Clinical documentation improvement software: where AI helps and people decide

Clinical documentation improvement, usually called CDI, is the work of making sure the health record says clearly what the clinical team found and did. CDI specialists review charts during or after a stay, look for documentation that is missing, vague or in conflict, and ask the provider to clarify it. The clarification request is called a query.

CDI software is meant to support that work in four ways: it pulls the right charts into a worklist, highlights passages that may need clarification, helps a specialist write and send a query, and tracks what was asked and answered.

The newer tools add AI. Language models can read a long record faster than a person and point to the sentence that matters. That speeds up review. The specialist is still accountable for the query and for what it says.

It applies to machines as well as people. The brief says "A query may be developed by a healthcare professional or through a computer autogenerated query process." It also says that whatever a tool calls it, a prompt, a nudge or an alert, "if it meets the above definition of a query, it is considered a query." An AI suggestion that asks a provider to document something is held to the same standard as one a specialist wrote.

It must not lead. The query statement should be non-leading, "clear, concise, and specific", and titles that providers see "should be non-leading in nature and not include impactful information (e.g., reimbursement, quality indicators, specific diagnoses ...)". A model rewarded for higher documented severity could drift into the leading wording the brief rules out, which is why a person reviews each query before it goes out.

It must be grounded in the record. Diagnosis options that are not already documented "must be supported by clinical indicators sourced from the medical record". For AI, that means every suggested option should cite where in the chart its support came from, so the specialist can check it.

Coding rules point the same way. Section I.A.19 of the ICD-10-CM Official Guidelines for Coding and Reporting, FY 2027, says a code follows the provider's diagnostic statement, not the clinical criteria behind it, and ends: "If there is conflicting medical record documentation, query the provider." AI can find conflicts. The provider resolves them.

How queries are routed and approved inside a team is a separate design question; our article on agentic CDI query orchestration covers that.

Treat each AI component in the CDI workflow as a system with an owner. Write down what it reads, what it is allowed to suggest, who reviews its output and how it is monitored. Keep the record of every suggestion and every human decision, because that record is what compliance and auditors will ask for.

CDI works with protected health information, so the delivery has to be HIPAA-aligned: least-privilege access to records, logging of who saw what, and agreements with any vendor that handles the data. Our page on what HIPAA requires of AI covers the regulation itself, and our page on auditing an AI agent's decisions in a clinical setting covers the review record in more depth.