How the Epilepsy Foundation uses Claude across the organization

The Epilepsy Foundation is a nonprofit serving people with epilepsy across the United States, with a mission to ensure that no one faces epilepsy alone. Its public assistant, Sage, is a Claude-powered AI epilepsy assistant that answers only from the Foundation's medically reviewed content. In addition to the assistant, the Foundation has built Claude into how its own teams work. David-Alexandre Jost, its Chief Technology and Innovations Officer, spoke with Anthropic about writing grants, understanding what the community needs at scale, and what he thinks AI should and shouldn't do at a mission-driven nonprofit. 

Anthropic: Sage only answers from content your editorial board has vetted, and it tells someone when it doesn't have an answer rather than guessing. That's a deliberate stance on what AI should do. Where does it come from?

David-Alexandre Jost, Epilepsy Foundation: We made a decision early on to have authoritative content from the start. There is a lot of unreliable information out there. People who are scared go looking for a cure or a treatment and end up on forums and sites no one has checked. So on epilepsy.com, nothing gets published until our editorial board and our marketing team have approved it, and that is exactly why we could ground Sage in that knowledge base. That grounding is the difference between Sage and someone just opening any AI chatbot and prompting it. What surprised us is how far that grounding goes. Sage started showing behaviors we never programmed, and when we dug deeper, we noticed they were baked into the content we had ingested. Our lifestyle articles offer real recommendations, such as how to handle a situation at school or how to make the most of a few minutes with your neurologist. Sage learned to offer that kind of guidance on its own. We do not have to prompt everything into the agent. A lot of it is already in the knowledge base we built.

Anthropic: You also use Sage to understand what your community needs. How does that work since every session is anonymous?

Jost: Because sessions are anonymous, we do not follow people from one visit to the next, which means a lot of the usual analytics simply are not available to us. So I use Claude Opus to write a weekly report on the behaviors we are seeing inside Sage. It has been a real eye-opener, because we can finally do this at scale. I have been running it for about six weeks, and the report gets better each time. The system has learned the kind of data I am looking for, partly because I give it feedback every week on which conversations to focus on. Now I get genuinely useful insights: which conversations to share with the team because they are a great story, and which ones point to a problem on the prompt side where we need to tighten a guardrail.

Anthropic: What has that weekly read shown you about how people actually use Sage?

Jost: What stood out is how many of our power users treat Sage as a diary. They want it to remember. A lot of people in our community live with cognitive challenges, so a conversation from a few days ago can be genuinely hard to recall, and if Sage can hold onto those details, that matters to them. Right now it can't, because every session starts fresh. That one insight is driving the next version. Our next iteration of this product is Sage Next, which adds authentication and long-term memory, so it can carry something a person shared weeks ago into a later conversation, and even flag a medication risk they might have forgotten to raise with their neurologist.

Anthropic: You bring that same grounding instinct to your own work. The clearest example is the grant you wrote for AWS Imagine. Walk us through it.

Jost: I have been using Claude on a daily basis for about six months, so by the time I sat down to write that grant, it already had all the context of how I work around Sage: how I talk about the technical pieces, how I think about the strategy. Drawing on all of that, I put together a compelling submission in a single day. I could not have done that in a day on my own. Without Claude, I would have likely had to spend 4x as much time on this grant. The value was not only speed. It meant the day went into shaping a strong story about Sage rather than staring at a blank page. There were a lot of lessons in it, and I am taking them back to our core and development teams so more of us can work this way.

Anthropic: What else are you trying to build while piloting Claude across the Foundation?

Jost: The internal rollout of Claude is a trial to see how it fits for adoption across the organization. On the IT side we already have the governance to support it. Where I want to take it next is broader AI fluency. We are planning to use the courses available for nonprofits and teach our core and development teams how to build context in Projects, share that context, and work more collaboratively rather than everyone starting from scratch. The goal is a way of working where you can be confident that what you create with AI is accurate and good enough to share with the community. That is the future I see for the organization: everyone building on the same shared context, whether that is a grant, a piece of content, or the next version of Sage.

Anthropic: You're also making this case beyond the Foundation, at conferences. What are you arguing for?

Jost: We just submitted an abstract to the American Epilepsy Society. If it's accepted, we'll be on a panel in December on how empathy is required when you have a conversational chatbot like Sage, and how important it is to ground it in content you own, with a team behind it that can do the editorial process. I'm also hoping to be on a panel at another conference on the technical side: how you audit the iterative process to get a better chatbot out there, with observability.

Anthropic: All of this runs on sensitive health information. What lets you keep expanding what AI touches?

Jost: It goes back to why we built our foundation on AWS in the first place. We created an environment that can comply with HIPAA and other requirements, where we decide what personal or health information gets shared and what does not, how it is stored, and whether developers or researchers can access it. That is already baked into the system. For us, it comes down to governance: we know how to store this information, and we know how to share it based on the policies we have set.

Anthropic: Across grants, analysis, and everything you're rolling out, what's the principle you keep coming back to?

Jost: Ground it in what you can stand behind, and keep humans in the loop. Everything we have built with Claude, from Sage to the way I write a grant, starts from content and context we trust. Get that right, and the tools compound. The same context that helped me write one grant in a day can help across the organization once people know how to use it. That is what I want for us: an organization where AI helps us do more of the work that matters, and where we stay confident that what we put in front of our community is valuable and accurate.