AI in Clinical Documentation: What It Is Good For, and Where It Should Not Be Trusted

Why this post is cautious

We sell an AI add-on. That is exactly why this is written the way it is: the fastest way to make clinicians distrust a tool is to oversell it, and the fastest way to make it dangerous is for someone to believe it more than they should.

Where AI assistance genuinely helps

These have something in common — the clinician remains the source of truth and the AI is reorganising or drafting.

• Turning your own notes into structured text. You know what happened; the work is typing it up.
• Summarising a long record you are about to read anyway. Useful as an index, not as a replacement for reading.
• Drafting patient-facing explanations. Turning clinical shorthand into something a patient can understand, for you to check and correct.
• Finding things in a record. "When was this medication started" across years of notes.

Where it should not be trusted

• As a source of clinical fact. These systems produce fluent, confident text whether or not the underlying claim is right. Fluency is not accuracy, and the failure mode is a plausible sentence rather than an obvious error.
• For diagnosis. Suggestions can be a prompt for your own thinking. They are not a conclusion, and they do not carry your clinical responsibility.
• For drug doses, interactions or anything where being nearly right is dangerous. Check these against a proper reference, every time.
• On anything that ends up in the record unread. The single worst pattern is AI-generated text saved without a clinician reading it. If a note is worth keeping it is worth reading first.

Two specific traps

The confident summary. A summary of a record you have not read is not a shortcut — it is you accepting a claim you cannot check. Summaries are most useful for records you will read, where they tell you what to look at first.

Silent authorship. If AI-generated content is stored without being marked as such, then six months later nobody can tell which parts of a record a clinician asserted and which a model drafted. That distinction matters clinically and it matters if the record is ever examined. In Roshtah, AI-generated content is marked as suggested at the point it is stored, on the server rather than in the interface, so the label cannot be lost by a client that forgets to show it.

Questions to ask before enabling any AI feature

• Is generated content clearly labelled in the stored record, or only on screen?
• What is sent to the model, and does it include identifiable patient data?
• Where is that processed, and is that acceptable under the rules that apply to you? This is a question for your own compliance advice, not for a vendor's assurance.
• Can I turn it off per user? Not everyone in a practice should necessarily have it.
• What happens when it is wrong? If the answer is "the clinician catches it", check the workflow actually gives them a moment to.

The short version

Treat it as a fast, tireless assistant that writes well and is sometimes wrong, whose work you sign. That framing gets you the benefit without the risk, and it is the only framing under which the clinical responsibility — which is still entirely yours — stays where it belongs.