One of our 2025 deployments ran into a wall in week three. The CDSS we were standing up needed access to a specific note type the EHR didn’t expose through its public API. We had a signed BAA, executive sponsorship, and a willing clinical team. The vendor still said no.
This happens more than the industry admits. Here’s what we did, and what we wish we’d done.
What we did
We built a thin extraction layer that ran inside the EHR’s allowed extension surface. It pulled the structured fields we could get, then prompted the clinician at chart-close for the two unstructured fields we couldn’t. The agent ran on the merged record.
It worked. Acceptance hit 71% by week six. The chart-close prompt added an average of 14 seconds per encounter, which the clinical lead deemed acceptable in exchange for the documentation savings downstream.
What it cost
- Three engineering weeks we didn’t budget for.
- A maintenance burden that lives forever — every EHR upgrade is a regression test.
- A clinician-facing UI we had to design and own, where we’d planned to be entirely backstage.
What we’d do differently
We’d push the API request to the vendor before signing the engagement scope, in writing, with a stated deadline. If the answer is no, the engagement either accepts the workaround in scope (with the corresponding price) or doesn’t start.
The hidden cost of vendor friction is the part of healthcare AI nobody plans for. It’s also the part that determines whether your roadmap survives contact with the calendar.