Editing AI Notes: How the Provenance Colour System Works

When a note is drafted from a conversation, the obvious question is: which parts can I trust at a glance, and which need a closer look? Reviewing a flat block of text where everything looks equally certain is slow and, worse, encourages skimming. A provenance colour system answers that question visually, so your attention goes where it is actually needed.

The idea

Rather than presenting a uniform wall of text, the note shows where each part came from. At a glance you can tell what was captured directly from the conversation, what the system inferred or structured, and what you have edited yourself. The note carries its own audit trail, visible while you review, so nothing important hides in plain sight.

Three states

A simple three-state system distinguishes three kinds of content: content grounded directly in what was said, content the system has inferred or structured from the conversation, and content you have changed during review. Each state is shown clearly and distinctly. The point is not decoration; it is that these three kinds of content carry different levels of certainty and therefore deserve different levels of scrutiny.

Why this matters for trust

The central risk with any AI drafting is that an inferred detail looks identical to a directly captured one, so you cannot tell where to focus. If everything reads with the same confidence, you either re-read the whole note with equal suspicion, which is slow, or you skim and risk missing the one part that needed checking. Surfacing provenance breaks that trade-off. It puts your review exactly where it belongs: on the parts that warrant a second look, not on every word equally.

Faster, and safer

This is what makes reviewing a draft both quick and careful at the same time, which usually feel like opposites. You are not re-reading the whole note with uniform suspicion; you are checking the parts the colours tell you to check. The clinician stays fully in control, with the information needed to exercise that control efficiently, and signs off a note they have genuinely reviewed rather than one they have skimmed and hoped about.

A worked example

A draft note contains a history taken almost verbatim from the conversation, a structured treatment section the system has organised from what was said, and a couple of small corrections the clinician makes on review. With provenance colours, the clinician can see at once that the history is directly grounded and needs only a confirming read, that the structured section is an inference worth a careful check, and that her own edits are marked as hers. Her review is focused and fast, and she signs off knowing exactly what she relied on and what she verified.

Frequently asked questions

What does the provenance colour system actually show?

Where each part of the note came from: directly captured from the conversation, inferred or structured by the system, or edited by you.

Why are there three states rather than just AI versus human?

Because directly captured content and inferred content carry different certainty and deserve different scrutiny, so collapsing them would lose the most useful distinction.

How does this speed up review?

It directs your attention to the parts that need checking, rather than forcing you to read the whole note with equal suspicion.

Does the clinician still control the note?

Entirely. The system surfaces provenance to support review; the clinician edits and signs off.