Topflight Apps

AI in EHRs often fails because the focus is on the product, not the provider’s workflow. You’re building inside someone else’s system so no matter how great your UI is, it won’t matter if it pulls providers out of their EHR. The key to success? Scott Rossignol at Topflight suggests meeting providers within their existing workflows. While Smart on FHIR and CDS Hooks offer potential, API-based write-backs often prove more effective by embedding insights directly into the tools providers already use. The golden rule in healthcare tech: never add extra steps. Your solution should fit seamlessly into the provider’s routine—not interrupt it.

Creating a chatbot with LLMs is one challenge. Designing one that meets the demands of regulated industries like healthcare or finance is a whole different game. Topflight's Terry Woodward shares this high-level process to make it work: 1. Start with the use case. Defining what you're solving for drives everything else. 2. Map your data needs, especially where PHI or sensitive data is involved. 3. Choose the right infrastructure. This includes vector databases, memory handling, and agentic frameworks that support context retention. 4. Build in traceability. You'll need robust logging and visibility for audits, QA, and continuous improvement. 5. Plan for escalation. Not everything should be automated. Knowing when and how to bring in a human is essential. This kind of structure isn't just technical hygiene. It's how you build reliable, compliant, and scalable AI systems that people trust.

Smart on FHIR requires careful implementation to succeed in clinical settings, shares Scott Rossignol from Topflight. When does a standalone UI make sense? Three specific scenarios: for sensitive information that shouldn't become part of the permanent record, for workflows extending beyond the EHR, or when your application manages a complete sequence of actions that significantly reduces provider time. What providers reject: clicking a button, waiting for your app to load, viewing information, then manually incorporating it elsewhere. This exchange doesn't justify the workflow interruption. The better approach? Embedded views that allow for side-by-side navigation between your application and EHR functionality. This creates a natural workflow for functions like ambient listening where providers need to move between systems.

When you're building a chatbot for a specific domain - especially something high-touch like telehealth - there’s always a tension between structure and flexibility. Topflight's Preston Hoang discusses the importance of structure to guide the conversation toward an outcome. In a scheduling flow, for example, there are certain pieces of information you must collect: name, phone number, appointment reason. Without structure, you risk never getting the data you need. But you also want to preserve the open-ended, conversational feel that LLMs make possible. That’s what makes the experience feel natural and human. The best systems blend both. You keep the guardrails in place to ensure reliability, but allow for enough flexibility so the chatbot doesn’t feel robotic. 

As the demand for imaging services grows, healthcare organizations face radiologist shortages, delaying patient care. Ryan Majoria, M.D., a practicing radiologist in Baton Rouge, encountered this firsthand. His radiologists were overwhelmed by surges in demand with no way to efficiently bring in extra staff or streamline compensation. Ryan needed a solution that could dynamically connect radiologists to available work, distribute studies efficiently, and ensure prompt payments—all while maintaining HIPAA compliance. That's where Topflight came in.

Clinical decision support in EHRs has been failing for 20 years. AI won't magically fix that. The sobering reality: generating recommendations doesn't mean you can deliver them effectively to providers. You might suggest an order but can't help place it, or recommend against a treatment the patient already received at another health system. What separates success from failure? User activation. Scott Rossignol from Topflight shares this key insight: You get ONE chance to prove your value to busy clinicians. If they try your solution and don't immediately see benefits, they'll never return. This moment in healthcare AI is a once-in-a-lifetime opportunity.  The organizations that succeed understand that perfect algorithms mean nothing without flawless implementation and provider adoption.

If you're building with LLMs, especially in high-stakes industries, reducing hallucinations and managing risk is foundational. Terry Woodward from Topflight shares the importance of a layered approach. First, look at the built-in guardrails available through the cloud provider where the application is hosted. These native controls can offer a strong baseline for protection. Next, bring in a specialized guardrail software that monitors the context of the conversation itself - checking whether answers are grounded in the right provenance and align with what should actually be shared. Finally, leverage vector databases where needed. Some platforms, like Vectara, offer built-in hallucination reduction or elimination features, which adds another layer of reliability. The right stack depends on your application, your data, and your risk profile. But getting this right is key to building trustworthy AI experiences.

The future of healthcare interoperability isn't APIs - it's AI agents talking to each other. Imagine AI agents representing your application and the EHR negotiating data exchange in real-time, customizing payloads based on specific needs rather than rigid predefined APIs. But two major roadblocks stand in the way. First, EHR vendors may resist this approach, fearing third-party applications might replace profitable modules. Second, outdated HIPAA regulations like "Minimum Necessary" create inconsistent implementation requirements - one health system allows full data access while another restricts it severely. According to Scott Rossignol from Topflight, what is needed is an industry consensus on data-sharing standards that balance security with functionality, eliminating the burden of maintaining multiple integration approaches.

One of the biggest challenges in building AI agents for healthcare isn’t the model itself - it’s the fragmented data infrastructure, says Topflight's Preston Hoang. That’s why developments like Model Context Protocols (MCPs), currently being explored by Anthropic and others, are worth watching closely. MCPs create a layer of standardization that makes it easier to pull in relevant data from different sources. When you have that kind of structure in place, it lowers the barrier to integrating diverse datasets without forcing every source to conform to the same rigid format. In industries like healthcare, where data silos and legacy systems are the norm, this kind of interoperability could be key to unlocking more intelligent, responsive, and effective AI agents - especially those built on RAG or LLM foundations. It’s not just about better models. It’s about making better data accessible.

API-based healthcare data integration breaks down when you need volume. A single patient record can require numerous individual FHIR queries. Multiply that across your patient population with monthly refreshes - you're pushing beyond practical limits of API-based approaches. According to Topflight's Scott Rossignol, this creates two critical problems: excessive load on the EHR system and increased potential for errors or timeouts. The solution requires precision: use APIs only when your LLM needs real-time data inputs. For large historical backloads, you need specialized approaches to scale both your application and the health system's API capacity. While we've successfully executed millions of API calls to retrieve healthcare data, it requires meticulous planning and continuous monitoring throughout the process.