The risk in clinical AI does not end when the algorithm produces an answer. CARA uses a patient-specific map to help clinicians navigate near the cardiac conduction system. Suppose the displayed guidance is stale or wrong, but still looks credible. A clinician acts on it during an intervention. The model has a name for one part of that sequence: clinical automation bias. It estimates a $6.39 million single-loss expectancy for that modeled event. The figure is an estimate from public information, not a measured CARA loss. The decision it points to is real for any company building procedural guidance: What evidence tells the clinician when to trust the display, when to question it, and when to stop using it? A risk model should help the product and clinical teams answer that together. A number alone cannot.

A patient-specific AI model is only valuable if it stays connected to the right patient. CARA’s workflow begins with a CT angiogram. A clinical expert prepares the 3D model. The model then supports guidance during an intervention. Now follow the handoffs: patient → scan → model → procedure. If the wrong patient’s anatomy enters that chain, the guidance could be internally consistent and still be wrong for the person on the table. Our model calls this a patient-data mismatch and estimates a $3.26 million single-loss expectancy. This is a modeled scenario, not a finding about CARA’s actual workflow. The countermeasure question is specific: where is patient identity checked across the full chain, and what stops guidance when the match cannot be confirmed? That is why I like writing risk as a story. It shows exactly where a control has to work.

The most dangerous guidance can look perfectly precise. The CARA System builds a patient-specific map of the cardiac conduction system from a CT angiogram, then overlays it on live fluoroscopy during a procedure. Imagine the map is correct during planning. During the procedure, the overlay drifts. It still looks precise, so the clinician continues to use it while working near the conduction system. That is a plausible loss event, not a reported CARA incident. In our public-information risk model, “overlay drift” has an estimated single-loss expectancy of $7.2 million. The number is a model estimate. The useful question behind it is more concrete: How would the team know the overlay had stopped pointing to the right place—and what would the clinician do next? That question connects software performance, clinical judgment, and the fallback workflow. It is the kind of question a list of technical requirements can miss.

Tired and tested: Can AI help you sleep better? Onera SleepMap is an AI-enabled device that offers a cloud-based, automated analysis of PSG signals with high accuracy, unlike competitors relying on third-party software. What is the cost of a single loss event due to bugs in the AI? @Onera Health has competitors like Somnoware - Somnoware Diagnostics, X-trodes - X-trodes Sleep System and Neurobit - Neurobit PSG What is the loss in $ of Physiologic signal artifact misclassification ? What is the loss in $ of Population performance drift ? What is the loss in $ of Respiratory event underdetection ? What is the loss in $ of Respiratory event overcalling

Do you have a FDA-cleared medical device? Your attackers, customers, and competitors are moving faster than you. We've built 1501 risk models using Open CRO the AI chief risk officer Reply and I'll send you your risk model

An elephant never forgets Have you ever been in a busy public place - looking for car keys parking ticket, passport? And you could not find for the life of you? Here's an idea for a cool agent product - "The elephant has the keys" This morning my wife and I were in a shopping mall. She needed the house key. I took it out of my backpack and handed it to her. She put it in her bag. Ten minutes later neither of us could remember where the key was. Two intelligent adults. Both present for the event. One handed over the key. The other received it. And neither of us remembered. As we grow older, this kind of memory failure becomes more common. Especially in noisy public places where your attention is degraded. The problem isn't necessarily remembering where you put something. The problem happened earlier. We never really encoded the event. So I thought of a different handoff. When I give my wife the key, we can imagine an elephant standing between us. The elephant takes the key from me. Then the elephant hands it to her and she puts it in her bag. We both see the same ridiculous image. Now try forgetting the elephant. The interesting thing isn't the acknowledgment. I could say, “I'm giving you the key.” She could say, “Got it.” And ten minutes later we might still forget. The elephant is what makes it work. It turns an ordinary, forgettable event into something vivid and absurd. That got me thinking about AI agents. What if an agent observing two people didn't just remember things *for* them? What if it helped them form the memory in the first place? It sees an important handoff. Keys. Passport. Medication. Tickets. A document. And at the moment of transfer it creates a strange shared image that reinforces the event. Not: “Danny gave Ofra the key at 10:37". But: “The elephant just put the house key in Ofra's bag.” The AI doesn't become your memory. It helps your memory do its job. I think there's a design pattern here: Use elephants to reinforce memory Because it's an absurd stupid image you cannot forget

You can roll up meaningless numbers Or drill down to the most severe single loss event that can hit your business. A medical device can have dozens of credible loss scenarios. A company can have dozens of products. Adding all those numbers together doesn't give you enterprise risk. It gives you a very large, mostly meaningless number. So I don't do that. For a company like Kestra Medical, OpenCRO can show the top single loss events across its products, ranked in dollars. Then you can drill down: Company → product → threat → loss event → countermeasure. No enterprise risk score. No roll-up.

We love big numbers $89.2M is a big number. But it doesn’t answer 3 big questions: • What creates the exposure? • What should I do first? • What happens if I do nothing? For the Kestra Medical ASSURE System, we modeled the loss events in dollars and mapped them to prioritized countermeasures. Change the countermeasures and see the residual exposure. That's how we make risk analysis an operational tool - not just regulatory report.

Is device risk created by device superiority? The American Heart Association says that 430,000 people die at home from sudden cardiac arrest every year in the US. Rapid defibrillation is the only proven therapy to reverse sudden cardiac arrest. I analyzed the Kestra Medical ASSURE System against competitors — ZOLL LifeVest and Element Science Jewel Patch. What stands out is ASSURE's garment-based approach to patient comfort and compliance. The flip side is the risk created by that superiority. Using public filings, we modeled the financial impact at $89.2M for a single loss event based on a $1.96B valuation.

Why don't healthcare organizations do more to improve their production systems quality? Let’s examine commitment to quality at three levels in an organization: end-users, development managers and senior executives. Users are conditioned to accept unreliable software on their desktop. Development managers are inclined to accept faulty software as a tradeoff to meeting a development schedule. Senior executives, while committed to quality of their own products and services, do not find security breaches sufficient reason to become security leaders with their enterprise systems because: They usually receive conflicting proposals for new information security initiatives with weak or missing financial justifications. The recommended security initiatives often disrupt the business. And who likes to disrupt their business? Now that's an executive decision.