Nov 6, 2025

Andrew Crowder, Vice President of Engineering at Carta Healthcare Generative AI is no longer a novelty, it’s a necessity. For healthcare leaders, the conversation has evolved beyond proof of concept. The real challenge now is integration: embedding these models into clinical workflows in ways that are safe, scalable, and operationally sound. Success demands more than impressive outputs. It requires systems built for healthcare’s complexity, with the agility to support frontline decisions and the clarity to preserve the human judgment at the heart of care. For most people, AI begins and ends with chat. Whether it’s ChatGPT or a hospital support bot, the pattern is familiar: ask a question, get an answer. It’s intuitive, but it’s not enough for healthcare. Clinical environments are flooded with complex, high-stakes data, much of it unstructured and constantly evolving. One-off prompts can’t keep up. What healthcare needs is agentic AI : systems that don’t just respond, but proactively observe, synthesize, and act. These models operate quietly in the background, parsing records, surfacing insights, and resolving nuanced clinical questions, without waiting to be asked. That’s the shift from interaction to intelligence. Context is King: Why Information Flow Shapes Model Performance Even the most advanced AI models need guidance, and that guidance comes from context. In healthcare, context is anything but simple. It’s not just what data you provide, but how, when, and in what form. Clinical systems generate vast volumes of information: vitals, free-text notes, repeated entries, and layers of noise. The answer to a critical question might be buried somewhere in that mix, but without clear direction, the model won’t find it. To deliver meaningful results, AI must be engineered to navigate complexity, knowing not just what to analyze, but where to look and why it matters. The real challenge is knowing what information to send to the model and when. A single medical record often contains more data than a model can process at once, and most of it won’t be relevant to the question at hand. What matters for one task might be meaningless, or even misleading, for another. Therefore, good performance requires an understanding of clinical intent. A patient note might look like a list of facts, but it’s also a narrative. If you don’t capture the right part of that narrative, the response might be technically correct, but clinically useless. True Intelligence is the Ability to Use Tools There was a time when building AI meant writing out rules: if this, then that. Then we moved to training models to find patterns. Now, the next step is giving those models tools and letting them decide how to use them. For example, a model might need to retrieve structured medication data from a database or scan a PDF discharge summary before answering a clinical question. Each of these tools, retrieving, parsing, cross-referencing, helps the model go beyond language and solve more complex problems with greater accuracy. This shift changes the role of a model entirely. Instead of just answering questions, the model can reason about how to answer them. For example, we could feed the model a clinical note and ask whether a patient received a certain medication. The model might read the note and conclude that it lacks enough information. Rather than guessing, it can now go look: query a medication log, check a database, or cross-reference lab results. The key is that it decides. This orchestration is especially important in clinical data abstraction workflows, where responses often depend on multiple sources and subtle context. You need components that can parse documents, fetch data, validate outputs, and move information between steps. A tool-using model is more adaptable. Rigid systems can break under variability. Tool-using systems can flex, retrieve what they need, and return results that are more accurate and durable across use cases. Writing Love Letters to AI: The Art and Science of Prompt Tuning The way a question is phrased impacts how accurately the model responds. Getting it right is less about writing style and more about engineering – testing, refining, and adjusting language to find what works in each scenario. Think of it like writing a love letter, they are personal. Your structure, tone, and even length depend on who you are, what you want to convey, and who’s on the receiving end. Prompt design works the same way. You’re crafting language not only to share information, but to guide behavior. Some tasks require logic; others call for interpretation. As models evolve, the same prompt might perform differently across updates, requiring ongoing tuning and maintenance. Producing consistent results means understanding how language drives behavior in systems built to mimic how we think. Lessons from Scaling AI Integration Scaling any generative AI model brings new challenges. Throughput, latency, and cost are all obvious. In healthcare, the bigger concern is trust. When a model returns a response, clinicians want to know where it came from, whether it’s accurate, and how confident the system is. Studies suggest that trust increases when outputs are explainable, when models are transparent about uncertainty, and when systems are tailored to local data and workflows. Without that trust, even the most accurate models can struggle to gain traction in real-world care. The safest clinical-grade systems have guardrails: workflows that link model outputs to evidence, citations, and an audit trail. This is Hybrid Intelligence : a deliberate division of labor between machine and expert. The model is the engine, and it moves fast. But the human is still holding the wheel, making sure it’s pointed in the right direction. Shaping the Future of Applied Intelligence Intelligence doesn’t come from the model alone. It comes from the system around it: the tools, workflows, people, and decisions that guide how and when the model is used. Deploying AI in healthcare isn’t just a technical challenge, it’s a real-world imperative. Success demands systems that can extract, structure, and validate data at scale, while embedding safeguards that keep clinicians in control. But technology alone isn’t enough. What’s needed are solutions that understand the full spectrum of clinical, technical, and operational complexity, and fit for the demands of everyday patient care. About Andrew Crowder Andrew Crowder leads the engineering team at Carta Healthcare . Andrew is a results-oriented software engineering executive and AI innovation leader. Andrew drives the integration of cutting-edge AI into impactful healthcare solutions. His focus on “Applied AI” delivers tangible efficiency gains in medical data analysis. He champions the transformative power of thoughtful technology in healthcare, always prioritizing the user experience and workflow improvement.