AI Can Find the Signal. Healthcare Still Has to Act on It.
A few reflections from my time at the 2026 Cleveland Clinic AI Summit
AI can recognize a suspicious lesion, identify a missed screening, flag a medication conflict, summarize a medical record, and draft a response to a patient. Its ability to find patterns in enormous volumes of clinical data is advancing at an extraordinary pace.
But one lesson stood out to me at the 2026 Cleveland Clinic AI Summit: finding the answer is not the same as improving the outcome.
Once AI produces an insight, healthcare still has to decide whether the insight is right, determine what should happen next, identify who is responsible, and place the right information into that person’s workflow. In many cases, the work has only just begun.
AI as a sparring partner, not an admirer
The conversation at the summit was appropriately optimistic about AI’s potential. Health system leaders were deeply engaged in exploring how it could improve safety, patient experience, costs, and the care team experience. They were also candid about its limitations.
One of the most important problems is the conversational AI's agreeable nature. These systems are designed to be helpful, engaging, and responsive. Without careful prompting and critical review, that tendency can become a form of confirmation bias: the system accepts the premise of the question, reinforces the user’s viewpoint, and produces an answer that sounds more certain than the evidence warrants.
That is especially consequential in healthcare. A reassuring, fluent response may still encourage unnecessary worry, overdiagnosis, inappropriate testing, and the wrong next steps. The problem is no longer simply whether an AI system can generate a plausible answer. It is whether the human using it knows how to challenge the answer.
We should teach clinicians, students, executives, and patients to use AI as a sparring partner: ask it to identify weaknesses, argue the opposite case, surface missing evidence, explain uncertainty, and say what would change its conclusion. AI literacy must include disciplined disagreement.
Augmentation can become dependency
The same principle applies to clinical skill. AI can extend human capability, but poorly utilized, use can also weaken it.
Research discussed at the summit included an observational study of colonoscopy practice. After clinicians had routine access to AI-assisted polyp detection, their adenoma detection rate during unassisted procedures declined from 28.4% to 22.4%. The study does not prove that every AI tool will de-skill every clinician, but it raises a serious implementation question for health systems: how do we gain the benefits of assistance without surrendering the expertise needed to evaluate it?
Education offers a similar lesson. AI that simply supplies answers can reduce the productive struggle through which people learn. AI used as a tutor—asking questions, coaching reasoning, and providing feedback—can support learning instead. The decisive issue is not merely access to AI. It is the role we design AI to play.
Health systems may need a formal AI education function, much as they have established disciplines for quality, safety, informatics, and compliance. Stanford’s appointment of Jonathan Chen, MD, PhD, as Director for Medical Education in Artificial Intelligence reflects the kind of institutional capability that will become increasingly important. Organizations need people responsible not just for approving tools but also for teaching caregivers how to use them without eroding their judgment.
If sanctioned AI is unavailable, caregivers will find their own
There is also a practical reality health systems cannot ignore: caregivers are already using conversational AI. If appropriate, sanctioned tools are not available, some will turn to public applications on personal devices. Entering protected health information into an unapproved service creates serious privacy and security exposure.
The answer is not to pretend use can be prohibited out of existence. Health systems need governed alternatives, clear policies, training, and workflows that make the safe option the useful option. The rapid adoption of ambient documentation offers a helpful example. Where the tool fits naturally into clinical work and requires little training, adoption can move quickly. Integration elsewhere remains far more difficult.
AI can answer the ‘what.’ Humans remain essential to the ‘so what?’
AI is increasingly capable of matching or improving upon clinicians on bounded tasks involving facts, pattern recognition, differential diagnoses, or test-taking. Yet care requires more than producing the ‘what.’
A clinician sees the whole patient, including what is not captured in the prompt. A clinician understands context, values, history, urgency, and the consequences of acting or not acting. A clinician can say, ‘No, you do not have any of these worrisome diagnoses,’ when a system optimized to be helpful continues generating (often worrisome) possibilities. A clinician can align a recommendation with what a particular patient needs and is prepared to do.
The same distinction helps explain why patients sometimes rate AI-generated messages as more empathetic than physicians’ responses. In a well-known JAMA Internal Medicine study, evaluators preferred chatbot responses and rated them higher for empathy. That does not mean a machine experiences empathy. It may mean that an AI has unlimited patience to provide a longer, reassuring, and mildly flattering explanation, while a physician is answering dozens of messages between other responsibilities.
That is an important design signal. AI may help caregivers communicate with greater clarity and completeness. It should create more room for human judgment and connection—not become a substitute for either.
Every AI insight creates a communication problem
Much of the healthcare AI conversation focuses on analysis: Can a model detect hidden cancer risk? Identify a patient who has missed a screening? Detect poor or incomplete management of hypertension, diabetes, or heart failure? Recognize a medication conflict? Reinterpret an image obtained for another purpose?
Those are valuable capabilities. But most of their value remains theoretical until someone acts.
Imagine an algorithm analyzing clinical and claims data and identifying patients who appear overdue for colorectal cancer screening. The output may be accurate. It may even be prioritized perfectly. Yet the system must still determine the appropriate next step. Should it notify the primary care physician, an advanced practice provider, a population health nurse, a scheduling team, or the patient? Which organization currently owns the relationship? What communication channel does that recipient actually use? What information must accompany the alert? How will anyone know whether it was received and acted upon? And how can we overcome or mitigate the very real alert fatigue identified across decades of formal informatics science?
A similar problem follows nearly every operational AI use case. The model’s output must be incorporated into a communication flow or a managed workflow. The right care-team member needs the right information, in the right format and location, with a clear next action.
This is the last mile of healthcare AI, and it is not solved by intelligence alone.
From insight to action
Operational AI depends on data moving in both directions. It may need local EHR data, records from outside organizations, a patient-clinician conversation, claims history, imaging, lab results, and other fragmented inputs. AI is particularly useful when those inputs are poorly structured or have a low signal-to-noise ratio.
The resulting insight must then become appropriate action.
When the appropriate next step is complex or requires longitudinal follow-through, the AI output may need to create a task in a coordinated workflow, with ownership, deadlines, dependencies, and escalation. That is the kind of work careMESH NAVIGATE enables.
When the next step is a notification or request, the information may need to reach a provider, an advanced practice provider, a nurse, a referral coordinator, a scheduler, or another member of the extended care team. It must arrive through a channel the recipient can use, in a format that fits the receiving workflow, with visibility into whether delivery succeeded. That is where careMESH CONNECT, and the provider and communication intelligence behind it, becomes essential.
The opportunity for AI in healthcare is enormous. But we should judge implementations by more than model performance or the elegance of a generated answer. We should ask whether they expand or at least preserve human expertise, support safe use, fit into clinical workflows, or easily enable new ones, and close the loop between insight and action.
AI can find the signal. Better outcomes begin when the right person receives it, knows what to do next, and actually does it.
Sources and editorial references
Cleveland Clinic. 2026 AI Summit, Cleveland, OH (and online), August 28, 2026.
Budzyn, K. et al. “Endoscopist deskilling after exposure to artificial intelligence in colonoscopy: a multicentre observational study.” The Lancet Gastroenterology & Hepatology 10, no. 10 (2025): 896-903. https://doi.org/10.1016/S2468-1253(25)00133-5
Ayers, J.W. et al. “Comparing Physician and Artificial Intelligence Chatbot Responses to Patient Questions Posted to a Public Social Media Forum.” JAMA Internal Medicine 183, no. 6 (2023): 589-596. https://doi.org/10.1001/jamainternmed.2023.1838
Stanford Medicine. Jonathan H. Chen, MD, PhD faculty profile. https://med.stanford.edu/profiles/jonc101
Contact careMESH today to learn how careMESH can accelerate your product's interoperability strategy.
