Insights

AI Governance

A doctor's homemade AI scribe is a prototype, and the gap to a governed tool is specific

A physician who gets a working AI scribe out of a general chatbot has built a prototype, not a clinical workflow. A side-by-side checklist shows what a governance review will ask for and turns the doctor's initiative into a requirements list.

November 19, 2026 · 5 min read

A doctor's homemade AI scribe is a prototype, and the gap to a governed tool is specific

A prompt that turns a visit recording into a decent note proves the idea works. It does not prove the workflow is safe. Hubi Zieliński put it well in a post on X on November 11: "If you're a doctor building your own AI scribe with ChatGPT, I respect the initiative. But a working prompt is a prototype." He goes on to name what is missing: "Patient-data controls, error checks, permissions and EHR integration are what make it a usable clinical workflow."

I want to take that seriously in both directions. The doctor who builds a scribe has done something useful. They found a real pain point, they tested a fix on their own time, and they now know more about what a good note looks like than most product teams. That is valuable. It is also not something a health system can approve as-is, and the reasons are specific enough to write down.

Why the prototype feels finished

A homemade scribe works on the cases its builder tried. The note reads well. It saves time. The builder trusts it because they wrote it and they check every output.

That last part is the hidden control. The prototype is safe in one person's hands because that person reviews everything, knows its quirks, and knows which visits not to use it for. None of that transfers when a colleague copies the prompt. The safety lived in the builder's habits, not in the tool.

A governance review exists to move those habits into the system, so the tool stays safe when the person who built it is on vacation.

The side-by-side checklist

Here is how I would lay out the gap. The left column is what a typical homemade scribe has. The right column is what a governance committee will ask for before it lets the tool near patient care.

| Area | Prototype | What a governance review asks for | | --- | --- | --- | | Where the audio and text go | A personal or general-purpose chatbot account | A named tool under an agreement that covers patient data, with known retention and storage location | | Who can use it | Whoever has the prompt | Named users with accounts the organization controls and can turn off | | Patient consent and notice | Up to the individual clinician | A consistent way to tell patients and record that they were told, per the organization's policy | | Error checks | The builder reads every note | A defined review step before signing, plus a way to report and track errors | | Known failure modes | Discovered one at a time | A written list (missed negations, wrong medication names, invented findings) and a test set that checks for them | | Getting the note into the chart | Copy and paste | An integration or a controlled transfer that records where the text came from | | Changes over time | The model or prompt changes without notice | Version tracking and re-testing when the model, prompt, or vendor changes | | Who owns it | The builder | A named clinical owner and a named technical owner | | When to stop | When the builder notices a problem | Criteria for pausing use and a person with authority to do it |

None of those rows says the prototype was wrong. Each one says something that worked for one careful person now has to work for everyone.

The rows that come up first

If I had to pick the rows a committee will raise in the first meeting, they are the first two: where the data goes and who can use it. A general chatbot account is the part of the prototype most likely to stop the conversation, because the committee cannot answer basic questions about storage, retention, and who else can see the content. Ask the vendor and your privacy officer, in writing, before anyone records another visit.

Error checks come next. A scribe that is right most of the time is still producing notes that a clinician signs. The question is not whether errors happen. It is whether the workflow catches them before signature and whether anyone tracks the ones that slip through.

How to turn initiative into requirements

The best outcome is not to shut the doctor down. It is to use what they learned.

Ask the builder for three things. The prompt itself, because it encodes what a good note looks like for their specialty. A short list of the cases where it failed, because those become test cases. And their honest view of where they would not use it, because that becomes the scope.

Then hand that package to whoever evaluates commercial scribe tools, or to the team building one. The doctor's prototype becomes the acceptance criteria. If a vendor's product cannot match the homemade version on the builder's own test cases, that is worth knowing before you sign.

What I would tell the doctor

Keep going, but stop using real patient data in a personal account until someone with authority has said yes. Write down what you learned while it is fresh. Bring it to your informatics or compliance team as a proposal, not a confession.

And for the committee: treat the homemade scribe as a signal, not a violation. A clinician who built one is telling you the approved tools are not meeting a need. The table above is the fastest way to find out what it would take to meet it safely.