Saturday, 12 September 2026
Artificial intelligence in healthcare is often presented through its most dramatic possibilities, from disease detection to advanced clinical decision support. Novant Health is taking a broader approach. Rather than starting with the technology itself, the US healthcare group is looking at where patients, clinicians and hospital operations encounter unnecessary delays, administrative burdens or complexity, then asking whether AI can remove some of that friction. Speaking at AI4 2026, Vijay Sankararaman, Chief AI Officer at Novant Health, described a strategy built around three areas: patient experience, clinical care and operational performance. The objective is not to replace doctors, nurses or other employees, but to reduce repetitive work, improve access to information and allow healthcare professionals to spend more time with patients.
Healthcare contains enormous amounts of administrative activity surrounding the actual delivery of medicine. Patients may struggle to determine which clinician is available, schedule appointments, understand bills, prepare for surgery or obtain reliable answers outside normal office hours. Clinicians, meanwhile, spend substantial amounts of time reviewing records, preparing notes, responding to messages and completing documentation. Many of these processes have accumulated additional systems and workarounds over the years, creating layers of complexity that healthcare organisations and patients have effectively learned to tolerate. Novant’s approach is to identify these points of friction rather than introducing AI simply because the technology is available.
One of the clearest examples involves patients preparing for surgery. They frequently have practical questions long before entering hospital, ranging from medication and diet to preparation at home and post-operative recovery. Traditionally, many of these questions reach nurses or clinical teams through telephone calls and messages. Clinical involvement remains essential when medical judgement is required, but routine questions can arrive outside normal working hours and add to already substantial workloads.
Novant launched an AI-enabled virtual assistant called Aubrey in June 2026 as part of a phased perioperative care programme. The system can provide reminders and guidance to eligible patients, including pre-operative instructions, information about what to expect before and after procedures, scheduling assistance and practical details surrounding a hospital visit. The technology is intended to complement rather than replace clinical teams, with human oversight and existing routes for patients to contact healthcare professionals remaining in place. This provides a possible model for healthcare AI in which technology creates an additional communication layer around existing care, allowing routine information to be available around the clock while more complex questions continue to be handled by clinicians.
Novant is extending the same principle into primary care. In July 2026, the organisation announced an expansion of virtual primary-care services across North and South Carolina, including AI-supported health screening before appointments. Patients can provide information about symptoms and concerns before speaking with a clinician, giving the medical team more structured information at the beginning of the consultation. The broader opportunity is to make healthcare access less fragmented. Patients often encounter separate systems for appointments, medical questions, billing, medication and follow-up care. AI potentially creates a more continuous interface across these interactions, provided the underlying systems communicate securely and the information remains medically reliable.
The same logic is being applied to clinicians. A physician preparing for an appointment may need to review laboratory results, medication histories, previous notes and recent medical changes distributed throughout an electronic medical record. AI can increasingly assemble and summarise this information before the consultation. Novant is also using ambient clinical documentation, where conversations can be captured with patient permission to assist with preparing medical notes, together with tools supporting clinician inboxes and drafting responses to routine messages. The intention is to reduce administrative work outside normal clinical hours without transferring responsibility for the medical record to an algorithm. Physicians remain responsible for confirming that information is accurate.
Similar opportunities exist for nurses, pharmacists and other members of clinical teams. Hospital care depends on information moving continuously between different professions, making handovers and coordination potential areas for AI assistance. One example presented at AI4 involved patient telemetry. Continuous monitoring provides important information about a patient’s condition but can also generate large numbers of alerts, contributing to alarm fatigue among staff while sometimes leaving patients connected to equipment longer than necessary. Novant has used predictive modelling to help clinicians determine when particular monitoring services may no longer be required, potentially reducing unnecessary monitoring while releasing equipment for other patients.
The example demonstrates that useful healthcare AI does not always need to be highly visible. An algorithm that improves equipment allocation, reduces unnecessary alarms or helps clinicians organise information may be almost invisible to patients while still improving hospital capacity and workflow. This is particularly important because healthcare facilities contain expensive infrastructure that cannot easily be expanded whenever demand increases.
Novant is also applying predictive systems to population health. Sankararaman described work using consented electronic medical-record information to identify patients who may have elevated risks of serious conditions and could benefit from further screening. One example involved lung cancer. According to the presentation, Novant identified roughly 20,000 people within its patient population who potentially warranted additional outreach based on risk indicators. The figure represents a programme result reported by Novant rather than an independently validated diagnosis, and identification by such a model does not mean that an individual has cancer.
The value of the system is in prioritisation and speed. A health network serving a large population cannot manually contact every patient with the same intensity. Predictive models can help identify groups where screening may be particularly important, after which conversational AI can support outreach through voice or text and help patients arrange appointments. Novant said it is examining similar approaches for other conditions where earlier detection can materially affect treatment, including colorectal and breast cancer.
AI-supported screening can also expand access to services that traditionally require specialists. Sankararaman highlighted diabetic retinopathy, an eye complication associated with diabetes that can lead to vision loss. AI-assisted retinal imaging can enable screening in primary-care settings, with patients referred to specialists when results indicate further assessment is necessary. Such models could be particularly relevant in rural communities where access to specialist healthcare is more limited. These applications demonstrate that healthcare AI extends well beyond generative systems and can include predictive modelling, computer vision, conversational interfaces and workflow automation.
The third part of Novant’s strategy concerns hospital operations. Healthcare facilities are complex environments where expensive assets, specialised employees and limited capacity must be coordinated continuously. Even relatively small inefficiencies can therefore have significant financial and clinical consequences. Procedure cancellations provide one example. A colonoscopy involves more than reserving a room: clinical staff, equipment, supplies and often anaesthesia services must be coordinated around a particular time. When a patient cancels shortly before the procedure, that capacity can be difficult to refill.
Sankararaman described work using AI to understand the factors contributing to cancellations and improve the allocation of clinical resources. The objective is not simply to predict that someone may fail to attend but to identify where additional communication or preparation could prevent a cancellation and where capacity can be reassigned when a slot becomes available. The same operating model can potentially extend across hospitals, and Novant is examining applications in contact centres, supply chains, revenue management and wider hospital operations where better forecasting could help match resources with demand.
This represents an important shift in the healthcare AI discussion because the financial value of the technology may come as much from improving the utilisation of existing hospitals as from creating entirely new medical capabilities. Operating theatres, diagnostic equipment, hospital beds and clinical staff are constrained resources. If AI allows a healthcare provider to use those resources more efficiently, additional effective capacity can potentially be created without equivalent physical expansion.
That has implications for healthcare real estate and capital investment. Hospitals and medical campuses are among the most capital-intensive property types, with new capacity requiring substantial construction expenditure, lengthy development periods and complex approvals. Technology that improves the productive use of existing facilities could influence decisions about where and when additional physical capacity is required. It is unlikely to eliminate the need for new healthcare property, however. The more probable outcome is that digital intelligence and physical infrastructure develop together, with AI helping healthcare organisations operate both existing and future facilities more efficiently.
Patient communication could become another important source of efficiency. Someone preparing for surgery may need to change a laboratory appointment, ask a billing question and understand rehabilitation options. Traditionally, these issues might require several calls to different departments. Novant’s longer-term vision involves conversational systems capable of dealing with several parts of the patient journey through a single interaction. An assistant could potentially reschedule an appointment, explain administrative information and direct a patient towards appropriate follow-up services without requiring separate contacts with multiple departments.
The effectiveness of this approach depends heavily on integration. A conversational interface is only useful if it can securely interact with the scheduling, billing and clinical systems required to complete the requested action. Underlying data architecture and interoperability may therefore prove as important as the AI model itself.
Healthcare also places unusually high demands on trust. An inaccurate retail recommendation may be inconvenient, but an error involving medication, clinical guidance or a medical record can have much greater consequences. Novant says its AI programme therefore operates within a trust and safety framework covering internally developed systems, external models and technology partners, with clinical safety, monitoring, fairness, transparency and human oversight forming part of the approach.
These controls become increasingly important as AI moves from providing information towards performing actions. A system that summarises a medical record presents one level of risk, while a system capable of changing appointments, communicating clinical instructions or influencing treatment decisions requires stronger controls over what it can access and what it is authorised to do. The challenge for healthcare organisations will be to increase automation without creating invisible decision-making. Doctors and nurses need to understand where AI is involved, patients need confidence that their information remains protected, and healthcare organisations need mechanisms for monitoring systems after deployment.
The broader lesson from Novant’s programme is that successful healthcare AI may depend less on spectacular demonstrations than on solving persistent operational problems. Instead of asking where the newest model can be deployed, the organisation says it starts by identifying where patients struggle, where clinicians lose time and where operational processes create unnecessary delay or cost. Technology is then applied to those specific problems.
That approach could become increasingly important as healthcare systems face rising costs, workforce pressures and growing demand. Hospitals cannot simply add clinicians, beds and administrative employees indefinitely, making the productivity of existing resources increasingly important. Artificial intelligence offers one route to improving that productivity, but its value will ultimately depend on whether it simplifies healthcare rather than adding another technological layer to an already complicated system.
For Novant Health, the emerging model is one in which machines organise information, manage routine communication, identify potential risks and automate repetitive processes while people remain responsible for care. If that balance can be maintained, some of the most transformative healthcare applications of AI may not be those attempting to replace clinicians, but those that give clinicians more time to practise medicine and make healthcare considerably easier for patients to navigate.
Source: CIJ.World Research & Analysis Team
