AI Claim Management Software for Public Adjusters: What to Look For in 2026

On June 12, 2026, the Texas Department of Insurance issued Commissioner's Bulletin B-0003-26, and the address line matters more than the subject line: "To: All regulated entities and their agents and representatives."¹ Read that line closely before assuming it skips you. The bulletin never says "public adjuster," and it never cites Chapter 4102, the chapter that licenses you.¹ The opening phrase is what catches you: Section 4102.051 requires a license to act as a public adjuster in Texas, and a licensee the commissioner licenses sits inside "all regulated entities" no matter what the qualifier after it says. If any part of your file passes through an AI tool, the department that would examine that file is the department that wrote this.

Most AI claim software is sold on throughput. Faster estimates, auto-summarized policies, drafted correspondence, fewer billable hours per file. All of that is real, and none of it is what a regulator will ask you about. The regulator's question is narrower and much harder for software to answer after the fact: on this file, on this date, did a licensed human review the machine's output and agree with it before it went out?

Buy against that question.

What the Texas bulletin actually requires

B-0003-26 reads like housekeeping until you reach the operative sentence: "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."¹ Reviewing can be passive. Agreeing cannot. Agreement is an affirmative act by a named person, and an affirmative act either got recorded at the moment it happened or it did not happen in any way you can later evidence. Most claim software records the output. Far less of it records the agreement.

The rest of the bulletin is governance language, and it is the ordinary kind: controls proportionate to the risk of the use case, verification and testing sufficient to identify errors and unfair bias, written documentation of how the system is governed. TDI issues it against Chapters 401, 541, 542, 544, 560, 751, 831, 4001, 4101, and 4201 of the Insurance Code.¹ One chapter on that list looks like it names you and does not. Chapter 4101 licenses adjusters working the carrier's side of the file: Section 4101.001(a)(1) defines an adjuster as a person who investigates or adjusts losses "on behalf of an insurer." Your license sits one chapter over in 4102, and 4102 appears nowhere in the bulletin.

So how far does the address line actually reach? On the narrow reading, "their agents and representatives" means the agents and representatives of a regulated entity, and a public adjuster represents the insured rather than the carrier, which leaves you outside that clause. The clause in front of it carries you instead. TDI licenses public insurance adjusters under Section 4102.051, disciplines them, and under Section 4102.110 may examine the records they must keep.⁸ That combination, licensed, sanctioned, and examined by the commissioner, is what a regulated entity is in the plain sense of the phrase. TDI has published nothing settling the point, so treat the reach as unresolved and the standard as settled.

The bulletin creates no new statutory duty; it tells regulated parties how TDI reads duties that already exist, which is what a commissioner's bulletin does. The practical consequence lands on your file cabinet rather than on your license: TDI states it may request AI governance documentation during an examination or investigation.

The citation habit that gets public adjusters in front of the UPL committee

Texas has barred licensed public adjusters from practicing law since 2003, when the legislature created the license and wrote the bar into the same act. Section 4102.156 of the Insurance Code carries that bar today, added in the 2005 nonsubstantive recodification: a license holder "may not render services or perform acts that constitute the practice of law, including the giving of legal advice to any person in the license holder's capacity as a public insurance adjuster."² The Texarkana Court of Appeals decided Kubala Public Adjusters, Inc. v. Unauthorized Practice of Law Committee, 133 S.W.3d 790 (Tex. App.—Texarkana 2004, no pet.), seventeen months before that section number existed, reviewing an injunction against sixteen of a public adjusting firm's business practices under the unauthorized-practice-of-law rules that governed the conduct at issue. The line has been settled long enough that most licensees can recite it.

Where licensees actually get caught is narrower than the statute. Larry Bache of Merlin Law Group, writing on the Property Insurance Coverage Law Blog in February 2014, identified the recurring pattern as a public adjuster citing case law in letters to the carrier.³ Twelve years on, that is still the practitioner consensus about what draws a referral, and it happens to be the default behavior of a general-purpose language model too. Ask one to argue a disputed exclusion in a demand letter and it will reach for authority, because reaching for authority is what a model trained on legal and quasi-legal text does when the prompt sounds adversarial. Nobody has to instruct it. No prompt has to say "cite something." The model infers that a persuasive letter about coverage ought to carry a case name, produces one in correct citation format, and hands it back with the same flat confidence it applies to the deductible math. The habit arrives in the box, and it arrives looking like competence.

The second failure mode is the citation that does not exist at all. Damien Charlotin's AI Hallucination Cases database, maintained at HEC Paris, had logged 1,810 court decisions addressing AI-fabricated material as of July 25, 2026, with 1,251 of them in U.S. courts. Of those, 312 carried a monetary penalty and 126 carried a professional sanction such as suspension, bar referral, or removal.⁴

The money moved with the volume. The sanction in Mata v. Avianca, Inc. (S.D.N.Y., June 22, 2023), the case that first made this a national story, was $5,000 against two attorneys and their firm. The largest U.S. monetary sanction in the database is $110,204, in Couvrette v. Wisnovsky (D. Or., March 23, 2026), roughly twenty-two times the Mata figure.⁴

Read those numbers with the right amount of caution. Zero of the 1,810 decisions are categorized in an insurance legal field, and none of them sanction a public adjuster.⁴ The pattern is a courtroom phenomenon, and the exposure to your license runs by extension rather than by direct precedent. What the extension looks like in practice is a carrier's counsel reading a fabricated citation in your demand letter and forwarding it, along with a complaint, to the department that issued your license.

Why the NAIC's insurer-only bulletin doesn't reach you

The National Association of Insurance Commissioners adopted its Model Bulletin on AI on December 4, 2023, and 25 jurisdictions had adopted some version of it as of April 1, 2026⁵, the document your vendor most likely means when it says its product is built for AI regulatory compliance. The title of the model is "Use of Artificial Intelligence Systems by Insurers." Its scope is insurers. A public adjuster is a licensee representing the insured, and the model bulletin's governance program, board oversight, and third-party vendor obligations all attach to a carrier's AI systems program.

The adoption curve also stopped. Jurisdictions went from zero to twenty-four in the fifteen months after adoption, then added exactly one more (Hawaii, on December 10, 2025) over the following nine.⁵

The gap this leaves is specific. A product built to the model bulletin will arrive with governance artifacts around model risk, bias testing, and third-party oversight, all framed as an insurer's program obligations. None of that framework speaks to whether the tool keeps a public adjuster clear of the practice-of-law bar, whether it preserves an adjuster's records for the statutory five years, or whether it recorded that you personally approved a settlement position before it went to the carrier. Those are licensee duties. They attach to the individual holding the license rather than to an enterprise compliance function, which is why a vendor's NAIC posture answers a question you were not asked.

Texas doesn't appear on that list of twenty-five — the NAIC's own status map handles Texas separately, under a heading for state-specific insurance regulation and guidance, and the entry it cites there is Bulletin B-0036-20, dated September 30, 2020.⁵ That map, current as of April 1, 2026, predates B-0003-26 by more than two months, which is a fair reminder that the national compliance picture your vendor is selling against lags the state rule you are actually licensed under.

What discovery in Lokken means for your own AI tool

In Estate of Gene B. Lokken v. UnitedHealth Group, Inc., No. 23-cv-3514 (JRT/SGE) (D. Minn.), Magistrate Judge Shannon G. Elkins granted in part a motion to compel on March 9, 2026, in an order docketed at 162. The discovery reached how the defendant's nH Predict algorithm functions and who sat on the company's internal AI Review Board. The order treats oversight documents as "directly relevant to its alleged conduct, as is the identity of the individuals who conduct that review and oversight."⁶

Lokken is a Medicare Advantage case against a health insurer in a Minnesota federal court. It involves no property claim, the defendant is a carrier rather than a policyholder representative, and no part of it binds a public adjuster in Texas or anywhere else. Nothing in that order creates a duty you did not already have.

What travels is the logic, and the logic is portable to any dispute in which a party used software to shape a consequential decision. Once oversight is asserted as a defense, oversight becomes a subject of discovery, and the party asserting it has to produce documents and names. Your tool's audit trail is what would get produced. If it stores a boolean where a name should be, or overwrites the reviewer field on the next edit, or exports a formatted PDF summary in place of the underlying record, you would be answering that request with an explanation instead of a document.

One honest gap runs through all of this. No survey of AI adoption among public adjusters exists, so nobody can tell you what share of your peers are already running these tools or how they are configured. The Bain figures in the panel above (78% of 160 surveyed insurers adopting generative AI somewhere, 4% scaling it in claims) describe carriers, and carriers are the other side of your file.⁷ The regulatory record is real and dated, the litigation record is real and dated, and the adoption baseline for your own profession is simply missing.

Seven things to check before you buy

Each row below is something a vendor can demonstrate on a live file during a thirty-minute call, or cannot. Ask for the demonstration on a real claim rather than a sandbox, and watch what the software writes down as it happens.

An examiner, a UPL committee, and opposing counsel with a document request will all ask about one specific claim rather than about the product in general: what the software produced on that file, who read it, and what that person did next. Nothing on a feature slide answers that. The answer, if you have one, is a dated line in the record with a name attached to it.

Start with the export. Ask for the full five-year record set on a closed claim before you ask about anything the software generates, because every other answer on that list is only worth what the export can carry out.⁸