What Texas's AI Law Requires Before You Roll Out AI in Operations

Texas's first broad artificial-intelligence statute, the Texas Responsible Artificial Intelligence Governance Act (TRAIGA, House Bill 149), was signed in June 2025 and took effect on January 1, 2026. If your company does business in Texas and uses an AI tool, the law reaches you. But for the back-office AI most operators are actually rolling out—reading invoices, drafting purchase orders from emails, matching customer payments, scheduling—TRAIGA asks less than the headlines suggest.

The real work is short. Confirm your use is not one of a handful of prohibited purposes, add one disclosure if you deliver health care, and keep a dated, framework-based review of each AI workflow so you can raise the statute's own legal defense. This article walks a Texas operations or finance sponsor through those three steps, with the section numbers to check and a worked Houston example. It is a practical reading of the statute, not legal advice; confirm edge cases with Texas counsel.

Does the Texas AI law apply to the AI you're adopting?

Yes, almost certainly—but "applies" is not the same as "heavy duties." TRAIGA covers any person who develops, deploys, or offers an AI system in Texas, and it defines an "artificial intelligence system" as a machine-based system that infers from the data it receives how to generate outputs such as content, decisions, predictions, or recommendations (Sec. 551.001). A deployer is simply a person who puts such a system to use in the state (Sec. 552.001); you do not have to build the model to be covered. There is no small-business exemption, no registration, and no general impact-assessment requirement for private companies.

See how an AI assessment settles scope, ownership and boundaries before you deploy

So an invoice-reading tool or an order-intake assistant is an "AI system" under the law. The useful question is not whether the law applies but what it asks you to do with that system—and for ordinary operations, the answer is narrow.

What TRAIGA actually prohibits

The law's private-sector teeth are a short list of intent-based prohibitions, not a general duty of care. A person may not develop or deploy an AI system to manipulate a person into self-harm, harm to others, or crime (Sec. 552.052); with the sole intent to infringe constitutional rights (Sec. 552.055); with the intent to unlawfully discriminate against a protected class (Sec. 552.056); or to produce unlawful sexual deepfakes or child sexual abuse material (Sec. 552.057). Two further prohibitions—social scoring (Sec. 552.053) and capturing biometric identifiers without consent (Sec. 552.054)—apply to governmental entities, not private businesses.

Intent is the hinge. The discrimination ban states plainly that "a disparate impact is not sufficient by itself to demonstrate an intent to discriminate" (Sec. 552.056(c)). That is why a reconciliation model or a document-reading workflow deployed to speed up ordinary operations does not trip these provisions: no one is deploying an invoice reader to discriminate, manipulate, or produce illegal content. The table below maps common back-office uses to the duty, if any, that TRAIGA creates.

This table is Clairvance's reading of the statute to help you triage; it is a method, not a legal opinion, and unusual uses deserve a closer look.

The one disclosure that lands on health care

If you deliver health care, tell patients when AI is used in their care—no later than the date the service is provided, or as soon as reasonably possible in an emergency (Sec. 552.051(f)). This is the clearest private-sector disclosure duty in the law, and it lands squarely on a city like Houston, where the Texas Medical Center anchors a large base of medical, dental, and veterinary groups now adding AI to intake, scheduling, and documentation. The parallel rule for government agencies interacting with consumers (Sec. 552.051(b)) does not bind private businesses, and TRAIGA imposes no general "you're talking to a bot" disclosure on a private company's customer chatbot.

How the safe harbor actually works

TRAIGA gives you an affirmative defense, and earning it is the single most useful thing an operator can do. Under Sec. 552.105(e), a defendant is not liable if it discovers a violation through its own testing or internal review and was substantially complying with the most recent NIST AI Risk Management Framework: Generative AI Profile (NIST AI 600-1) or another recognized risk-management framework. The statute also starts you with a rebuttable presumption that you used reasonable care (Sec. 552.105(c)).

Be clear about what this is and is not. It is a defense raised after a claim, not immunity, and "substantial compliance" is not defined in the statute or certified by NIST. A single boilerplate memo will not carry it. What does carry it is a living record: an inventory of each AI workflow, its purpose, its data inputs and outputs, its known limits, the person who reviews results, and a dated review mapping the workflow to the framework's functions. Keep it current as the workflow changes, because the defense rests on what you were actually doing when a problem surfaced.

Who enforces it, and what it can cost

Only the Texas Attorney General can enforce TRAIGA, and there is no private right of action (Sec. 552.101). Before any penalty, the attorney general must give written notice and a 60-day period to cure the problem (Sec. 552.104). Civil penalties then run from $10,000 to $12,000 for a curable violation, $80,000 to $200,000 for one that cannot be cured, and $2,000 to $40,000 per day for a continuing violation (Sec. 552.105). One nuance operators miss: a state agency that licenses your business can add its own sanctions—including a monetary penalty up to $100,000 or action against your license—but only after the attorney general has found a violation and recommended that agency action (Sec. 552.106). For companies that want to test higher-risk AI, the Department of Information Resources runs a regulatory sandbox that allows supervised testing for up to 36 months (Chapter 553), though it cannot waive the prohibited-use rules.

A short review to run before you switch an AI workflow on

Here is the whole method applied to one case. Hypothetical example: a Houston multi-site clinic group is adding an AI assistant that books appointments and drafts visit summaries for a clinician to review. Walk the same three steps any operator can.

1. Screen for prohibited intent. Booking and drafting are not manipulation, discrimination, or any purpose listed in Sec. 552.052 through Sec. 552.057, so no prohibition applies.
2. Check disclosure. Because the tool touches patient care, the group adds a plain line to intake—"AI assists with scheduling and draft notes; your clinician reviews them"—delivered by the date of service (Sec. 552.051(f)).
3. Bank the defense. The group writes one page per workflow: the purpose, the data in and out, the clinician who signs off on summaries, the failure cases to test (a wrong patient match, an invented history), and how each maps to the NIST profile. It dates the review and schedules a refresh.

That is the entire obligation for this workflow. The example is illustrative; a group that adds biometric check-in, or AI that decides coverage or denials, would have more to document and should treat those as higher-risk uses.