Texas named a target variable your rating model may not have

On September 2 the Texas Department of Insurance told every company it regulates that setting a premium by how likely a policyholder is to shop is unfairly discriminatory. Two of the three code chapters that bulletin rests on are chapters the department’s June bulletin on artificial intelligence had already named as law that AI supported consumer decisions must satisfy.

This week’s story: what a rating model is allowed to be aimed at

Commissioner’s Bulletin B-0007-26 went out on September 2 to “All insurance companies and their agents and representatives.” It defines price optimization as varying premiums on factors unrelated to the policyholder’s risk of loss or the company’s expense, and says the practice includes considering “the elasticity of demand” to predict whether a policyholder will renew or shop. The department’s news release that day describes it as covering home and auto rates. The bulletin’s own addressee line carries no such limit.

The definition has two halves: factors unrelated to risk of loss or expense, used “so that the company can charge the highest price that policyholders will tolerate before shopping for a new policy.” The test is an outcome test, “any practice that results in two policyholders with the same risk profile getting different premium increases,” and the bulletin states the reach directly: “Any use of price optimization in the ratemaking or pricing process, or in establishing a rating plan, is unfairly discriminatory and violates the Insurance Code.” Nothing in it turns on which technique produced the number.

B-0007-26 never uses the words artificial intelligence, model, or algorithm. It does not have to. In June the same department issued B-0003-26, addressed more broadly to all regulated entities and their agents, saying that decisions or actions “impacting consumers that are made or supported by advanced analytical and computational technologies, including artificial intelligence (AI), must comply with all applicable insurance laws and regulations,” then listing the chapters it meant. Chapter 544 on prohibited discrimination and Chapter 560 on prohibited rates are on that June list, and are two of the three the September bulletin rests on. Our report sets out what it expects: what that June bulletin asks of a governance program.

The two bulletins do not describe Chapter 544 the same way. June’s calls it a prohibition on “refusing or limiting coverage on the basis of race, color, religion, national origin, age, gender, marital status, geographic location, or disability.” September’s calls it a prohibition on unfair discrimination “between individuals of the same class and of essentially the same hazard.” Our reading is that a program built around the first description has no reason to look at an elasticity term: the policyholders it separates are in the same class before the model runs.

What to do this week

• If you have a filing in force in Texas, find the part that lists its factors and says what each one is there for. Read it for anything that measures a policyholder’s behavior rather than their risk: renewal odds, shopping, tolerance for an increase. That reading takes a minute, and it is the half of the bulletin most likely to touch a live filing.
• Whoever keeps your model documentation can say which rating models carry a demand component and which do not. Get it as a list, one line per model, with a name beside each line. The test in the bulletin is applied to a rating plan, not to a modeling team, so the list is the thing that has to answer. Texas is one of four jurisdictions with a framework of its own rather than the NAIC model bulletin, which our record puts at 25, so run the list against where each state stands before assuming one answer travels.
• Version 5.0 of the NAIC supplement closes for comment on September 29, and the working group has now dated the meeting that follows it, October 8. Decide this week which of the two gets your hour, because they ask for different work: a written comment, or an hour on a call.

On the Docket

• September 14, 2026: TDI owes the Governor’s office the additional administrative actions it can take immediately, and any statutory changes it thinks are needed. Anyone writing home or auto in Texas, since the price optimization bulletin came out of the same letter. Governor’s directive to TDI

• September 29, 2026: The comment period on version 5.0 of the AI Risk Evaluation Supplement closes at the end of the business day. Anyone who has read the draft and not yet written. NAIC working group page

• September 30, 2026: Under the California constitution a bill passed before September 1 and in the Governor’s possession on or after that date becomes a statute if it is not returned on or before this day. AB 1405, the AI auditor registry bill, was enrolled on September 3, and the legislature’s record does not yet show it presented to the Governor. Anyone who expects to buy an outside review of a model. California Constitution, article IV, section 10

• October 8, 2026: The Big Data and Artificial Intelligence (H) Working Group meets by public Webex for one hour at 11:00 AM ET, to continue public discussion of the supplement. No agenda or materials were posted for that session as of September 8. Anyone who wants to hear what the comments produced. NAIC working group page

This week in brief

Governance

Eleven letters on the NAIC’s vendor registry drew five general themes, the first of them “concerns with the attestation asking vendors to certify compliance they cannot control.” The minutes leave revisions to a drafting group, and the working group session listed next is regulator only; our account of the August 12 session has the exchange: NAIC vendor registry draws eleven comment letters. Committee minutes

A House bill introduced September 1 would keep an AI system from issuing a clinical denial under an employer or individual plan, and would count AI used in utilization review as a treatment limitation for parity purposes. Denial notices and claim files are where that lands, and our report on the bill walks through both: the Doctors Not AI Act would bar AI from clinical denials. H.R. 10210

California would let only registered auditors sell one kind of AI review. Beginning January 1, 2029, AB 1405 bars an unregistered person from offering, selling, or conducting a “covered AI audit,” which it defines as one assessing “internal controls, processes, or systems implemented for an AI system or model that are necessary for compliance with state law,” and our reading is that an insurer’s AI governance controls, kept because a department expects them, are controls of that kind. AB 1405 text

From the Guides

One chapter, described twice. Our entry on unfair discrimination covers the sense a governance program is built to test for: protected traits, and the neutral variables that stand in for them. Chapter 560 gets two descriptions from the same department: June’s bulletin has it requiring that rates “must be just, fair, reasonable, and adequate,” and September’s carries the rest of that sentence, “and may not be confiscatory, excessive, or unfairly discriminatory.”

One question runs through this week’s bulletin: what is the model aimed at? Reply to the email this issue arrived in with the answer for one of yours, even a partial one. We read every reply.

– The Editor, InsureAI Wire