Keeping a clinician in the loop: how AI EHRs should handle decisions

AI in an EHR should prepare the work, not make the clinical decision. The principle is simple to state — the AI drafts, a clinician decides — but it's the single most important thing to get right, because it's the line between a tool that helps and a tool that quietly does harm. Here's what "human-in-the-loop" actually means and how to tell whether a system respects it.

The principle: AI prepares, a clinician decides

Every clinical action — a diagnosis, a prescription, a signed note — carries responsibility that belongs to a licensed human. Good AI in healthcare doesn't try to take that responsibility. It does the preparation: pulling the relevant context together, drafting the note, suggesting the codes, proposing the orders. Then it stops and hands the decision to the clinician.

That handoff isn't a limitation to be engineered away. It's the design.

Why it's non-negotiable

The reason is measurable. A 2025 commentary in npj Digital Medicine notes that evaluations of modern LLM-based ambient scribes report overall error rates of roughly 1–3%, while cautioning that reported rates vary significantly with evaluation methodology — some studies count only factual inaccuracies, others include omissions and clinical inconsistencies. (Higher figures of 7–11% are still quoted, but those describe speech-recognition dictation rather than modern scribes.) In most software, a low single-digit error rate is a bug to schedule. In clinical software, it's a reason a human must verify before anything is signed, because the cost of the wrong error reaching a patient is not measured in inconvenience.

There's a professional dimension too: the clinician who signs generally assumes responsibility for what they sign, and exercising independent clinical judgment is part of that. A system that blurs the line — that lets AI output flow into the record without a clear human decision — isn't being advanced. It's shifting risk onto the clinician without giving them control.

What good human-in-the-loop looks like

Keeping the human in the loop is more than a disclaimer at the bottom of the screen. In a well-designed system it shows up as concrete behavior:

1. Review is a visible step, not an assumption. The clinician sees what the AI produced and actively confirms it, rather than discovering after the fact that something was filed.
2. The AI is transparent about what it did. Generated content is labeled as generated, ideally with what produced it and when, so the clinician knows what they're reviewing.
3. Actions are previewable before they happen. The clinician can see what the AI is about to do — draft an order, prepare a claim — and adjust it before it's committed, not after.
4. The default leans conservative for high-stakes work. The more consequential the action, the more the system should ask before acting, rather than acting and asking forgiveness.

The common thread: the clinician is never surprised by what the AI did, because the AI's role is to propose and the clinician's role is to decide.

What bad human-in-the-loop looks like

The warning signs are just as concrete. Be wary of a system that:

• Treats review as a formality — a single "accept all" with no real visibility.
• Files AI-generated content into the record without a clear authorization step.
• Can't tell you what was AI-generated versus human-entered after the fact.
• Markets autonomy as a feature — "handles it for you" — for tasks that carry clinical weight.

Any of these means the loop is open where it should be closed.

Why this is a workflow question, not just a safety one

Here's what's easy to miss: done well, human-in-the-loop isn't only safer — it's better to use. A system that surfaces exactly what the clinician needs to review, clearly and quickly, respects their time and their judgment. A system that hides the AI's work, or forces a blind rubber-stamp, wastes both. The goal isn't to slow the clinician down with verification; it's to make the right verification fast, so the loop is a moment of control rather than a chore.

That's the difference between AI that treats the clinician as an obstacle to automate around and AI that treats the clinician as the expert it exists to support.

Where Zenthea fits

Zenthea is built around a straightforward rule: the AI prepares, and a clinician decides. Thea can draft documentation, gather context, and propose the next steps in a workflow, but a human clinician reviews and authorizes anything that counts as a clinical action. The intent is to make that review fast and clear — a genuine point of control, not a rubber stamp — because keeping the clinician in the loop is the whole point of doing this responsibly.