Building an Editable Medication-Review Draft
A draft that can't be corrected easily isn't really a draft — it's a bottleneck wearing a draft's clothing.
Article
Starting from the edit, not the generation
It's tempting, when building an AI drafting feature, to focus almost entirely on generation quality and treat editing as an afterthought — a text box bolted on so users have somewhere to fix mistakes. We approached the medication-review draft the opposite way: we designed the editing experience first, and treated generation as the thing that populates it, not the thing the interface revolves around.
That ordering matters because a pharmacist's real workflow is fundamentally an editing workflow. They read a draft, they compare it against their own clinical read of the case, and they adjust language, add missing considerations, or remove content that doesn't fit. If the editing surface is clumsy, the drafting feature fails regardless of how good the underlying generation is.
Keeping provenance visible through edits
One of the harder design problems is what happens to citations and structure once a pharmacist starts editing generated text. We didn't want editing to quietly erase the traceability that the explainability layer works hard to establish. So the draft retains a record of which sections originated as generated content and which have been modified, without making that distinction intrusive to the everyday editing experience.
This matters for professional documentation. A review record that shows what the system proposed and what the pharmacist changed is more useful — and more honest — than one that flattens both into an undifferentiated final paragraph. It also supports later review of the tool itself: seeing where pharmacists consistently edit a certain type of statement is valuable product signal.
Structured fields alongside free text
Not every part of the draft is a paragraph. The review record includes structured elements — flagged items, source citations, intake summaries — alongside free-text sections, and each needs its own approach to editing. Structured fields get constrained, form-like editing controls that prevent accidental malformation, while narrative sections get a more conventional rich-text experience. Blending these into a single editing model, rather than forcing one paradigm onto both, was one of the more involved workflow decisions in building this feature.
We also made sure structured fields and narrative text stay visibly linked, so an edit to a flagged item is reflected sensibly in any narrative sentence that referenced it, rather than leaving the two out of sync.
Undo, versions and confidence to edit freely
Pharmacists need to feel free to substantially rewrite a draft without worrying about losing the original or making an irreversible mistake. The draft experience includes straightforward undo behaviour and a lightweight version awareness, so editing feels safe rather than high-stakes. A drafting tool that makes people cautious about editing defeats its own purpose.
None of this changes what happens at approval: the final, approved content is still the version a pharmacist has reviewed and signed off on, edited or not. The editing tooling exists to make that review step productive, not to add ceremony around it.
Insights articles explain how this product and its demonstration are designed. They are not clinical guidance and do not represent guideline recommendations.

