Governing AI Summaries of Outside Records Before They Enter the Chart
Treat the gap between summary generation and chart entry as an enforcement problem. Write policy for what the summarization step may access and produce. Enforce that policy at defined control points (input handling, redaction, output checks, human approval), each able to allow, redact, block or escalate. Retain evidence of every decision. Runtime enforcement covers individual interactions; whether a summary belongs in the legal record is a clinical and records-management decision.
A staged control model
- 1
1. Input handling
Policy defines which intake channels and document types may feed the summarization step, and what content the step may read. Typical outcomes are allow or block. Documents that cannot be classified can escalate.
- 2
2. Sensitive data redaction
Policy names the categories to mask or remove before the model receives content, and the categories that cannot be safely redacted and therefore trigger a block. The outcome is redact, allow or block.
- 3
3. Output policy check
Policy defines permitted output content, for example that each statement points to a source document and that restriction labels from the source are carried forward. These checks can enforce structure and scope. They cannot establish clinical accuracy on their own.
- 4
4. Human approval
Policy states when approval is mandatory, who is authorized to approve, and what the approver sees. The approver's decision releases or withholds the summary, and the decision is recorded.
- 5
5. Chart-entry release
The write to the chart is itself an action under policy. Only summaries that policy allowed or an approver accepted proceed, and the outcome is recorded.
Evidence each summary should produce
- Policy version applied to the interaction
- Source documents processed
- Redaction and block decisions, with the rule that triggered each
- Output released or withheld
- Approver identity and time of decision
- Chart-entry outcome
Why the gap before chart entry needs enforcement
An AI summary of an outside record, such as a faxed document, a referral packet or a record received from another organization, is new content built from material the organization did not author. The concern is that the summary may omit, distort or add information, or may expose sensitive data. If it enters the chart without defined controls, the error can become part of the record of care.
Review after the fact addresses this late. Once a summary has been accepted, the organization often cannot show what the AI was permitted to do, what source content it processed, which summaries were released, or who approved them. Placing controls between generation and chart entry changes the question from whether someone caught a problem to what policy allowed through.
The HIPAA Security Rule offers a useful frame. Regulated entities must ensure the confidentiality, integrity and availability of all ePHI they create, receive, maintain or transmit, and must protect against reasonably anticipated impermissible uses or disclosures. Outside records are ePHI an organization receives, and summaries are content it creates from them, so the integrity and confidentiality duties are a natural starting point for both.
Write policy before configuring controls
Start with an inventory of intake channels for outside records, including fax, referral packets and inter-organization exchange, and decide which may feed an AI summarization step at all. Controls configured before this decision tend to encode assumptions nobody has approved.
Then write policy for each summarization step: permitted inputs, permitted output content, redaction categories, and the conditions that trigger a block or an escalation. Access scope is where minimum necessary thinking applies. The 2013 HIPAA final rule applies the minimum necessary standard directly to business associates using, disclosing or requesting PHI, and the Security Rule separately requires protection against impermissible uses. A summarization step that reads a whole packet when it needs one section is a scope decision that policy should make explicitly.
Restriction labels deserve a deliberate decision. The ONC § 170.315(b)(7) criterion text describes summary records tagged as restricted and subject to re-disclosure restrictions at the document, section and data-element level. It is an existing pattern for labeling at that granularity. Whether an AI summary must preserve such labels is an open design question for each organization. Decide whether the output check blocks, flags or escalates when a summary draws on labeled content.
Finally, define who approves what. Mandatory approval conditions, authorized approvers and accountability for reviewing escalated items should be named roles, not assumptions. Plan periodic review of block and redaction rates against policy so that policy changes are deliberate.
What runtime enforcement covers, and what it does not
Runtime enforcement can decide, at the moment of each interaction or tool call, whether content is allowed, redacted, blocked or escalated, and can produce evidence of that decision. Trussed enforces policy in the execution path across models, agents and tools, with allow, redact, block and escalate-for-human-approval outcomes, and generates audit-ready evidence from that enforcement. Policy is enforced at runtime, not reviewed after the fact. Its scope is individual interactions and tool calls. It does not provide workflow-level or transaction-level governance, and it does not describe model or dataset lineage.
Several decisions sit outside any enforcement layer. Clinical, records-management and compliance owners must decide whether an AI summary becomes part of the legal record. They must set the conditions for mandatory human approval and assign people to work the escalation queue. They must also track regulatory change, including the ASTP/ONC proposal and the HIPAA Security Rule rulemaking, before citing either as a current obligation.
A practical test for any implementation: for a given summary in the chart, can you show what the step was permitted to access, which source content it processed, what was redacted or blocked, and who approved release? If any answer depends on someone's memory, the control point is missing.
Control points between outside record and chart
Input handling
Decide which records may reach the summarization step.
Sensitive data redaction
Remove defined categories before the model sees content.
Output policy check
Test the generated summary against permitted content rules.
Human approval
Escalate defined cases to an authorized approver.
Chart entry
Release only what policy allowed or an approver accepted.
Define your control points before summaries reach the chart
Review how policy enforcement in the execution path produces audit-ready evidence for each AI interaction.
See How Evidence Is ProducedHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.