From Chaos to Control: Dr. Mo Canellas on AI, Emergency Medicine & Why Most “AI Companies” Fake It

Dr. Maureen "Mo" Canellas wears a lot of hats. She is the associate chief medical officer at UMass Memorial Medical Center, a researcher in machine learning and health care operations, a faculty collaborator at MIT, and a longtime contributor to the Emergency Department Benchmarking Alliance. But the thread connecting all of it traces back to a college thesis on the Google search algorithm and a question she has been asking ever since: how do you bring order to chaos? That question is what drew her to emergency medicine, and it is what keeps her at the center of serious conversations about AI in hospital operations.

Canellas sat down with Dr. Kevin Stevenson on the I Don't Care podcast to talk through where health care AI actually stands, what separates real innovation from slide-deck theater, and why the financial structure of hospitals may be the most uncomfortable problem that data science still needs to solve.

Most "AI companies" are really doing data science at best

Canellas did not mince words on the state of vendor claims. The field is crowded with companies that attach the AI label to products that rely on little more than basic analytics. "AI is what you tell your CEO you're doing," she said. "Machine learning is what you tell other people you're doing. And what you're really doing a lot of the times when they're AI enabled, it's data science. It's not, like, a linear algorithm at best." Her test for separating credible vendors from the noise is straightforward: skip the polished pitch and ask granular questions. If someone can think on their feet about data inputs, outputs, and model mechanics, that signals genuine understanding. If they can only navigate the slides they rehearsed, that tells you something too.

For hospital CEOs evaluating AI investments, Canellas recommends bringing a cross-functional team into every vendor conversation rather than letting a single executive be the sole judge. She also points to the Coalition for Health AI, known as CHAI, as a practical resource. The organization is building governance and encryption standards for health care AI by mapping from established frameworks like IEEE and aligning with existing regulatory bodies, including the Joint Commission, which has already published a white paper on responsible AI. Canellas believes Joint Commission surveys will eventually require evidence of CHAI alignment, making early adoption a competitive advantage rather than a compliance burden.

Clinical intuition, patient flow, and the limits of current tools

One of the more provocative ideas Canellas raised is how AI might capture what emergency physicians call clinical gestalt, the experienced gut sense that something is wrong with a patient before the data catches up. Rather than waiting for AI to replicate that instinct autonomously, she proposed a simpler bridge: prompt the physician at the start of an encounter with a single binary question about likely admission. Feed that response into the algorithm. "Why not take that gestalt and see if that improves the outcome?" she asked. "That way the docs are involved, increased adoption, increasing using their brainpower to help the algorithm." It is a practical, incremental approach that reflects her broader philosophy: build AI with clinicians, not around them.

Her work on patient throughput connects directly to a problem she described as one of the most uncomfortable topics in any hospital boardroom: boarding. When patients who need inpatient beds remain parked in the ED, it strains every downstream resource. Canellas has researched the cost differential and presented data-driven financial models to make the case for investment, arguing that simulation AI can project a return on investment before a single dollar is committed. The discomfort, she noted, comes from the fact that the revenue structures governing inpatient and emergency care are fundamentally misaligned, and fixing boarding requires confronting that misalignment directly.

Where health care AI stands and what comes next

Asked to put a baseball inning on health care AI's current progress, Canellas called it the second or third inning. Enough early runs to generate real interest, but far from the decisive middle innings where the best tools separate themselves from the field. She sees the proliferation of AI products across every conceivable use case as analogous to going through a batting order once: now the game starts for real, and teams will begin identifying what actually works. The institutions that will benefit most are those that have already done the foundational governance work, built internal AI literacy, and selected vendors who can answer hard questions without a script.

For health care operators, the practical takeaway from Canellas is consistent across every topic she covers: resist the rose-colored glasses, demand specificity from vendors, invest in standards before regulators require it, and treat AI as a tool that works best when clinicians remain active participants in the loop. The chaos of emergency medicine, she has argued since graduate school, is not a problem to be eliminated. It is a system to be understood well enough to control.