I just finished a Non-Emergency Medical Transport demo on Pega Customer Service for Healthcare, and the build turned out to be as interesting as the result. I did it working side by side with Claude in Cowork mode, and it changed how fast I could move.
We split the work in a way that made sense. I handled the Pega side — Blueprint, data transforms, views, the integration wiring. Claude handled the thinking around it: what the data model should look like, the exact values to use, what was likely to break, and whether things were actually right. When I got stuck on an error, I pasted the screen, it read the message, and it toldme the next step.
My one rule the whole way through was that the demo couldn’t be smoke and mirrors. If I was going to say it on stage, it had to be true. It is:
- The provider list is a live call to the public CMS NPPES registry. Real providers, nothing seeded.
- Coverage, the provider taxonomy, and the candidate list are all worked out at runtime.
- The recommendation actually reads the member’s situation — uses a walker, needs door-to-door help, how far the trip is against the coverage cap — figures out the right kind of transport, and lands on a provider that really is in the member’s ZIP.
- The advocate can override that pick from a dropdown of every provider and their NPI, pre-set to the AI’s choice.
The debugging was the fun part. At one point the recommendation kept looking random, and Claude traced it to a data transform bug that was only ever handing the AI the last provider in the list. Another time an out-of-state provider snuck in because the registry query wasn’t pinned to the practice location. Small fixes, but they’re the difference between a demo that holds up when someone pokes at it and one that falls apart.
What I liked was that it didn’t just agree with me. It checked provider records against the live registry, told me where the story would sound thin if someone pushed on it, and owned it when it got a Constellation detail wrong. I made the calls; it kept me honest.
One more thing, and it matters: none of this is really about Claude specifically. The useful part was the way of working — a partner that sweats the details and checks the facts while you build. You’d get the same thing pairing with Pega’s own AI assistant.