If your adjuster cannot prove they reviewed the AI output, the AI output is the decision. And it's discoverable.
That is the operating implication of the March 23, 2026 ruling in Estate of Gene B. Lokken v. UnitedHealth Group, Inc., No. 23-CV-3514 (D. Minn.). A Minnesota federal court granted plaintiffs' motion to compel discovery into the development, deployment, and use of UnitedHealth's nH Predict algorithm in post-acute care decisions. The court allowed plaintiffs to seek documents about how the program was built, what its anticipated benefits were, and "whether it was designed to supplant physician decision-making." Internal investigations were largely shielded. Financial incentive data was kept out. Almost everything else about how the AI worked, who trained employees on it, and what outside investigators had asked about it, came in.
The decision is an evidentiary ruling about what the plaintiff can examine, well short of a final judgment on whether the AI did anything wrong. That distinction is what makes it so significant for the rest of the industry. A discovery ruling does not require the carrier to have done anything wrong with AI. It requires the carrier to produce the record of what the AI did and what humans did about it. Carriers that cannot produce that record have a problem long before they get to a merits argument.
The legal commentary picked this up immediately. As the Hunton Insurance Recovery Blog noted, the doctrine extends past health insurance: policyholders can now seek discovery into whether AI was used to "supplant the adjuster's decision-making authority" in property and liability claims as well. The National Law Review's analysis reached the same conclusion. If an insurer denies a claim based primarily on AI output without a meaningful human review on the file, the discovery the Lokken court allowed would support a bad-faith argument, especially when the AI produced an error or hallucination that went unchallenged inside the workflow.
For carriers writing property, casualty, and liability lines, the implications run in three directions. Each is worth understanding separately, because each requires a different operational fix.
What Lokken Actually Allowed, and What It Means for Property and Liability Adjusters
Strip away the case-specific facts and look at what the court approved. Plaintiffs got documents on how the AI program works, on its design goals, on the question of whether it was built to replace human judgment, on employee training materials covering AI use, and on the identity of employees involved in AI deployment. They also got documents on government investigations into the AI's use, with internal corporate investigations carved out.
That is a wide opening. The framework the court applied, particularly the question of whether the system was "designed to supplant" professional judgment, does not depend on the medical context. The same framework reads cleanly onto a claim adjuster's decision file in property, an underwriter's risk classification in commercial lines, or an SIU referral in fraud workflows. Any decision where the carrier asserts a human professional made the call is now an invitation, in a bad-faith suit, for the plaintiff to demand the documentation that proves it.
The doctrinal expansion is the most important downstream consequence. Plaintiffs' counsel will not need a second appellate-level case to start invoking Lokken in property and liability matters. They will cite it directly. Carriers facing first-party property bad-faith suits, third-party liability defense matters, and disability or workers' compensation disputes should expect AI-related discovery requests to land in their next several discovery cycles. The requests will not be exotic. They will look like the requests Lokken approved: how the model works, who trained on it, what governance documentation exists, what the human reviewer actually did.
A carrier whose claim file consists of an AI-generated severity assessment, an adjuster's signature, and a payment record is not going to enjoy that discovery cycle. The signature does not document the review. It documents the closure of the file.
The Privilege Trap Most Carriers Have Not Mapped
The Lokken court approved discovery into the AI program itself. A separate strand of decisions, more troubling for carriers, addresses what happens when AI is used in legal analysis or coverage review.
In USA v. Heppner, Judge Jed S. Rakoff held that documents created using a public AI tool were protected by neither attorney-client privilege nor work product doctrine, even after counsel was looped in on the analysis. The court's reasoning was that prompts directed to a public AI system, rather than communications directed to counsel, broke the privilege chain. The defendant's later consultation with attorneys did not retroactively cloak the AI-generated material in privilege.
The implication for insurance is direct. Coverage analysts, SIU investigators, claims professionals, and in-house counsel who use general-purpose AI tools to analyze coverage questions, draft denial language, or assess litigation posture may be producing materials that an opposing party can subpoena. The work-product fence the carrier assumed was in place does not extend to AI outputs created on platforms whose terms allow third-party access to the prompts and responses.
A carrier facing a bad-faith suit may now find itself producing not only the AI outputs that influenced the underlying claim decision, under the Lokken doctrine, but also the AI outputs the carrier's own coverage team generated while analyzing the claim, under the Heppner reasoning. Those two production obligations together describe a record set most carriers have never inventoried, much less retained, much less segregated by privilege status.
The fix is procedural, and it is not optional. Carriers need a documented policy on which AI tools may be used for which categories of work, with strict controls on the use of public AI platforms for legal-analysis-adjacent tasks. The policy should be backed by audit logs that prove what tool was used for what purpose. This is the substrate that allows in-house counsel to make defensible privilege assertions in the next round of litigation. Without it, every privilege log is a guess and every protective order is a fight.
The Operational Fix: Per-Decision Human Review as a Verifiable Artifact
The Lokken doctrine and the Heppner privilege analysis both point to the same operational requirement. Claim files, underwriting files, and SIU files need to contain a verifiable record, per decision, that documents what AI did, what the human reviewer did, what the human changed, and when.
That is the artifact specification. A system-events log does not satisfy it. A decision-level human action log does. Three components are non-negotiable.
The first is AI output capture. Every AI output that touched a decision must be retained in its original form, with a timestamp, a model version reference, and the input data that produced it. If the carrier cannot reproduce the AI output the adjuster reviewed at the time of the decision, the discovery request is going to be answered with a shrug, and the shrug will be expensive.
The second is human review attestation. The reviewer must take a recorded action, captured at the decision moment, that documents which AI output they reviewed, what they accepted, what they overrode, and on what reasoning. A workflow that asks the adjuster to "review and approve" without capturing the substance of the review produces a record that looks, in deposition, like a rubber stamp. The Lokken plaintiffs argued exactly this point in their motion to compel.
The third is decision finalization with attribution. The final claim disposition must be linked, in the audit trail, to the human reviewer who finalized it, with a reference to the AI artifacts they reviewed and the documented overrides. The chain has to close. A claim file that documents AI input and human output but does not connect the two is a file that produces, in discovery, the worst of both worlds: evidence that AI was involved and no evidence that a human meaningfully engaged with it.
Carriers that build examination-grade audit trails for AI decisions into their claim and underwriting platforms now will be ready when the discovery requests land. Carriers that retrofit them after the first Lokken-style motion to compel arrives will be operating on the plaintiff's timeline. The artifacts the operational fix produces are the same artifacts that satisfy the NAIC AI Systems Evaluation Tool's Exhibit C and Exhibit D requirements, and the same artifacts that defend against the bias-testing claims that will travel alongside bad-faith allegations as the litigation pattern matures.
Where the Litigation Wave Is Heading
Lokken is the visible front edge of a much larger shift in how courts handle AI-influenced insurance decisions, and it will not be the only ruling of its kind in 2026. The plaintiff's bar has spent two years building the legal theory. The court rulings catching up to that theory, in 2026, are going to remake bad-faith litigation in ways that compound through every state where similar suits are filed.
The carriers that come through this period in good shape will not be the ones who used AI most cautiously. They will be the ones whose AI use is accompanied, on every decision, by a human review record they can produce on a 30-day discovery deadline. The technology choice is secondary. The artifact choice is primary.
A claim denial without a defensible audit trail is a discovery problem waiting for a complaint. The carriers building that audit trail today will read the next round of Lokken-style rulings as ratification. The carriers without one will read them as forecasts.
