I recently worked with a team on a POC for a customer who was interested in leveraging AI within their Open Negotiation case type for the ACA No Surprise Bill. Open negotiation requests most often come in from Providers via email with multiple attachments and must be acknowledged within 30 days of receipt. For those who know, you know the volume of these requests!
Our customer came to us with a goal of automating the intake (and much more within the case type, but this post is focusing on this initial step) of these email since the process is 100% manual requiring dual keying into two back end systems. To solve this, we instituted a two step process:
An email listener that monitors the email box
A Pega Agent (leveraging Claude Sonnet 4.6) that then reads the email(s) and associated attachments and creates cases with the content in the attachment. After much analysis, we derived each case to be a combination of a claim number and specialty code to be worked individually. IMHO all Agents within Healthcare should have HITL oversight and that is how we built this agent to enable Coordinators to have final approval.
We did have a couple of lessons learned that I will share:
Because the Provider name was of Indian origin, the LLM assumed the dollar to be in Rupees instead of dollar.
Be very succinct with instructions to the Agent and when mapping the values from the attachment directly into the case
e.g. Initial Payment Amount = {.InitialPaymentAmount}
The next step is to leverage Pega’s AI Assistant in the next iteration of our development of an Agent like this. Additionally, the ultimate benefit in this approach is the hundreds of hours our customer will now be saving every year!
I’m curious what types of healthcare Agents others have you created to help make Advocate, Examiners, Coordinator or Investigator lives easier?
This is a great example. Healthcare agents can provide realtime use in many processes for speeding up the Intake for care coordination as well as Agents for Admission and Discharge. Helping to reduce the manual keying of information from one system to another. Similar process just in the front end process of admission requests or assisting with discharge planning.
Our claim product team built something similar but further down the lifecycle — an agent that works pended claims by literally following the desktop procedure step by step: doing external lookups/validations as it goes, reporting what it found at each step, then deciding the next step based on that — same branching logic an examiner would follow. At the end it hands the examiner a recommended resolution plus the full trail of what it checked, so nothing gets submitted without a human sign-off.
We also learned a lesson — job aids that read fine to a human (with all the implicit context) can send an agent down the wrong branch if you don’t spell that context out explicitly.
@dayts1 This is a fantastic application and shared insight into using agents in the healthcare space. I appreciate the callouts to HITL and significant time savings as one of your outcomes. By solving for hundreds of hours from tedious dual-keying, you’ve now freed up time for coordinators to focus on the work where their experience and judgment are really needed. Using the agent to do the heavy lifting, while keeping the human in control to make the final call, is exactly how we should be using AI.
@Janet_Rappa this is a great example and one that can be applied across the healthcare payer spectrum. There is so much power in the traceability of the logic the agent followed, ensuring claims are paid properly! Thank you for sharing!