AI Tools and Bad Faith Risk in Insurance Claim Handling: Lessons from Lokken
Artificial intelligence (AI) is increasingly used in claims handling through predictive analytics, automation, fraud detection, and cost estimation. While these tools provide speed, consistency, and accuracy, they also raise litigation risks—plaintiffs may challenge both outcomes and the AI-driven process. A 2025 case, Estate of Lokken v. UnitedHealth Group, Inc., illustrates how plaintiffs may plead that AI replaced individualized judgment and how courts may treat those allegations. The main takeaway: replacing human judgment with AI may increase exposure to allegations of bad faith and invasive discovery into insurer claims handling processes.
The Case of Lokken
In Lokken, the plaintiffs were insured under Medicare Advantage plans sold or administered by UnitedHealth entities. The policyholders sought coverage for post-acute care, were denied, and alleged that the denials caused serious harm, including worsening injuries and, in some instances, death. The factual centerpiece was the allegation that an AI tool—“nH Predict”—effectively substituted physicians’ judgment by applying rigid criteria, generating estimates based on comparisons to “similar” patients, and driving denials even when treating providers recommended additional care. The plaintiffs also alleged that the tool was inaccurate and pointed to high reversal rates on appeal, and that the appeals process was frustrated through repeated denial letters or late payments that allegedly avoided exhaustion of administrative remedies.
The Lokken court dismissed most state-law and statutory claims by applying the Medicare Act preemption. Critically, however, two claims survived: breach of contract and breach of the implied covenant of good faith and fair dealing. The court allowed those claims to proceed because they could be resolved without imposing state-law standards that regulate Medicare’s benefits framework.
Where AI Risk Shows Up
Lokken did not tackle the substance of the plaintiffs’ bad faith claims premised on the use of an AI tool, and its Medicare context includes doctrines—like preemption—that do not map neatly in the property and liability insurance context. But it does show how courts and litigants are approaching AI in claims-adjacent decisions: by applying traditional legal concepts to modern tools and focusing on whether automation replaced individualized judgment.
In Lokken, the plaintiffs framed AI as replacing individualized professional judgment. Similar allegations of overreliance or rubber-stamping can arise in other types of claims when AI tools are used to set scope and pricing, flag coverage issues, recommend causation conclusions, or drive SIU referrals. The more an adjuster’s role looks like confirming an AI recommendation, the more a plaintiff may argue the carrier failed to conduct a reasonable, claim-specific evaluation.
Another risk is explainability, which quickly becomes a discoverability problem. If AI materially influenced a claim decision, counsel should expect discovery demands aimed at model configuration, thresholds, training data sources, vendor communications, override rates, and internal guidance on how staff should use the output. Weak governance can fuel arguments that the investigation was unreasonable—even if the carrier ultimately prevails on coverage.
Data quality and bias are also risks. If an AI tool is trained on historical data embedded with past adjusting practices or relies on unsuitable “similarity” comparisons, it may introduce systematic estimate errors. From a litigation standpoint, pattern-based inconsistencies can become the narrative, even when any single claim decision appears defensible.
Finally, Lokken highlights risks arising from operational incentives and control failures. Productivity metrics and workflow performance standards are not inherently problematic, but if they functionally penalize adjusters for deviating from AI outputs or taking time to investigate exceptions, they can be recast in litigation as institutional pressure favoring speed and cost containment over accuracy.
What Is An Insurer To Do?
The defense posture is familiar: the file should demonstrate the facts gathered, the policy language applied, what the AI tool contributed, whether the output was tested against claim-specific evidence, and why the final decision is reasonable. Adjusters should document their reasoning and any reliance on or deviation from AI recommendations to ensure transparency for neutral reviewers.
For claims professionals, the most durable takeaway is that AI will be evaluated as part of the claim-handling process and will be discoverable like other decision inputs. As AI capabilities develop and legal standards evolve unevenly across jurisdictions, today’s AI-driven workflows will be examined in litigation through the familiar lens of reasonable investigation, policy-based decision-making, and good-faith conduct.
