Texas SB 1188 AI-Generated Records Checklist: Engineering Evidence for Practitioner Review
As of November 2026, Texas SB 1188 (89th Legislature) adds Chapter 183 to the Health and Safety Code on electronic health record requirements. Bill text calls for practitioners to review all records created with artificial intelligence to ensure the data is accurate and properly managed. The text does not prescribe an evidence format, so teams should be able to produce six artifacts for any sampled record: the link to its AI interactions, the policy applied, the action taken, the review event with reviewer and time, the standard referenced, and durable, retrievable storage with known location and access controls.
The evidence checklist
What the bill text asks for, and what is testable
SB 1188 is scoped to electronic health records and health care practitioners, not records in general. Bill versions place a review requirement in new Chapter 183: practitioners review all records created with artificial intelligence to ensure that the data is accurate and properly managed. A Senate Research Center analysis frames the scope more narrowly, describing review where AI may be used for diagnostic purposes only or to recommend treatments, with review consistent with medical records standards developed by the Texas Medical Board. The bill also carries provisions on U.S. physical storage of electronic health records and on limiting access to people who need the information for treatment, payment or health care operations. Because versions differ, have counsel confirm which text applies to your organization and whether you are covered.
The word the text relies on is review, and it does not define what proof of review looks like. Everything below is an engineering inference: a set of artifacts that let you demonstrate review to an auditor or regulator. They are not statutory requirements, and producing them does not establish legal compliance with SB 1188. The bill also contemplates rules from Texas boards and the Department of Insurance, which may later shape evidence expectations.
Why evidence should come from enforcement, not reconstruction
Most teams start from application and model logs and try to assemble a story afterward. That approach tends to fail on the questions an auditor asks first: which policy was in force for this interaction, was the output changed before it reached the record, and did a person approve it before it was committed. Logs written for debugging rarely carry policy version, enforcement action or reviewer attribution in a consistent shape.
A more reliable design generates the evidence at the moment of each decision. When policy is enforced in the execution path, the same step that allows, redacts, blocks or escalates an interaction can write the decision record, including the policy applied and the outcome. When the action is escalation for human approval, the approval or rejection becomes another event on the same interaction, with the reviewer and time attached. The evidence exists because enforcement happened, so there is no separate reconstruction exercise and no gap between what was enforced and what was recorded.
This is a design choice, not something the bill specifies. It is useful to compare any approach, including tooling you already run, against the checklist above. Trussed is one implementation of this pattern: it enforces policy in the execution path at the level of individual interactions and tool calls, supports escalation for human approval, and generates audit-ready evidence as a by-product of enforcement.
Retention, residency, access and integrity
The evidence store is itself a sensitive system. AI interaction data and record content may contain protected health information, so the store needs the same attention as the records it describes. The bill's analyses describe a requirement that electronic health records be physically maintained in the United States or a U.S. territory, including records held by third-party computing facilities or cloud providers. Whether interaction logs count as electronic health records is a question for counsel, but a conservative engineering posture is to know and document where interaction data and evidence reside, and to be able to show it.
Set retention deliberately. The material available does not specify a retention period, so choose one with counsel and make sure it is configured rather than assumed, including how long approval events are kept relative to the records they support. Restrict access to evidence to people with a defined need, and record who reads it.
Integrity needs an owner. Decide who can alter or delete evidence, require that such actions are themselves recorded, and define how tampering would be detected. A decision record that an administrator can silently edit carries little weight in an inquiry. Self-managed deployment in your own cloud, with configurable retention and data residency, is one way to keep evidence under your own storage and access controls.
What runtime enforcement does not cover
Runtime enforcement can evidence the policy applied, the action taken and the approval events for interactions it sees. It does not show that a reviewer read the content carefully, that the clinical judgment was sound, or that the review met Texas Medical Board standards. Those remain matters for practitioners, supervisors and quality processes, and your evidence should reference the procedures that govern them rather than imply the system proves them.
Two further limits apply. Enforcement evidence covers only interactions that pass through the enforcement point, so any AI use outside it, such as an unsanctioned tool, produces nothing. And enforcement at the level of individual interactions and tool calls does not by itself give you transaction-level governance or end-to-end tracing of an agentic workflow. If your auditors expect that, plan for it separately.
Finally, do not describe any tooling as delivering compliance or certification with SB 1188. The accurate statement is narrower: certain controls produce evidence that supports a practitioner's ability to demonstrate review.
How to assess your current gaps
Run the assessment as a sample-based test rather than a document review. The goal is to find out whether the evidence can actually be produced, not whether a policy says it should be.
- Inventory AI-touched record workflows: List every workflow where AI creates or alters a patient-related record, and separate diagnostic or treatment-recommendation uses from other uses.
- Check each checklist item per workflow: Record whether interaction ID, record ID, policy version, action, reviewer and timestamp are captured, and where they are stored.
- Confirm approval is a system event: Verify that human approval is captured with identity and time and linked to the record, not held in email or a separate sign-off.
- Verify the evidence store: Document retention settings, storage location, access controls and tamper detection, and name an owner for each.
- Test retrieval end to end: Pick a sample record and produce its AI interactions, policy decisions and approval events on request, timing the exercise.
From review expectation to evidence artifact
Record linkage
Record ID tied to every AI interaction that created or altered it
Policy and action
Policy version applied and the result: allow, redact, block or escalate
Human review
Approval event with reviewer identity, timestamp and outcome
Evidence custody
Retention, storage location, access control and tamper detection
Test your evidence against the checklist
See how policy enforcement in the execution path produces audit-ready evidence for individual interactions and tool calls, and where your own review process still has to carry the load.
See How Evidence Is Produced