Inside the Medicare Conversation Agent We Built for Geisinger
Almost a year ago, we wrote on this blog about the potential for large language models (LLMs) to transform static websites into interactive conversations. At the time, it was a strategic argument. LLMs would make it possible to move beyond one-size-fits-all information architectures into experiences that could adapt to each visitor in real time. The premise was that a website could feel less like a brochure and more like a conversation.
Since then, we’ve put that premise to work. Leading up to the 2027 Medicare Annual Enrollment Period (AEP), we built an AI-powered tool that served as a key feature of Geisinger Health Plan’s pre-heat landing page. It was designed to help visitors decide whether to shop for a new Medicare plan this year or stay with the one they had. If shopping made sense, it gave them a plan of action for getting the most out of AEP.
From Chatbot to Conversation Agent
What we built for Geisinger is what we call a conversation agent, and it works differently from a chatbot. Chatbots are often transactional. They live at the FAQ layer, handling defined questions, routing to defined answers and capturing defined data. They’re built to handle high volume at low complexity.
A conversation agent operates at a higher level. It carries context across a conversation, exercises a form of judgment within guardrails set by our development team and adapts to what each visitor actually needs — not what a decision tree assumes they need. It’s the difference between a simple search and an interaction. That difference matters most in categories where visitors arrive with real stakes and no simple answers. Medicare shopping is one of them.
How We Built It
Geisinger’s conversation agent didn’t pitch Geisinger Gold. It didn’t lead the visitor toward enrollment. It didn’t compare Geisinger to competitors even when the visitor invites it to. Instead, it helps a visitor evaluate their current coverage and identify factors to look for if they intend to shop.
The tool architecture was achieved through three layers of system design; each built in partnership with the Geisinger team:
System instructions: the operating rules that govern the tool’s voice, scope, guardrails and refusal behavior. These are the rules the tool follows on every turn, without exception.
Training materials: the curated content library the tool draws from when answering. Nothing outside this library gets referenced. There’s no open-web browsing, no invented answers, and no drift into topics the tool wasn’t built to handle.
A decision framework: the discovery flow that guides the conversation from opening questions through personalized recommendations. This is where the tool’s consultative posture is operationalized, and where the discipline of not sellingactually lives.
None of those three layers are remarkable on their own. What matters is how they work together, and how carefully each must be authored to hold the strategic posture the tool was built to hold.
The three layers don’t hold themselves in place. Ongoing monitoring does. We review the tool’s conversations regularly, watching for opportunities to sharpen how it responds — a place where the tone could land warmer, a question where the flow could feel more natural and a topic where the framing could land better with the audience. When we see one, we update the system instructions, training materials or decision framework in real time. That’s what keeps the tool performing as planned over the life of the campaign, not just at launch.
There’s also a compliance envelope around these tools that reaches beyond CMS marketing rules. As conversation agents move deeper into regulated categories, plans and their agencies are navigating an evolving set of questions around consent, tracking and third-party data processing — including how state privacy laws apply when a website’s chat is powered by a vendor’s AI. Executing this kind of tool means keeping a live line of sight to that frontier as it evolves.
Earning Trust One Implementation at a Time
Older adults are increasingly willing to use AI. Our 2026 Consumer In Sights research found AI adoption among adults age 63 to 75 more than doubled in a single year, and digital confidence, not age, is the strongest predictor of who’s using it.
But adoption at the category level doesn’t automatically transfer to trust in a specific tool. Every implementation has to earn its own. The tools in this space will keep getting better. The constraint isn’t the technology — it’s the discipline of designing an interaction the audience will actually trust in a category they approach with real caution. Thoughtful design and ongoing human stewardship are what make that trust possible.
What Comes Next
Geisinger’s conversation agent came down on September 30, at the close of the pre-heat window. Since then, we’ve completed a full analysis of how it performed — what worked, where it needed adjustment and what patterns emerged. That analysis is informing how we approach the next generation of these tools with our health plan clients.
