Policy Engine in Medical Record Summarization, which turns a policy document into a decision model that evaluates clinical records against prescriptive rules.

A policy turns a written document - a standard operating procedure, a scoring methodology, or a list of rules - into a decision model that Medical Record Summarization can apply to a clinical record. Where a Summary section asks the AI to narrate what a record contains, a policy asks it to assess the record against prescriptive criteria and return a defined outcome.

Common policy documents include payer coverage policies, hospital or health system protocols, and medical coding guidelines. Any document that carries rules or criteria that must be evaluated consistently and accurately can serve as a policy document.

A policy has three parts, and each depends on the one before it.

When you create a policy, the AI reads the policy document and builds the decision model from it. That run takes a few minutes. You can then review the generated rule tree and refine it before it is used in production.

A decision profile pins down what question is being asked and what answers are permitted. Because a profile is reusable, any number of policies covering the same determination can share one. The profile also determines what the reviewer is presented with: a discrete, predefined outcome that the AI selects by applying the decision model's rules and scoring logic.

Profiles come in two forms:

For a direct decision profile, each outcome is labeled and assigned a polarity - Positive, Negative, or Neutral - which controls only how the outcome is styled, not how it is chosen. In the reviewer experience, Positive outcomes appear green, Neutral outcomes grey, and Negative outcomes red. For a scored decision profile, each outcome is a score band with a minimum and a maximum. Bands are matched in order from top to bottom and their bounds are inclusive.

A decision model is a tree. Rule groups nest inside other rule groups and bottom out in conditions, each of which is a single check against a field extracted from the record.

Every rule group declares how its children are judged:

In a First match group, the order of the items decides which one is applied, so set that order deliberately. For guidance on picking a qualifier, structuring requirements, and keeping outcomes mutually exclusive, see Refining decision model rules.

Conditions are backed by a catalog of extracted fields. Each field has a data type and a description of how the AI locates it in the record, and records which scoring criteria use it. Avoid changing an extracted field to text that the policy document does not contain.

Saving an edited decision model creates a new numbered version. Versions are sequential and permanent, and cases that have already been decided keep the version they ran against - regenerating a policy affects new cases only.

Each version carries a status, which is also aggregated on the Policies page:

A policy becomes part of a summary through a template. In the template editor, a section set to the Policy output type is bound to one published policy. When a record is processed, the AI evaluates that policy's rules against the record and the section renders the outcome, the score for scored profiles, a confidence rating, a written rationale, and a breakdown of which rules were met.

Reviewers can then accept or reject the outcome independently of the overall review, which feeds the policy's own feedback loop.