
A proposal with its evidence is useful. An assigned code is not the assistant's to make
Coding from clinical documentation is skilled work, and a lot of it is repetition. Identical clinical situations result in the same codes. An experienced coder spends a great deal of time on cases that are not ambiguous.
The ambiguous cases are the ones that need judgment: the documentation supports more than one code, or the documentation is so thin that it does not support the code in question.
The more important of those cases is the one an automated system is most likely to ignore. If the documentation does not support a code, that is a documentation problem. It is better to highlight the problem than to produce the code.
That is why this assistant offers suggestions but does not make decisions. A certified coder reviews each proposal. Any code assigned to a claim is associated with the person responsible for it. The page makes it clear that the output is only a draft for review.
How the assistant is built in Agent Studio
The coding team maintains prompt skills that include the coding guidelines and previous decisions. A date is assigned to the reference set so you can identify proposals based on a superseded reference.
The system prompt is the strictest in this library. When you propose a code, you must include the relevant passage as supporting evidence, state your confidence, and list other codes that the documentation could have supported. You must never treat an output as an assignment. The prompt states clearly that a certified coder determines the code.
It also requires an explicit result when the documentation does not support a code, and treats that outcome as normal rather than a failure.
The flow sends the extracted documentation to the Agent Chat node, scoped to the coding task.
Notion or your coding queue receives proposals that include passages, confidence levels, and alternatives. A Human in the Loop node sits between each proposal and any claim.
When documentation is poor, it is passed to the clinician as a documentation query rather than coded.
The proposals recorded in Google Sheets also serve as a quality measure and provide the evidence trail that an audit would require.
What this assistant is built from
- Prompt skills: your coding guidelines and prior decisions, with a dated reference set.
- System prompt: propose with a quoted passage, state confidence, list alternatives, never assign.
- System prompt: an explicit outcome for documentation that supports no code, treated as normal.
- Agent Chat (Util): receives extracted documentation, scoped to the coding task only.
- Human in the Loop: a certified coder reviews every proposal before any claim.
- Google Sheets (Integration): logs proposals against coder decisions as a quality measure and audit trail.
- When an ActionFlow is enough: If a routine encounter type already has an established code set, build the workflow. A coder still reviews the match.
Frequently asked questions
Start building AI workflows
Create a free account, open a template or a blank canvas, and run your first ActionFlow.
Newsletter
Get product updates
New nodes, agents, and product notes. We send mail only when we have something worth opening.