Implementation Guide

    How to Keep EHR Data Reaching AI Systems Inside the US Under Texas SB 1188

    Treat US-only handling of EHR data as a runtime enforcement problem. Map every path from the EHR to an AI component, pin models, vector stores and log sinks to US regions, restrict sub-processors, and enforce policy in the execution path so each request is allowed, redacted or blocked based on where it would go. Keep a per-interaction record of the decision. Contracts show intent, but only runtime enforcement and evidence show what happened. Legal and compliance should own the interpretation of the statute. This guide is not legal advice.

    Map the paths where EHR data can leave the US

    1. 1

      Prompts and context windows

      Clinical text sent to a model endpoint whose serving region is not fixed. Control: region-pinned endpoints and redaction of identifiers before the request leaves your boundary.

    2. 2

      Retrieval-augmented generation

      Embeddings, vector indexes and caches hold EHR-derived content at rest. Control: host indexes in US regions and confirm where any managed service replicates data.

    3. 3

      Tool calls and MCP servers

      Agents call tools that fetch or forward EHR content to hosts you do not run. Control: an allow-list of tools and destinations, with least-privilege access for agents.

    4. 4

      Third-party model and vendor APIs

      Sub-processors may route, process or retain data outside the US. Control: contractual location terms plus runtime checks on the actual destination.

    5. 5

      Coding assistants and AI sites

      Developers and staff paste records or extracts into tools outside the sanctioned path. Control: route these tools through the same enforcement point.

    6. 6

      Logs, telemetry and fine-tuning sets

      Prompts and outputs copied into observability stacks, support tickets or training data in other regions. Control: US-only sinks and defined retention.

    1. Complete the EHR-to-AI path inventory and tag every hop with region and vendor
    2. Block or re-route destinations whose region is unconfirmed
    3. Place in-path enforcement on model calls, tool calls and MCP servers first
    4. Keep AI-generated content out of the chart until practitioner review is complete
    5. Exclude credit score and voter registration data from AI-accessible EHR content

    Evidence to retain for each AI interaction

    • The destination region of each model, tool, MCP server and vendor endpoint contacted
    • The data classification of the content and whether it contained EHR-derived data
    • The policy decision applied (allow, redact or block) and the policy version that produced it
    • The identity or role of the user, agent or tool and its mapping to treatment, payment or health care operations
    • The outcome, including what was redacted or why a request was blocked
    • The region and retention settings of caches, indexes and log sinks that hold derived content

    What SB 1188 requires and what it leaves to your architecture

    Texas SB 1188 added Chapter 183, Electronic Health Records, to the Texas Health and Safety Code. As of November 2026, the chapter is enacted and codified, and it directs the named state agencies to enter into a memorandum of understanding and adopt rules as necessary. Two duties matter most for AI architecture.

    First, a covered entity must ensure that the EHR information of Texas residents, other than open data, is accessible only to individuals who need it for duties related to treatment, payment or health care operations. Second, the bill is described as requiring EHRs under a covered entity's control to be physically maintained in the United States or a US territory. That duty extends to records stored by a third-party or subcontracted computing facility or a cloud computing services entity, and to records held in technology through which patient information may be electronically retrieved, accessed or transmitted.

    The statute states outcomes. It does not prescribe a topology, a vendor pattern or a control. Everything below is a design choice that supports those outcomes. Ownership of the covered-entity determination and the reading of 'physically maintained' belongs with legal and compliance. Engineering's job is to make the data paths knowable and the boundary enforceable, so the design should treat any component that stores or processes EHR content as in scope for region controls.

    Two adjacent duties affect AI design as well. A practitioner must review records created with AI under Texas Medical Board standards, so AI-generated content should stay out of the chart until that review is complete. The bill also prohibits collecting, storing or sharing credit score or voter registration information in an individual's EHR, so keep those fields out of anything AI systems ingest from the record.

    Apply the layers in order

    No single layer is sufficient. Topology narrows where data can go, region pinning fixes where it is served, vendor controls limit who can touch it, and runtime enforcement catches what the first three miss. The order below puts the cheapest structural decisions first.

    Flag any endpoint, sub-processor or telemetry destination whose region you cannot confirm as US or a US territory, and block or re-route it until it is confirmed. An unknown region should be treated as a failed check, not a pass.

    • 1. Choose the deployment topology: Run AI components in your own US cloud environment where feasible, so models, indexes, caches and logs sit inside a boundary you control.
    • 2. Pin regions: Configure model endpoints, vector stores, queues and log sinks to specific US regions, and disable cross-region replication and failover to non-US regions.
    • 3. Constrain vendors and sub-processors: Require documented processing and storage locations, a list of sub-processors, and notice of changes. Contracts set the expectation but do not prove behavior.
    • 4. Restrict access by role: Tie agent and tool access to EHR data to roles that map to treatment, payment or health care operations, and document the mapping.
    • 5. Enforce at runtime: Place an enforcement point in the request path that allows, redacts or blocks based on destination and data classification.
    • 6. Retain evidence: Store the destination region, policy decision and outcome for each interaction with defined retention.

    Why runtime enforcement is the layer that closes the gap

    Topology and vendor terms are decisions made in advance. Production traffic changes underneath them. A model provider adds a fallback region, an agent is given a new tool, or a developer points a coding assistant at a different service. A procurement review does not see any of this, and a quarterly attestation sees it late.

    An enforcement point in the execution path evaluates each interaction before it proceeds. If the destination is an approved US endpoint and the data class is permitted, the request is allowed. If the data class is permitted only with identifiers removed, it is redacted first. If the destination is unknown or outside approved regions, it is blocked. The same logic should apply to model calls, tool calls and MCP servers, because EHR content moves through all three.

    There are tradeoffs. An in-path control adds a dependency on the request path, so it needs to meet your availability targets. Redaction can reduce the clinical usefulness of a prompt, so decide per use case whether to redact or block. Overly broad blocking drives staff to unsanctioned tools, which is why coding assistants and AI sites belong behind the same enforcement point.

    Trussed is one way to implement this layer. It enforces policy in the execution path across models, agents, tools, MCP servers, AI sites and coding assistants, with allow, redact, block, rate limit and escalation for human approval, applied to individual interactions and tool calls. It supports self-managed deployment in the customer cloud with configurable retention and data residency, and its enforcement generates audit-ready evidence mapped to HIPAA.

    From EHR to AI: where US-only handling has to hold

    Map

    Inventory every prompt, retrieval, tool, vendor and telemetry path that touches EHR content.

    Pin

    Fix model endpoints, indexes and log sinks to US regions.

    Enforce

    Allow, redact or block requests in the execution path.

    Evidence

    Retain the destination, decision and outcome for each interaction.

    Enforce US-only handling of EHR data in the execution path

    Talk through your EHR-to-AI data paths and see how policy enforcement and audit-ready evidence work together in a regulated enterprise.

    Talk to an Expert