small part of what is required to transform patient care. The more difficult challenge is redesigning clinical workflows, upgrading infrastructure, governing rapidly expanding volumes of data and demonstrating that technology produces measurable improvements for patients. That was one of the central messages from a healthcare leadership panel at AI4 2026, where executives and clinicians representing major medical institutions and rural healthcare discussed what happens when AI moves beyond experimentation and becomes part of everyday operations.
The discussion brought together Dr Caroline Chung, professor of radiation oncology and diagnostic imaging and co-director of the Institute for Data Science in Oncology at MD Anderson Cancer Center; Caleb Hoar, president and CEO of Boone County Health Center in Nebraska; Alpin Patel, an interventional radiologist and former healthcare innovation executive; and Morgan Jeffries, a neurohospitalist and medical director for AI at Geisinger. Despite representing very different parts of the healthcare system, the speakers repeatedly returned to the same principle: providers should begin with a clinical or operational problem rather than an ambition to deploy artificial intelligence.
That distinction is becoming increasingly important as hospitals face pressure to adopt AI quickly. New applications are appearing across clinical documentation, medical imaging, patient communication, revenue management and diagnostic support. Yet installing technology without redesigning the surrounding workflow can increase rather than reduce the workload. Patel described an example involving automated detection technology on a CT scanner. Instead of simplifying the radiologist’s work, the system initially produced an additional set of images that clinicians then had to review alongside the originals. The technology therefore created more work because the implementation had not been designed around how radiologists actually operated.
The lesson is relevant far beyond medical imaging. Hospitals are complex organisations where a change in one process can affect clinicians, IT departments, administrative employees, billing systems and patients. AI consequently needs to be treated as an organisational transformation rather than simply another software purchase.
One of the clearest examples of technology successfully crossing from experimentation into large-scale use is ambient clinical documentation. These systems can listen to conversations between clinicians and patients, with appropriate controls and consent, and produce draft clinical notes for review. Geisinger has deployed ambient documentation across its organisation, according to Jeffries. He said measurements showed approximately two minutes of documentation time being saved per note, although the precise figure varies. Multiplied across numerous consultations during a working day, relatively small individual savings can become substantial.
The response from clinicians has been equally important. Jeffries described unusually strong enthusiasm from physicians using the technology, with some reporting that they did not want to return to their previous documentation process. Ambient systems can also change the interaction between doctors and patients. Instead of dividing their attention between the patient and a computer screen, clinicians can potentially maintain more direct conversation while the technology assists with documentation in the background.
The experience at Boone County Health Center demonstrates that this is not only relevant to large healthcare networks. The Nebraska provider serves a small rural community and has also introduced ambient AI into its clinical operations. Hoar said the technology has helped create more comprehensive medical records while allowing clinicians to spend more of the consultation engaging directly with patients. It has also produced operational benefits around documentation, quality reporting and revenue-cycle processes.
But the implementation exposed another issue: the digital infrastructure supporting smaller hospitals may not have been designed for the volume of information created by modern AI applications. Boone County discovered that introducing ambient technology substantially increased the amount of data moving through its systems. Hoar said parts of the hospital’s IT infrastructure had not undergone major upgrades for many years, meaning additional capacity was needed to accommodate the new information flows.
The experience highlights an overlooked aspect of healthcare AI investment. Artificial intelligence is often discussed as a software transformation, but widespread adoption also creates demand for stronger networks, storage, cloud infrastructure, cybersecurity and data-management systems. For rural hospitals in particular, this can become a significant constraint. Smaller providers may have fewer technical employees and less capital available for infrastructure investment, even though they may have some of the strongest reasons to use AI because of persistent workforce shortages.
Boone County serves a town of approximately 1,800 people and a wider population of around 10,000. In such communities, recruiting specialist clinicians, nurses and technical employees can be difficult. AI therefore offers the possibility of supporting existing staff and extending scarce clinical capacity. This does not necessarily mean replacing healthcare workers. Hoar argued that rural healthcare currently faces such significant labour shortages that automation could initially help fill gaps rather than eliminate occupied positions.
Some administrative or routine responsibilities could move away from highly qualified employees, allowing nurses and clinicians to concentrate on areas where their skills are more urgently required. That could gradually change healthcare staffing models. Tasks traditionally performed by registered nurses in lower-acuity settings, for example, may increasingly be supported by technology or reassigned to other employees, while nurses concentrate on emergency departments, inpatient units, obstetrics and other areas requiring more complex clinical judgement.
The potential productivity benefit is considerable, but the panel cautioned against measuring success purely through time savings. Chung argued that healthcare organisations need to determine whether they are measuring outcomes that actually matter. A model can achieve impressive technical accuracy without producing any meaningful improvement in care.
The same applies to efficiency measurements. One frequently discussed benefit of ambient documentation is the reduction of after-hours administrative work by doctors. However, measuring whether physicians are logged into an electronic health-record system late in the evening does not necessarily reveal whether AI has improved their lives. Some doctors deliberately postpone documentation until after spending time with their families. Simply recording when they use the system could therefore produce a misleading interpretation of productivity or wellbeing.
Hospitals need to determine what outcome they are actually trying to achieve before deciding how to measure an AI deployment. If the objective is efficiency, the organisation needs to establish whether the overall workflow genuinely became faster. If the objective is diagnostic improvement, speed cannot come at the expense of accuracy or patient safety.
Context also matters. An AI system that is insufficiently accurate for a specialist academic medical centre could still potentially provide value in a rural community where access to specialist expertise is extremely limited, provided it is deployed with appropriate safeguards and used for the right purpose. The relevant benchmark is therefore not necessarily perfection. It is whether the technology improves upon the existing standard of care in the environment where it will actually operate.
This question becomes particularly important because humans already make mistakes. Requiring every healthcare AI system to approach perfect accuracy while comparing it against an imperfect existing process can create unrealistic expectations. However, the speakers cautioned that performance reported by a technology supplier or demonstrated during regulatory evaluation may not automatically translate into equivalent results inside another hospital. Patient populations, workflows, equipment and data can differ substantially between institutions.
Local validation consequently becomes essential. Hospitals need to understand how a model behaves using their own data and patient populations rather than relying exclusively on performance figures generated elsewhere. Continuous monitoring is equally important. AI systems can change in effectiveness as the data surrounding them changes. Patient populations evolve, clinical practices change and information entering the model can gradually differ from the material on which it was originally evaluated.
This creates the possibility of data drift and model drift, where performance deteriorates after deployment. Healthcare organisations therefore need mechanisms capable of monitoring AI systems throughout their operational lives rather than treating approval as a one-time event.
The scale of this challenge could become significant. Jeffries argued that healthcare organisations are moving towards a world in which AI functionality becomes embedded inside many existing technology products rather than arriving as clearly identifiable standalone systems. That means hospitals may eventually be responsible for monitoring hundreds of AI-enabled functions supplied by numerous vendors.
Keeping track of how those systems operate, what risks were identified, what mitigation measures were introduced and whether performance remains acceptable could require substantial additional resources. The challenge becomes even greater as healthcare AI moves from clinician-facing applications towards direct patient interaction.
A physician can review an AI-generated recommendation and apply years of medical training before acting on it. A patient interacting with an automated healthcare assistant cannot necessarily be expected to perform the same level of verification. This raises important questions for the next generation of patient-facing AI. A hospital cannot simultaneously tell patients that an automated system is trustworthy while expecting them to independently check every answer it provides.
Providers will therefore need stronger evidence about how patient-facing systems behave, when they fail and whether they reliably escalate situations requiring human intervention. The financial case for healthcare AI is another unresolved issue. Some applications produce relatively straightforward returns. Technology that improves coding, billing or revenue collection can be compared directly with its purchase and operating cost.
Clinical applications are harder to evaluate. Reducing administrative work could improve physician satisfaction and retention, but determining whether a particular doctor remained with an organisation specifically because of an AI tool is difficult. The economic benefit may nevertheless be substantial. Recruiting and replacing specialist physicians can be extremely expensive. If technology reduces burnout or increases the number of patients clinicians can treat, part of its return may appear indirectly through lower turnover and increased clinical capacity rather than a simple reduction in payroll.
This makes the definition of value highly dependent on the problem being addressed. An AI project designed to improve hospital operations can be evaluated differently from a system intended to detect disease earlier or improve the patient experience. Healthcare providers therefore need to determine the intended outcome before implementation rather than searching for financial justification afterwards.
Data ownership is becoming another major strategic issue. As hospitals connect AI systems to clinical records, enormous quantities of highly valuable healthcare information can pass through technology vendors and external platforms. Hoar said his organisation has begun examining vendor agreements much more closely to understand where data travels, who can access it and what external organisations may ultimately receive information derived from hospital operations.
For smaller healthcare providers, the issue represents a significant change in how technology contracts need to be negotiated. Clinical information is no longer simply something stored within a hospital’s electronic records. In an AI environment, data becomes an economic and strategic asset that can potentially support research, commercial products, insurance analysis and future algorithms. Hospitals therefore need clearer contractual controls over how their information can be used, shared or incorporated into external systems.
The growing use of unofficial AI tools adds another complication. Healthcare employees are already experimenting with general-purpose AI assistants for work-related tasks, sometimes outside formal enterprise systems. Attempting to prohibit every such application may push activity further out of sight. Jeffries suggested that organisations need to understand which tools employees are finding useful and, where appropriate, provide secure enterprise alternatives that allow experimentation without exposing protected or commercially sensitive information.
This points towards a broader cultural challenge. AI adoption cannot be managed solely through technology departments and compliance policies. Clinicians, nurses and administrative employees need enough understanding of the systems to recognise both their capabilities and their limitations. The panel repeatedly emphasised the importance of involving frontline employees in implementation. AI introduced to a clinical team without understanding how that team works can easily create resistance or additional workload. Technology developed with clinicians, rather than simply imposed upon them, has a much greater chance of becoming useful.
Education therefore becomes part of the infrastructure required for AI adoption. Employees need to understand what the technology is doing, how its conclusions are generated, what information it can access and where human judgement remains necessary. For healthcare executives, this means the transformation is considerably larger than purchasing licences for AI software. Successful deployment can require changes to networks, data architecture, contracts, cybersecurity, clinical workflows, staffing models, governance and performance measurement.
The physical healthcare environment may also be affected. Hospitals increasingly depend on high-capacity digital infrastructure, resilient connectivity and secure data environments. As more clinical processes become AI-enabled, those requirements could influence future investment in hospitals, outpatient facilities and rural healthcare infrastructure.
Smaller hospitals may face the greatest challenge. They have strong incentives to automate because labour shortages are particularly acute, but they often operate with older buildings, older IT systems and smaller capital budgets. AI could therefore expose a growing digital infrastructure divide between large healthcare networks and smaller regional providers. At the same time, the experience of Boone County suggests that smaller institutions can adopt meaningful AI applications when technology is targeted at specific problems rather than treated as a comprehensive transformation programme.
The panel’s final advice to healthcare executives reflected that pragmatism. Organisations should define the problem before selecting the technology, understand where their data is going, involve the people who will actually use the system, determine how success will be measured and avoid deploying AI everywhere simply because it has become available.
Healthcare may ultimately become one of the industries most profoundly changed by artificial intelligence, but the transformation is unlikely to happen through algorithms alone. Every successful AI deployment sits inside a much larger system of people, technology, buildings, data and clinical processes.
For healthcare providers, the competitive advantage may therefore come not from possessing the largest number of AI applications, but from knowing which problems are worth solving and having the infrastructure, governance and organisational discipline required to turn those technologies into better care.
Source: CIJ.World Research & Analysis Team
