Most physicians have tried using AI for a clinical question by now. You open ChatGPT, type something in, and get back a response that sounds confident, cites no sources, and may or may not reflect Canadian practice. It’s useful enough to be tempting and unreliable enough to be dangerous.
We’ve spent the past year building something different. Asklepius is a clinical AI assistant from Virtual Hallway, designed for how physicians actually think through clinical problems. This post explains what it is, why we built it the way we did, and how to get the most from it.
What general-purpose AI gets wrong about clinical questions
Large language models are remarkably good at generating plausible text about medicine. That’s the problem. Plausible and accurate aren’t the same thing, and in clinical work the gap between them matters.
When you ask a general-purpose AI a clinical question, several things tend to go wrong. The model draws on American guidelines when you’re practising in Canada. It can’t tell you where its answer came from. It has no sense of which clinical context you’re operating in, so it treats a straightforward textbook question the same as a nuanced management decision that really needs a specialist’s input. And it has no idea when it’s out of its depth.
That last part is the one we think about most.
How Asklepius works
Asklepius is a clinical AI assistant you can talk to the way you’d talk to a knowledgeable colleague. Ask it a clinical question in plain language and it will give you a sourced, structured response grounded in Canadian clinical practice guidelines.
The core experience is a conversation. You ask a question, you get an answer with sources, you follow up. But unlike a general-purpose chatbot, Asklepius is built on clinical knowledge infrastructure. It knows what the CPS says. It knows what Canadian guidelines recommend across specialties. And it’s transparent about what it’s drawing on, so you can evaluate the answer the way you’d evaluate any clinical information: by looking at where it came from. When it’s working from broader evidence rather than a Canadian guideline, it tells you that too.
We’ve also built a dedicated Drug Mode for pharmacotherapy questions, with structured tools for drug information, interactions, comparisons, switching protocols, indications, and deprescribing. Drug Mode draws on CPS monograph data and specialized clinical sources matched to each question type. We’ll write more about Drug Mode in a future post.
An AI that knows when to stop being an AI
Asklepius is built on Virtual Hallway. That means it sits inside a network of over 10,000 physicians, including specialists across every major discipline. We are training Asklepius to recognize when a question exceeds what any AI should answer alone, and to escalate to a real specialist consultation through the Virtual Hallway network.
We call this escalation intelligence. It’s the difference between a system that gives you its best guess regardless of confidence and one that says “this is a question you should discuss with a cardiologist, and I can connect you to one now.”
Most AI tools have no way to act on the realization that a question needs a human. Asklepius can, because it’s built into a consultation network where the next step already exists. This capability is still in the research phase, but the architecture is in place: an AI that knows when to hand off, connected to a network where the handoff works.
How we hold ourselves accountable
We run ongoing validation studies where specialists review Asklepius responses against real clinical questions. This isn’t a one-time benchmark. It’s a continuous process where specialist feedback directly improves the system. When a respirologist tells us an Asklepius response about COPD management missed a nuance, that feedback gets incorporated.
In our initial validation with 14 physicians across 70 evaluations, physicians preferred Asklepius over ChatGPT and OpenEvidence 59% of the time (Cohen’s d of 1.1). More useful than the headline number is what drove the preference: source transparency, Canadian clinical context, and responses that matched how physicians actually think through problems.
We’ll continue to publish validation data as it matures. If we’re going to ask physicians to use this tool, we owe them ongoing evidence that it works.
How to get the most out of Asklepius
Ask it like you’d ask a colleague. “67-year-old on apixaban and amiodarone, new diagnosis of CKD, how should I adjust” will get you a better answer than “apixaban renal dosing.” Give it the context that would change your management.
Look at the sources. Every response is grounded in identifiable clinical sources. If it draws on a Canadian guideline, check whether the recommendation applies to your patient. We’ve built the tool to make that easy.
Use your judgment, then tell us. Asklepius is not flawless and no clinical AI system ever will be. If something doesn’t match your clinical reasoning, trust your reasoning. Then flag it, because that feedback is how the system improves.
Notice when it escalates. When Asklepius suggests a question warrants a specialist consultation, that’s the system working as designed. Sometimes the best clinical decision support is a real conversation with the right specialist.
What we’re building toward
Asklepius is launching with clinical chat grounded in Canadian guidelines and a Drug Mode for pharmacotherapy. The vision is broader than any single feature, but the principles stay the same: Canadian context by default, source transparency, ongoing specialist validation, and the ability to escalate to a real physician when AI isn’t enough.
If you’re on Virtual Hallway and want to try Asklepius, it’s available now in closed beta. If you find something that doesn’t look right, tell us. That’s how it gets better.