HYPERNOVELTYINSTITUTE

Technical intelligence
for accelerating change.

Search the record

Home/Analysis

Daily Hypernovelty Lead · November 14, 2026

The Insurer Still Owns the Algorithm’s Decision

An insurance adjuster reviews the remains of a storm-damaged home.
Photo: Andrea Booher / FEMA (public domain), via Wikimedia Commons.

California’s insurance regulator announced on November 13 that it is opening a structured inquiry into how artificial intelligence is being used across the state’s insurance industry. A virtual public discussion workshop is scheduled for December 15, and an insurer survey is due December 30. The California Department of Insurance made clear that this announcement changes no existing policies, coverage, or rates. No regulatory text has been proposed.

This is the information-gathering stage. Its design reveals where the department expects difficult questions to emerge before it decides whether to propose new rules.

Insurance Commissioner Ricardo Lara supplied the organizing principle: “You can outsource technology, but you cannot outsource responsibility.” That puts the insurer, rather than a model developer or software vendor, at the center of the accountability chain.

The decision needs a record

The department says AI and other advanced analytical tools can support claims processing, fraud detection, customer service, underwriting, and operational decisions. Those uses can shape ordinary interactions between an insurer and a policyholder. A consumer may receive an outcome influenced by a model without knowing which data, system, employee, or vendor contributed to it.

The workshop invitation asks how responsibility should be allocated among insurers, boards and senior management, employees, vendors, and other participants in the AI lifecycle. This matters because a complicated vendor chain can make an adverse outcome harder to reconstruct. The insurer may control the decision while relying on data or software supplied by someone else.

A useful decision record would connect five elements: the outcome that affected the consumer, the insurer responsible for it, the data and model or vendor involved, the human review that occurred, and the correction path available if the outcome was wrong. California has not said that every insurer lacks such a record. The survey is meant to produce a clearer picture of current deployments, governance structures, third-party roles, and operational effects.

Human review alone is too vague to serve as a safeguard. A reviewer needs access to the evidence that shaped the outcome and authority to change it. Otherwise, the person may be present in the process without being able to test or correct the system’s work.

Verification bottleneck

  • What moved faster: AI use and experimentation spread across insurance functions before regulators had a common picture of where the systems operate, what oversight surrounds them, and how third-party vendors fit into the decision chain.
  • Who has to verify: Insurers must be able to examine systems they use, including vendor-supplied tools. Regulators need operating evidence from insurers, while the public workshop gathers consumer experience and priorities. These are different evidence streams, and both matter.
  • What to watch next: The workshop is a prenotice discussion, not a formal hearing or proposed regulation. The next concrete signals are published survey findings, a Notice of Proposed Action, enforcement evidence, or proof that complaint and appeal processes can resolve AI-influenced disputes accurately.

Opportunities

An AI decision receipt could give a policyholder and reviewer a practical starting point. It could identify the insurer, the decision date, the business function, whether an automated system materially informed the outcome, the data categories considered, the responsible reviewer, and the route for correction. It would not need to expose source code or every proprietary model detail to make the decision traceable.

An appeal packet could then ask for the specific record behind an outcome rather than forcing a consumer to challenge an opaque process. The department’s workshop explicitly asks how transparency and explainability can be balanced with security, confidentiality, and intellectual property. That creates room for disclosure focused on consequence and evidence instead of a demand for total technical access.

There is also room for a vendor oversight checklist built around the department’s own questions: What information and access does an insurer receive? How does it test performance over time? Who investigates a failure? Who can stop the system? How are corrections recorded and carried into later decisions? Value may appear in helping insurers answer those questions before a regulator or consumer has to ask them after harm occurs.

The current picture would change with concrete evidence. A proposed rule could define duties that do not yet exist. Survey results could show where governance is already strong or where access to vendor systems is weak. Enforcement cases could reveal recurring failures. Complaint-resolution data could show whether existing processes are enough. California is now building the record regulators will use to decide whether safeguards need to change.

Sources

Back to the archive