How Health Systems Should Be Thinking About Picking an AI Platform Right Now
Between October 2025 and March 2026, three major AI platforms launched specifically for healthcare. Claude for Healthcare from Anthropic, ChatGPT for Healthcare from OpenAI and Amazon Connect Health from AWS. All purpose-built, HIPAA-aligned and targeting the same health systems.
The instinct when three platforms launch in 90 days is to run a comparison. Which one is best? That’s the wrong question and I say that as someone who just finished evaluating all three for clients.
The more useful question is: what does your environment need to look like before any of them can move from pilot to production?
Because in my experience, that’s where most healthcare AI deployments actually fail, at the point where the AI meets your real EHR configuration, your actual compliance posture, and your live clinical workflows.
Platform selection is about 20% of the challenge
I’ve started saying this in every healthcare AI conversation I have, because it’s the thing that consistently surprises people: the platform you choose determines maybe 20% of whether the deployment succeeds.
The other 80% is three things.
Get Arunkumar Ganesan’s stories in your inbox
Join Medium for free to get updates from this writer.
EHR integration — Two of the three platforms Claude and ChatGPT for Healthcare rely on FHIR API connectors to access your EHR. Amazon Connect Health has direct EHR coupling. For health systems where FHIR adoption is partial, or where legacy EHR configurations are non-standard, the API-based approach requires integration engineering work before any AI workflow goes live. That work is not included in the platform contract. I’ve seen it take anywhere from a few weeks to most of a quarter, depending on the environment.
Compliance operationalisation — A signed BAA gets you past the procurement conversation. It doesn’t get you to production. PHI boundary enforcement, audit log architecture, data residency configuration, role-based access controls, these are architectural decisions the platform vendor doesn’t make on your behalf. Every organisation I’ve seen treat BAA execution as the finish line has encountered gaps at their first audit. The compliance work starts after the contract is signed.
Workflow instrumentation — This is the one that kills the most pilots. Prior authorization automation, ambient scribing, clinical documentation tools, all require calibration against your actual clinical workflows, not the generic templates in the vendor’s implementation guide. The distance between a vendor demo environment and a live clinical environment is significant. AI connected to an uninstrumented workflow doesn’t fail dramatically. It produces outputs nobody trusts, adoption stalls, and six months later the conclusion is that “AI isn’t ready for this.”
What to do before you pick a platform
Get your governance infrastructure in place first. Before platform selection, before vendor conversations, get the cross-functional group together IT, Legal, Clinical Informatics, Finance and align on approved use cases, data handling standards, and BAA requirements. Platform selection is a much faster conversation once the governance boundaries are defined.
Then match one platform to one workflow and run a bounded pilot with defined success metrics without a broad proof-of-concept. One use case, one timeline, one set of numbers you’re trying to move.
And treat integration as its own workstream with its own engineering resourcing from Day 1. The platform contract determines the ceiling. The integration work determines whether you reach it.
The full comparison
For the actual platform breakdown what Claude, ChatGPT for Healthcare, and Amazon Connect Health are each optimised for, which workflows map to which platform, and what the readiness checklist looks like across EHR integration, compliance architecture, and workflow fit, I put that together over at ideas2it.com.
That piece covers the detail this one doesn’t, and it’s worth reading before you sit down with any of the three vendors.
