Building a Technology-First Healthcare Company with Matthew Woo of Summer Health
This is part of our executive insights series where Elion CEO Bobby Guelich speaks with healthcare leaders about their tech priorities and learnings. For more, become a member and sign up for our email here.
Role: Co-Founder & President
Organization: Summer Health
Can you start by telling me a little about your background and your approach at Summer Health to technology?
I’m the co-founder, president, and chief product officer at Summer Health, where I oversee product, engineering, and parts of business operations. Summer Health is a telehealth pediatric service. We started in cash pay but have recently expanded into payers, particularly Medicaid.
Because we see ourselves as a technology-first healthcare company, we’ve taken a vertically integrated approach to AI. We built our own EHR, directly integrate models into our workflows, and use AI throughout the chat experience and beyond, so we really own the full experience for the parents that we serve on our platform.
Where are you applying generative AI today? How has that evolved over the past year?
We started with medical note documentation. Since we’re an asynchronous-first platform, we already had full chat transcripts, which made it easy to improve note completion speed and quality.
Since then, we’ve expanded into:
Patient engagement: AI helps draft personalized follow-ups, care plans, and messages for providers to review before sending. From there we’ve begun to fine-tune AI models based on each provider’s tone and style (e.g., formal vs. casual, using first names or not).
AI co-pilots for session management: We’re exploring how AI can assist doctors in real-time by surfacing relevant patient data. For example, if a doctor is about to prescribe a certain medication, the AI can pull the child’s most recent weight and recommend a dosage.
What lessons have you learned about building with AI over the past year?
The biggest surprise has been the variability in model outputs. It’s tempting to always upgrade (e.g., GPT-3 to GPT-4 and so forth), but sometimes newer models introduce changes that make outputs worse for providers.
We’ve addressed this by tracking provider acceptance rates of the AI-generated outputs. Going forward, we’re integrating a clinical review evaluation framework to ensure quality before full deployment. Because we always have a human-in-the-loop and have a self-reinforcing fine-tuning layer on top, we can quickly course-correct when needed.
You’re obviously very build-first, but are there areas where you’ve chosen to buy instead?
Every time we consider buying, it seems a new model comes out that makes building more feasible. The only AI-related things we purchase are foundational model access (GPT, Claude, DeepSeek, etc.) and development tools. Owning the interface gives us more flexibility—buying often requires heavy integration, and you’re beholden to the vendor’s development pace.
How extensible is your approach for larger, incumbent healthcare organizations?
It depends on three things:
Control over the experience: Do they own enough of the patient/provider workflow to inject AI meaningfully?
Technical capability: Do they have a development team that iterates continuously, or is it project-based with long cycles?
Risk appetite: AI can be safely integrated with human oversight, but incumbents need a culture shift to move faster.
Many larger organizations are focusing on narrow AI use cases, which makes sense. As a startup, we have to be vertically integrated to win. That level of ownership allows us to move faster and redefine care models.
Given the capabilities that AI is enabling in making development faster, has the type of talent you look for changed?
Yes, across all roles. Within user research, AI automates surveys, transcribes interviews, and even conducts simple user research conversations, so we’re starting to experiment with getting more frequent insights on smaller sets of questions. Then when it comes to design, you can rapidly mock up client-side apps using AI-powered tools, speeding up validation.
Of course, when it comes to engineering, everyone is talking about how AI accelerates coding, so we’ve adapted our hiring process. Instead of barring AI in technical interviews, we encourage candidates to use it and see how far they can get in an hour.
One thing we’re still figuring out is whether we need dedicated AI engineers. Initially, we thought we’d hire ML engineers, but given how fast models improve, an applied ML/ops role might be more valuable.
Anything else you think is interesting right now?
One thing I’m thinking about right now is obviously GenAI is really great for a lot of reasoning tasks, but there's sometimes a trap of thinking that GenAI and LLMs can do everything. There are other ML approaches that have proven to be very effective that we should be leveraging as part of a whole suite versus just throwing everything into GPT.
Have you experimented with any more agentic approaches?
Not yet, but we see potential in care coordination. For example, how can Summer Health go beyond virtual care and actively assist in referrals and care transitions? That’s an area where agentic AI could be valuable.