The Age Old Battle of Healthcare
2024 has been a great year for ambient scribing. There’s no question that ambient scribing offers relief from manual note-taking, giving clinicians more time to be present and focus directly on interacting with the patient. Ambient scribes have also been important because they have been the first widely adopted applications of AI in healthcare. This has set early expectations that AI, when used safely, can be both a benefit for clinician efficiency and patient satisfaction, generating real positive ROI. With widespread adoption and the quick entrance of many players across the industry, each scribe is trying to find new ways to provide value beyond "note generation." Unfortunately, the long term utility and viability of most of the scribe solutions, and AI in general, has an artificial ceiling. This ceiling comes from constrained access to patient context, the ability to operate in real time with data and user inputs, and the inability to control workflows or act on the behalf of the provider.
Why such limitations on the potential of ambient scribing? The dreaded EMR.
EMR vendors are (in some cases purposefully) dragging their feet on enabling AI efforts outside their own. Sometimes it’s due to legacy architectures, siloed data, and limited APIs. Many times it's because they want to defend their business from any threat that might disrupt their hold on their customers' data and workflow. When it comes down to it, legacy EMR vendors just aren’t designed to work with AI. AI companies are taking different approaches to get around some of these limitations, but it will be a long struggle to overcome all the blockers. This fundamentally limits the effectiveness, and most importantly, the safety of AI use.
Canvas - the Platform for AI in Healthcare
This is where Canvas stands apart. We’ve built our EMR from the ground up with a server-side SDK that gives developers, AI vendors, and healthcare organizations access to real time event streams and fine-grained control over both the data and the workflows involved in care delivery and payment. This open architecture enables:
- Composable Multi-Model Chaining: Different AI services—like a summarization model, a coding assistant, or an ambient scribe—can communicate within Canvas in real time. For instance, your scribe’s output can feed directly into a summarization tool, which then flags a care gap, which in turn prompts an agent to place an order or schedule a follow-up. Because all these steps happen inside Canvas, there’s no cumbersome copy-paste or external bridging.
- Human-in-the-Loop Control: Our event-driven architecture lets you decide where and how clinicians or staff should approve AI actions. Maybe you want a copilot to propose diagnoses but require a physician sign-off before sending them to billing. Or you want a nurse to validate medication orders flagged by the AI. By exposing all these triggers in the SDK, Canvas enables safe, responsible automation.
- Real-Time Data Access: Canvas’ Narrative Charting system and modern data model allow AI tools to see the complete patient context—diagnoses, medications, past encounters, lab results, etc. —in real-time. This drastically improves LLM accuracy and reduces hallucinations by giving each AI agent the right data at the right moment.
- Pluggable, Best-of-Breed Models: We don’t lock you into a single AI vendor or approach. With our open framework, you can swap in specialized models for cardiology, oncology, or behavioral health, tailoring the technology to your patient population. Think of it as “cognitive composability”—like building with LEGO bricks, except each brick is a different AI specialized for a specific task.
- Command Structure for True Automation: Beyond reading data, Canvas’ “command structure” means AI can place orders, add billing codes, and otherwise direct workflows programmatically. That’s how we move from passive scribing and summarization to active, agentic copilots that can truly transform care delivery.
Hyperscribe: A True Clinical Copilot
Today we are announcing Hyperscribe, a clinical copilot for Canvas. Hyperscribe is like having a real clinical assistant with you while you are seeing patients. Not only does it take notes, it does so with the full context of the medical chart of the patient you are seeing. It captures information in a structured form so it can be acted upon. It is aware of clinical guidelines and programmed to apply them. It works in real time and is able to tee up tasks like ordering labs or scheduling follow up appointments. And so much more…
We were able to build Hyperscribe out of the box using the publicly available Canvas SDK and tools from AI vendors like OpenAI,Google, and Anthropic. All of the capabilities we used internally to build Hyperscribe are the same tools any developer can use to integrate or build copilots. In fact, we are open-sourcing the Hyperscribe extension code so anyone can see how easy it is to build a copilot or plug in their own large language models or extend Hyperscribe with their own domain-specific logic.
Healthtech AI Needs Benchmarks
Not only are we making the source code behind Hyperscribe available to the public — we are also publishing the evaluation code and a growing set of benchmarks of how Hyperscribe performs across various dimensions. Our goal in publishing Hyperscribe evals is two fold. First, the delivery of care is serious business and requires equal rigor in studying the potential impacts of AI on workflows. To get a true sense of how well a system performs, the tests need to be transparent and easily reviewed. There is a lot of gaming when it comes to AI benchmarks. But when it comes to healthcare, we need to raise the bar very high. Second, we are investing to improve Hyperscribe continuously, and there is no way to do that without diligent, reproducible evals to guide those investments. We know foundation models will continue to evolve and we have excellent surveillance on performance over time so we and our developer community can adapt. By making eval code and benchmark data publicly available, we also invite the entire community to contribute more situational and case data to help everyone effectively test and improve their copilots and agents over time.
Safely Compounding Agents Toward a New Era of Effective Medicine
What makes the Canvas SDK so powerful is that each copilot or agent, like Hyperscribe, can work with other copilots and agents, intelligently. The real magic to unlock greater levels of clinical effectiveness and efficiency with AI will happen when summarization tools, scribes, and domain-specific copilots interact with each other—automating administrative burdens, highlighting relevant clinical insights, and gracefully escalating decisions to the right human at the right time. Some examples we see coming this year:
- Integration Overlays: Summaries feed into a coding agent, which adds nuanced reimbursement codes.
- Auto-Referrals: A specialized referral model can parse a summarized note, identify the need for a sleep specialist consult, and auto-schedule it—pending human approval.
- Agentic Workflows: Combine a summarization model with a patient outreach agent that can suggest or even send patient communications about upcoming labs or annual screenings.
With the rise of copilots and agents in care delivery, we must remain clear-eyed about investments in quality control, ensuring that repeated summarizations or agentic actions don’t inadvertently propagate errors. This again is where enabling chaining is critical. With the Canvas SDK, these chained AI interactions are straightforward to implement—and can be precisely configured for each organization's workflow rules and safety checks. One agent can be highly tuned to gather appropriate information from various sources. Another highly tuned agent can take that input and refine the output with appropriate context. A third copilot can reason and present the final output to a clinician for review or start the next chain of agents.
Get Involved and Transform Your Practice
If you’re exploring how multi-model AI can streamline your care delivery or if you want to see Hyperscribe in action, we’d love to connect!
Contact us via the Hyperscribe extension page and let us know if you are interested in trying it out. We’ll show you how Canvas’ open architecture can bring together best-of-breed AI solutions, empower human oversight at critical steps, and deliver the true AI clinical assistant experience clinicians need.
