Texas TRAIGA Nine Months In: What the Attorney General Has Actually Enforced

Nine months after the Texas Responsible Artificial Intelligence Governance Act took effect, the attorney general has announced zero enforcement actions under it. Every AI case Texas has made public ran on older consumer protection and biometric laws. If you sell AI software into Texas, the Texas AI law matters less for what it has punished than for what it lets the attorney general demand.

Direct answer. The Texas attorney general has not publicly announced any enforcement action under TRAIGA since it took effect on 1 January 2026. The law gives the attorney general exclusive enforcement power, a consumer complaint portal due by 1 September 2026, civil investigative demands, a 60-day cure period and penalties up to $200,000 per uncurable violation.

The Texas AI law enforcement record after nine months

The headline finding is an absence. We looked for every public trace of TRAIGA enforcement between 1 January and 10 October 2026. That meant law firm alerts, enforcement trackers, trade press and the attorney general's own releases as reported by others. We found no lawsuit, no assurance of voluntary compliance and no announced investigation that names HB 149.

The most authoritative statement comes from Sidley Austin's Texas chapter in the Chambers AI 2026 practice guide. It says that "to date, no enforcement actions have been brought under TRAIGA", as of its 21 May 2026 update. Later explainers from September and early October describe the process and report no cases either.

The portal is the one thing that moved

The statute ordered one concrete step with a date attached. Section 8 of HB 149 says the attorney general must post a consumer complaint mechanism online "not later than September 1, 2026". A Cloud Security Alliance research note dated 13 September 2026 reports the mechanism was activated on 1 September.

The detail matters less than it sounds. A Newtral review published on 14 August 2026 found that the attorney general's Consumer AI Rights page already carried a "File An AI Complaint Online" button. That button led to the office's general consumer complaint portal. So the AI complaint channel is a door into an existing intake system, not a new enforcement unit.

The statute defines a consumer as a Texas resident acting in an individual or household context, and excludes anyone acting in a commercial or employment context. The Cloud Security Alliance note draws the obvious consequence: businesses, employees and competitors are not the intended filers. For a B2B vendor, the complaint that starts an investigation will usually come from your customer's customer, not your customer.

What TRAIGA prohibits for private-sector AI vendors

Read the enrolled text before reading any summary, because the private-sector duties are narrower than most coverage implies. The final bill dropped the broad risk-assessment model of earlier drafts. What survived is a short list of intent-based prohibitions, plus disclosure duties that mostly bind government agencies and health care providers.

The definitions are wide. A developer is anyone whose AI system is "offered, sold, leased, given, or otherwise provided" in Texas. A deployer is anyone who deploys one for use in the state. If you sell software with any model inside it to Texas customers, assume you are at least one of the two.

Summarised from the enrolled text of HB 149, Business and Commerce Code chapters 551 and 552. Read the statute before relying on any row.

Intent is the load-bearing word

Four of the five private-sector prohibitions turn on intent, and two require sole intent. That is a high bar for a prosecutor. A hiring model with a skewed output is not, by that fact alone, a TRAIGA violation, because section 552.056(c) says disparate impact "is not sufficient by itself" to show intent.

Our view is that this design explains the empty docket better than any theory about political will. A statute that punishes purpose rather than outcome produces few clean cases. The attorney general can still reach outcome-based harm, but through the deceptive practices law, where the question is what you claimed rather than what you meant.

Why a quiet TRAIGA record is built into the statute

Zero public cases does not mean zero activity. The enforcement pipeline in sections 552.102 to 552.104 has 5 stages, and only the last one is reliably public. A complaint, a demand, a notice and a cure can all happen without a filing anyone can search.

The cure statement is the off-ramp

Section 552.104 bars a lawsuit if the vendor cures within 60 days and sends a written statement. That statement must say the violation was cured, include supporting documentation showing how, and confirm changes to internal policies to prevent a repeat. A vendor that does this well is never named in a press release.

This is why the visible record will always undercount pressure. The attorney general can announce an investigation, as it often does, but the statute does not require it. If you are tracking Texas risk by reading headlines, you are measuring the failure rate of the cure process, not its volume.

The AI cases Texas did bring used older laws

Texas has been the most aggressive state attorney general on AI for two years. None of that activity needed TRAIGA. The office used the Deceptive Trade Practices Act, the Capture or Use of Biometric Identifier Act and the state's privacy statutes, which carry no cure period of the TRAIGA kind.

Compiled from Baker Donelson and Privacy World (2024), Sidley Austin in Chambers (2026), Hudson Cook (2025) and the National Law Review (2026). The attorney general's own releases blocked automated access during research.

Pieces is the template that matters

The Pieces case is the one a B2B vendor should study. The company had advertised a "severe hallucination rate" of under 1 per 100,000 to hospitals. The attorney general alleged the metrics were likely inaccurate. The settlement required Pieces to disclose the definition of any accuracy metric and the method used to calculate it.

According to Privacy World, those duties run for five years and carry no civil penalty. Baker Donelson lists the disclosures: training data and models, intended use, known limitations, foreseeable misuse and the documentation users need to spot inaccurate output. That list reads almost line for line like the TRAIGA demand categories, two years earlier.

The lesson is blunt. Your largest Texas exposure is the accuracy number on your sales deck. If you publish a benchmark figure, the method behind it needs to survive a subpoena, which is the same discipline we argue for in case studies with verified outcomes.

We think the preference is rational, and it will persist. The deceptive practices law punishes a false claim, which is easy to prove from marketing copy. TRAIGA punishes intent, and it gives the target a 60-day cure. For a prosecutor choosing between them on the same facts, the older law is faster and harsher.

The TRAIGA civil investigative demand is a documentation checklist

Section 552.103(b) is the most useful part of the statute for a vendor, because it tells you what the attorney general may ask for. Once a complaint arrives through the portal, the office may issue a demand covering 8 categories. Treat them as the outline of a document you should already have.

Most of this exists in some form at any vendor that has completed a security review or an EU AI Act exercise. The gap is assembly. If your AI bill of materials and your agent audit log live in different teams, a demand with a short return date will expose it.

The deployer clause deserves attention. For a deployer, monitoring includes "the oversight, use, and learning process" set up to handle deployment issues. That is a request for evidence of a working feedback loop, not a policy document. A vendor that can show tickets, fixes and dates will answer it more convincingly than one that shows a governance charter.

Penalties, cure statements and the safe harbor in the Texas AI law

The penalty schedule in section 552.105 is where most coverage starts, and it is the least likely part to touch you this year. Penalties apply only to violations that were not cured. They also cannot apply to a system that has not been deployed.

The attorney general decides what is curable

The statute sets $10,000 to $12,000 for a curable violation and $80,000 to $200,000 for one the attorney general treats as uncurable. A broken cure statement falls in the lower band, but it also ends your claim to good faith. Continuing violations add up to $40,000 a day.

A licensed business faces a second layer. Under section 552.106, a state licensing agency may sanction a licensee, including a monetary penalty of up to $100,000, but only after a court finding and a recommendation from the attorney general. That matters to vendors whose customers are licensed, such as clinics and insurers.

The safe harbor rewards finding your own problems

Section 552.105(c) starts with a rebuttable presumption that you used reasonable care. Subsection (e) then says a defendant may not be found liable if it discovered the violation through its own processes. Those include adversarial or red-team testing, feedback, and an internal review while substantially complying with the NIST Generative AI Profile or another recognised framework.

This is the most valuable clause in the law for a vendor, and it is underused. It converts testing records into a liability shield. A standing red-teaming programme with dated findings is no longer only a security practice in Texas. It is a defence you can cite by section number.

What B2B vendors should change, and what nobody can yet show they changed

The brief for this post asked what vendors changed. Honestly, no public dataset answers that. We found no survey, filing or disclosure that measures vendor behaviour change attributable to TRAIGA, and we will not invent one. What we can do is say what the statute rewards, and what we would change first.

Change the claims before the controls

Start with marketing, because that is where Texas has actually enforced. Audit every accuracy, safety and hallucination figure on your website and in sales decks. Each needs a stated definition, a calculation method and a date. If you cannot write those three things down, remove the number. That is a one-week job with a large effect.

Build the demand pack once

Next, assemble the 8-category file from the table above for each AI feature. Store it where legal can reach it without opening a ticket with engineering. The same pack answers most of the AI procurement questionnaires your enterprise customers already send, so the cost is shared.

Map your customers' disclosure duties

If you sell to Texas health care providers or government agencies, they carry strict disclosure duties under section 552.051. They will push those duties into your contract. Ship a clear disclosure component and default it on. Vendors who already did this work for the EU AI Act transparency rules or for California's 2026 AI laws can reuse most of it.

Where this reading is weakest

The central claim of this post is a negative, and negatives are hard to prove. The attorney general's website returned access errors to every automated request we made, so we could not read its news index directly. Our "zero" rests on a May 2026 statement by Sidley Austin, later explainers and our own searches. An announcement we missed would overturn the headline, and we would rather say that plainly than hide it.

Second, the portal launch date rests on secondary sources. The Cloud Security Alliance note says 1 September 2026. Newtral saw an AI complaint button in August that led to the general portal, and a law firm alert from earlier in the year said the portal was not yet live. The statutory deadline is certain. The exact activation date is not.

Third, our argument that the deceptive practices law will stay the preferred tool is a prediction. A single high-profile TRAIGA case, perhaps over a chatbot and self-harm, would change the picture quickly. Section 552.052 is the prohibition most likely to produce one, because a chatbot transcript can show what a system encouraged in a way that a hiring model's output rarely can.

Finally, federal pre-emption could make the whole question moot. Sidley's chapter notes Executive Order 14365 of 11 December 2025, and reports that the Commerce Department list of conflicting state AI laws had not been published as of early April 2026. We do not know where TRAIGA would land on such a list, and neither does anyone else yet.

Frequently asked questions

Where to start this month

Pick the cheapest of the three moves and do it first: pull every accuracy or safety number from your public pages and sales decks into one sheet. Next to each, write the definition, the calculation method and the date measured. Anything left blank comes down before the end of the month. That single exercise addresses the only theory of AI liability Texas has actually used against a vendor.

Then give one person ownership of the 8-category demand pack, with a target of producing it for your highest-risk AI feature within 10 working days. If that target feels hard, you have found your real Texas exposure, and it has nothing to do with a statute that has not yet been tested in court.

Related reading