Texas Puts a Human on Every Consequential AI Decision

The Texas Department of Insurance told carriers on June 12, 2026, that a person must review and agree before any consequential AI decision takes effect. The June 12 bulletin on AI use, numbered B-0003-26, applies to every regulated entity, its agents, and any third party working with it.

TDI says it is not writing a new playbook. The line that will show up in exams is simpler: if the tool decides something that hits a consumer, a human has to own it first.

TDI Expects a Person to Sign Off First

The bulletin’s opening reminder is familiar. Decisions or actions that affect consumers, made or only supported by AI and other advanced tools, still have to meet Texas insurance law, including unfair trade practice and unfair discrimination rules. TDI’s stated interest is consumer protection, not a ban on the software.

Then the department gets specific about who is on the hook when the software is in the loop.

If a regulated entity uses AI to make a consequential decision, TDI expects a person to review and agree with all decisions before action is taken.

Texas Department of Insurance, Commissioner’s Bulletin No. B-0003-26

That is a sign-off rule, not a model-governance essay. The bulletin never defines “consequential.” It does say the decision cannot be inaccurate, arbitrary, capricious, or unfairly discriminatory, no matter which tool produced it. Firms are told to put controls in place because tools without guardrails raise those risks.

TDI also says it will watch AI through examinations and product filings, and that it will take consumer complaints about AI use. Entities that use the tools “should be able to furnish procedures and protections upon request.” The same notice says its goal is not to prescribe specific practices or documentation requirements. The file still has to exist when an examiner asks.

How SB 815 Treats Automated Health Denials

For medical-necessity denials, Texas already went further than a bulletin. Senate Bill 815 added Insurance Code §4201.156, and TDI’s AI notice points straight at Chapter 4201 when it says the code “prohibits using AI to make an adverse determination.”

The statute is blunt. A utilization review agent may not use an automated decision system to make, wholly or partly, an adverse determination. The commissioner may audit and inspect that use at any time. The Senate passed the bill on March 26, 2025, by 30 yeas, 0 nays, and one present not voting. The House passed it on May 24, 2025, by 116 yeas, 13 nays, and two present not voting. The governor signed it on June 20, 2025. It took effect September 1, 2025.

The operational date is later. The amendments apply to utilization review on health benefit plans delivered, issued for delivery, or renewed on or after January 1, 2026. Older plan years stay under the prior law. So the June bulletin did not invent a human-in-the-loop idea for Texas health denials. It spread a weaker version of that idea, review and agree rather than a flat ban, to rating, claims, underwriting, and the rest of the consumer file.

THE TEXAS AI INSURANCE CLOCK

1. September 30, 2020: TDI reminds carriers they own the accuracy of third-party data used in rating, underwriting, and claims.
2. December 4, 2023: The NAIC adopts its model bulletin on insurers’ use of AI systems, including a written AIS Program.
3. June 20, 2025: SB 815 is signed. Section 4201.156 bars automated systems from making adverse determinations, wholly or partly.
4. January 1, 2026: That ban applies to utilization review on plans issued or renewed on or after this date.
5. June 12, 2026: TDI issues B-0003-26, extending a human review-and-agree rule to consequential AI decisions across regulated lines.

TDI’s bulletin page for B-0003-26 was last updated July 16, 2026. The text of the notice did not change into an NAIC clone in that revision.

Texas Skipped the NAIC’s Written AI Program

The NAIC’s 2020 Principles on Artificial Intelligence are the high-level guide TDI says it recognizes. Fairness, accountability, legal compliance, transparency, and safe systems sit in that document as guidance, not as a statute. The 2023 model bulletin built a much thicker set of exam expectations on top of those principles.

Under the model, insurers are expected to develop, implement, and maintain a written AIS Program the model expects for AI systems that make or support regulated decisions. That program is supposed to cover governance, risk controls, and internal audit, scaled to consumer harm, human involvement, explainability, and third-party dependence. As of April 1, 2026, the NAIC’s own implementation map listed 25 jurisdictions as having adopted that model. Texas was still listed there under Bulletin B-0036-20, a 2020 notice about third-party data, not the 2023 template.

B-0003-26 borrows the model’s consumer-protection opening and its list of existing unfair-practice laws. It does not borrow the AIS Program. TDI says it will not prescribe formats. The practical substitute is the person who has to agree before the consequential output ships, plus whatever procedures the company can hand over when TDI asks.

THREE TEXAS TRACKS FOR INSURANCE AI

Carriers that already keep model files for credit or fraud tools have a head start on the Texas exam, because they are used to showing how a system was tested and who can override it. They still have to map that habit onto underwriting, pricing, and claims, where the June bulletin now says a person agrees before the action.

Vendors, Agents, and the File TDI Can Demand

The bulletin’s reach is wider than the company that trained the model. Expectations “extend to any third party working with a regulated entity.” Agents stay under Chapter 4001. Adjusters stay under Chapter 4101. Official acts listed there, from taking an application to investigating a loss, are still licensed-person work even if software drafted the first pass.

That vendor rule is not new in spirit. TDI’s 2020 notice on third-party data already told regulated entities they own the accuracy of data used in rating, underwriting, and claims, even when a vendor supplied it. B-0036-20 also said it did not create a new legal duty. B-0003-26 takes the same posture on documentation, then adds the human agreement step for consequential AI output.

The cost shows up in contracts and in the claim file, not in a new Texas AI license. If a third-party platform scores a risk, triages a claim, or drafts a denial letter, the regulated entity still has to show the outcome was not unfairly discriminatory and that a person agreed before anyone acted. Shops that sold claim speed as a hands-off product now have to show, file by file, that a licensed person reviewed and approved the consequential step. A rubber stamp on a system default is the fact pattern TDI’s wording is built to catch.

TDI lists the statutes it expects firms to keep in view when they turn the tools on: Chapters 541 (unfair methods of competition and unfair or deceptive acts), 542 (claims processing and settlement), 544 (prohibited discrimination), 560 (prohibited rates), 831 (corporate governance annual disclosure), 4001 (agents), 4101 (adjusters), 4201 (utilization review agents), 751 (market conduct), and 401 (audits and production of records). That is ten chapters. None of them is an AI statute. All of them now travel with the model.

What Examiners Will Ask About AI Tools

Because TDI refused a prescribed packet, the exam becomes the instruction manual. The bulletin tells firms to expect questions on governance, risk management, data and privacy protections, and internal controls. It also says TDI may ask about any specific use of AI or its application. Product filings are in the same monitoring bucket as examinations.

WHAT TDI SAYS IT MAY PULL

• Governance: How the company oversees development, purchase, and use of AI, including work done by vendors.
• Risk and testing: Controls meant to stop inaccurate, arbitrary, capricious, or unfairly discriminatory outcomes, plus methods used to find errors and bias.
• Privacy and data: Protections around the data that feeds the tools, consistent with the consumer-protection purpose TDI states at the top of the notice.
• The human step: Evidence that a person reviewed and agreed with consequential decisions before action, not after the letter went out.
• A named use: How a particular model is used in underwriting, rating, claims, or customer contact, if TDI asks about that application.

Chapter 401 already requires firms to keep, and produce on request, the books and papers that relate to the business. Chapter 751 is the market-conduct hook. The June bulletin is TDI saying those hooks now reach the AI file even though the department will not hand out a template for that file.

Health utilization review is the harder case. Section 4201.156(b) lets the commissioner audit automated-system use for that work at any time, which is a tighter inspection right than the general “furnish procedures upon request” line in B-0003-26. Carriers that use the same vendor stack for health denials and for property claims will be answering two different questions: whether an automated system made the medical-necessity call at all, and whether a person agreed with other consequential AI output before anyone acted.

TDI Pointed Firms to a State Ethics Code

The other document TDI tells companies to read is not an insurance form. It is the Texas Department of Information Resources’ AI Code of Ethics and Minimum Standards, adopted for state agencies and local governments under Government Code §2054.702 and housed in 1 Texas Administrative Code Chapter 219. TDI recommends it as a framework. The code itself is written for government users, not for carriers. TDI still lists its themes: human oversight and control, fairness, accuracy, redress, transparency, data privacy, security, and accountability.

DIR’s own ethics language treats human review as a lifecycle duty. Governmental entities must deploy AI so people can review inputs and outputs at appropriate intervals, with more oversight on systems that face higher risk. TDI’s insurance bulletin is narrower and, on consequential decisions, stricter. It wants agreement before action, not a periodic sample after the fact.

WHAT THE DIR CODE PUTS ON THE PAGE

• Human oversight and control: People stay able to review inputs and outputs; higher-risk systems get more of that review.
• Fairness and accuracy: Outputs are supposed to be checked against legal rights and against error, not only against speed.
• Redress and transparency: There is a path to challenge an outcome, and the use of the system is supposed to be explainable to the people it affects.
• Privacy, security, and liability: Data protections sit next to a named accountability trail for the system’s use.

Read together, Texas now has a statutory ban on automated health denials, a department bulletin that puts a person on every consequential AI insurance decision, and a state ethics code TDI is using as a pointer rather than as a rule. The missing piece is the written AIS Program that 25 other jurisdictions had already copied from the NAIC by April 1, 2026. Texas firms that waited for that template will not get one from Austin. They will get exam questions, complaint files, and a request to show who agreed before the system was allowed to act.

Frequently Asked Questions

Does the TDI AI bulletin create new legal duties for insurers?

TDI frames B-0003-26 as a reminder of existing Insurance Code duties, and it says the notice does not prescribe specific practices or documentation formats. That matches the posture of Bulletin B-0036-20 from September 30, 2020, which also said it did not create a new legal duty while still putting carriers on the hook for third-party data. The enforceable pieces remain the ten chapters TDI lists, plus whatever a market-conduct exam can show about how a firm used the tools.

Can a Texas health plan use AI to deny a medical-necessity claim?

Not to make the adverse determination, even in part, once SB 815’s amendments apply to that plan year. Section 4201.156(c) still allows an algorithm, AI system, or automated decision system for administrative support or fraud-detection functions. The denial itself has to come from the human process Chapter 4201 already required, now with the automated-system ban sitting on top.

Does Texas require a written AIS Program like the NAIC model?

No. The NAIC model bulletin adopted December 4, 2023, tells insurers to keep a written AIS Program for AI systems that make or support regulated decisions. TDI’s June 12, 2026, bulletin never uses that term and says it will not prescribe documentation requirements. Firms still need something they can produce, because TDI says it will ask for procedures and protections and may question any specific AI use.

What counts as a consequential AI decision in the TDI bulletin?

The bulletin never defines the phrase. DIR’s state-agency rules, which TDI tells companies to review, treat a heightened-scrutiny system as one meant to make, or be a controlling factor in, a consequential decision. TDI did not import that definition into B-0003-26. Until it does, the safe reading inside a carrier is any AI-supported action that changes coverage, price, payment, or a consumer’s access to a benefit.

Do third-party AI vendors have to follow the Texas bulletin?

TDI addresses regulated entities and says the expectations extend to any third party working with them. The department’s enforcement path still runs through the licensed insurer, agent, or utilization review agent, which is the same pattern as the 2020 third-party data bulletin. Vendor contracts that hide the model, block bias testing, or leave no record of human agreement are now an exam problem for the Texas licensee that hired the tool.

Questions on the bulletin go to GovernmentRelations@tdi.texas.gov, the contact TDI printed on B-0003-26. The exam request will not come in by that inbox. It will come as a demand for the procedures, the test results, and the name of the person who agreed before the system was allowed to move.

Disclaimer: This article is news reporting and analysis of a Texas Department of Insurance bulletin and related Texas statutes. It is for information only and is not legal, compliance, insurance, or regulatory advice. Readers who need to apply these rules to underwriting, claims, vendor contracts, or a TDI exam should consult a licensed Texas insurance lawyer or a qualified compliance professional. Figures, effective dates, and the status of TDI guidance reflect the official texts cited here and can change if the department issues a later bulletin, order, or rule.