As artificial intelligence becomes embedded in everyday hospital operations, health systems are confronting a new question about what happens when the AI disappears or quietly stops working as intended.
For CIOs and other technology leaders, the issue is becoming more pressing as clinicians and operational teams move beyond experimenting with AI and begin building workflows around it. Once a tool becomes part of how work gets done, an outage can affect far more than convenience.
“Once the loss of an AI capability can impact patient care, operational throughput, revenue, or safety, it should be treated like any other mission-critical system: documented, tested, and recoverable,” Cleveland Clinic CIO Sarah Hatchett told Becker’s.
Cleveland Clinic reviews and rehearses business continuity plans annually as new technologies are introduced, Ms. Hatchett said. The health system also tiers AI and other software based on critical functions, allowable outage times and recovery priorities.
At Los Angeles-based Cedars-Sinai, the threshold for requiring a formal continuity plan is largely based on dependency, Chief Health Informatics Officer Shaun Miller, MD, told Becker’s.
The health system evaluates factors including patient safety, clinical dependency, scale of adoption, time sensitivity, the availability of a safe alternative and the consequences of prolonged unavailability.
AI documentation offers one example. The technology may not directly make clinical decisions, but widespread use can still create significant operational dependence.
“If thousands of clinicians have redesigned their workflows around it, an extended outage can have significant downstream consequences for documentation, clinician workload, throughput, and potentially care delivery,” Dr. Miller said.
Daniel Kortsch, MD, associate chief AI and digital health officer at Denver Health, framed the question around whether a workflow can continue without the AI.
He separates AI tools into two categories: those that assist people who can still perform the work manually and those that have taken on work no one is routinely doing anymore.
“The trigger for needing a continuity plan is the moment the AI tool stops being an accelerator and becomes the only path,” Dr. Kortsch said.
Ambient documentation falls into the first category, he said. If the technology stops working, clinicians can return to manual documentation. The process may take longer and reduce capacity, but the underlying work can continue.
The situation changes when organizations route substantial volumes of work into an automated process and no longer maintain the same level of human capacity behind it.
“If that system goes down, the work does not disappear, it returns to a team whose capacity may already be fully committed,” Dr. Kortsch said.
That makes fallback planning increasingly important as AI takes on larger roles in clinical and operational workflows.
For documentation, the fallback may be the process clinicians used before AI. For AI-assisted inbox management, organizations need to know how messages will be triaged without automation. For decision support, the underlying clinical information and conventional tools should remain accessible whenever possible, Dr. Miller said.
Cedars-Sinai focuses on establishing three elements before scaling AI: clear communication, clear processes and clear return procedures.
Clinicians need to know when a service is unavailable, what alternative workflow to use and when normal operations have resumed, Dr. Miller said.
The return to normal operations can present its own risks.
Queued messages, partially generated documentation, delayed tasks or AI outputs created around an outage may need to be reconciled once a service returns.
Without a process for addressing that backlog, organizations could introduce new problems during recovery, Dr. Miller said.
AI is also expanding what hospital leaders must count as a technology failure.
Traditional downtime planning often centers on whether a system is available or unavailable. An AI-enabled application, however, may remain online while the AI behind it slows down, loses access to data or begins performing differently than expected.
An AI tool integrated with an EHR could be affected by an EHR outage, a vendor outage, a model provider outage, an integration issue or a change in the underlying model or data, Dr. Kortsch said.
Some failures may be harder to recognize because they do not trigger a conventional outage alert.
An AI system used to summarize clinical information, for example, could begin generating less accurate or useful summaries after a change to the model or its data while appearing to operate normally, Dr. Kortsch said.
“That is a failure mode the industry is still learning how to identify,” he said.
Cedars-Sinai is working on an AI dependency map that identifies which workflows use particular models, which vendors and cloud services support them, what data and interfaces they depend on and what happens clinically if any part of that chain fails.
Dr. Miller said organizations should also expand tabletop exercises beyond traditional scenarios such as ransomware and EHR downtime.
Those exercises could include the loss of a model provider, an AI service becoming unreliable or several AI-enabled workflows failing at the same time.
The potential for less visible failures is also changing how organizations evaluate AI before deployment.
Dr. Kortsch said Denver Health asks what feedback mechanisms are available to identify not only obvious technical problems but also issues related to data and model drift.
For Dr. Kortsch, that means looking beyond whether a tool is ready to launch to how it will be monitored after deployment.
“As an industry we have spent years building rigor around pre-deployment validation and comparatively less on what comes next,” Dr. Kortsch said.
AI requires ongoing monitoring after implementation, creating recurring work and costs that health systems have to account for as their dependence on the technology grows. That responsibility may fall to the organization rather than the vendor, he said.
For Ms. Hatchett, AI continuity planning ultimately extends a principle health systems have already applied to other critical technologies.
“However, while downtime is something we hope to avoid, we must be realistic that it happens,” she said.
