Hybrid AI for hospitals: private where it matters, managed where it helps

Not every task an AI performs in a hospital carries the same risk. Summarising a patient’s referral letter involves special-category health data. Tracking a newly published guideline does not. Treating both the same way either exposes patient data unnecessarily or denies clinicians useful capability.

Axis One: Private or Managed

In a hybrid architecture, every request passes a policy check before any model sees it. The question is simple: does this involve patient data?

• If yes, or if there is any doubt, the request goes to private models running in the hospital’s own cloud account, with outbound traffic blocked.
• If no, for example tracking new publications or comparing a revised guideline with the previous version, the task may use managed AI: a leading commercial model such as Claude, GPT or Grok, called through a managed service, preferably via the hospital’s own cloud provider, and billed per token.

The hospital sets the policy. The platform enforces it and records every decision, so it can be audited later. Because managed AI is billed per token, keeping it to this small share of non-clinical work also keeps its cost small and predictable.

When in doubt, a request stays private. That single default does more for patient safety than any contract clause.

Axis Two: Small or Large

Within the private lane there is a second choice. Most clinical requests, such as extracting lab values or answering a protocol question with a citation, are handled well by small language models. Running two small models in parallel, one general and one medical, lets them check each other. Only when they disagree, report low confidence or face a high-risk task does the request escalate to a larger model.

This keeps answers fast, keeps expensive capacity free for the cases that need it, and adds a built-in second opinion.

What Hybrid Is Not

Hybrid does not mean sending “less sensitive” patient data to an external service. It does not mean letting clinicians choose freely per question. And it does not mean a managed model quietly receiving context that includes a patient record. The boundary is defined by data, enforced by the architecture, and visible to your security team.

The Result

Sovereignty where it matters, frontier capability where it helps, and the right-sized model for each task. For hospital leaders, that is the difference between an AI strategy built on compromise and one built on design.