In our practice, we’re inundated with refill requests from pharmacies for prescriptions we wrote. To stay sane, my buddy Claude Code and I built a workflow that pulls key chart data (active meds, recent device data, last labs, whether a med can be uptitrated, etc.) and applies rule-based logic that mimics how a provider would handle refills. Easy enough.

But I really don’t enjoy refilling other physicians’ prescriptions—usually at a first encounter when a patient says they ran out of X, can’t reach their doctor, and asks me to refill it. It’s important to note that we operate outside an IDN, and our patients’ records are fragmented across practices/systems. In these scenarios, I stay in my lane and refill only meds I’d reasonably be expected to prescribe as part of our role/contract.

Still, problems come up in two common scenarios:

  1. The prior doctor actually discontinued the med (lab issues, new med-med interactions, …), but the patient didn’t realize it (and pharmacies, of course, have no clue).
  2. The patient shouldn’t have been on it in the first place (side effects or contraindications their provider was presumably unaware of).

In my experience, this happens often enough (maybe ~5% of cases). And unfortunately, if you refill it, you own it. “I didn’t prescribe it in the first place,” or “The patient told me they were on it,” doesn’t feel like a defensible position.

So my approach to refilling other providers’ prescriptions is:

  1. Given what I’d reasonably have access to (HIE, recent labs) + what the patient reports: should they be on this medication?
  2. If yes, should I refill it?

This is why I’m watching the recent buzz about Utah allowing a non–FDA-cleared AI agent (Doctronic) to run a direct-to-consumer refill workflow. After trying the interface, my sense (I could be wrong) is the default assumption is: the med was appropriately prescribed, the patient simply ran out, and the main check is whether they report new contraindications.

That will work for most cases, and we shouldn’t let edge cases derail access. But the unhappy path is one I see fairly often, and it’s hard to address when data is missing/fragmented. This isn’t really AI vs human—it’s a missing data problem. But the AI will only be as good as what it can see (and how it’s trained).