AI Spend Classification in Insurer Expense Ratios: ALAE, ULAE and Administrative Expense

Why AI spend does not map cleanly to expense categories

Insurers report claim-related and operating costs in categories that include allocated loss adjustment expense (ALAE), unallocated loss adjustment expense (ULAE), other underwriting expense and general administrative expense. The definitions and the principles for allocating shared costs are set out in NAIC Statement of Statutory Accounting Principles No. 55 and the Annual Statement instructions. Those texts are the authority for any classification decision. This guide does not restate or replace them.

AI spend complicates the exercise because a single deployment often serves several functions. NAIC reports that insurers use AI across claims adjudication, fraud detection, pricing and plan design, data processing, risk management, and sales and marketing. A shared model endpoint, a vendor platform fee, an engineering team and an oversight function can each touch claims, underwriting and operations at once.

The cost components themselves are familiar: model usage, vendor fees, engineering and oversight. What is new is that none of them arrives labeled with a function. The label has to come from how usage is attributed, which makes the quality of usage records the deciding factor in whether an allocation can be supported.

The factors that drive classification

Three questions do most of the work when finance and actuarial teams decide where AI spend belongs. Each one depends on knowing what the AI actually did.

The first is claim-specific versus general operations. Activity tied to the handling of an identifiable claim is analyzed differently from activity that supports the claims function or the company generally. An AI tool that summarizes documents for a particular claim file raises a different question than a shared model used to improve internal workflows.

The second is direct versus shared. A cost that serves one use case can be assigned to it directly. A cost that serves several needs an allocation basis, and that basis needs drivers that can be measured and reproduced.

The third is claims versus underwriting use. The same vendor contract may support both. Underwriting and pricing uses also carry their own regulatory documentation expectations, so attribution should keep the two from blurring.

The mapping from these factors to a specific statutory category belongs to the written method that finance and actuarial adopt against the authoritative text. The role of the evidence is to show which side of each question a given unit of usage falls on.

Cost components and the attribution question each raises

The table lists common AI cost components and the attribution question each one raises. It does not assign a statutory category. That assignment depends on the authoritative text and the insurer's approved method.

• Model usage charges: Can usage be tied to a use case and function? Evidence: interaction-level records with system, use case and timestamp.
• Vendor platform or subscription fees: Is the fee direct to one use case or shared? Evidence: contract scope, vendor usage reports reconciled to internal records.
• Engineering effort: Which use cases did the work support? Evidence: time or project attribution tied to the use case inventory.
• Oversight and control functions: Is oversight specific to a function or common to all AI? Evidence: documented scope of each oversight activity.
• Tool-call and integration costs: Which downstream systems did the AI act on, and for which function? Evidence: tool-call records with the policy applied.
• Shared infrastructure: What driver apportions it? Evidence: a written basis and the measured driver data behind it.

What examiners and regulators are likely to ask for

The NAIC Model Bulletin on the Use of Artificial Intelligence Systems by Insurers was adopted at the 2023 Fall National Meeting. It expects insurers to maintain a written AIS Program commensurate with risk, documenting the risk management framework and internal controls across the AI system life cycle. It also advises insurers of the documentation a Department of Insurance may request in an investigation or examination, including the AIS Program and its scope. How each state applies the bulletin is a matter for compliance review as of October 2026.

New York DFS Insurance Circular Letter No. 7 (2024) addresses AI systems and external consumer data and information sources in underwriting and pricing. It describes a documented assessment that guidelines are not unfairly or unlawfully discriminatory, and a documented search for a less discriminatory alternative.

Neither document addresses expense allocation. The connection is practical. An examiner reviewing how AI spend was categorized will want to know which systems were in use, for which functions, and under what controls. If expense attribution records disagree with the AIS Program inventory, the inconsistency becomes the finding. Aligning the two sets of records reduces that risk.

Where runtime enforcement produces the evidence

Examiner questions tend to come in two parts: what the AI was permitted to do, and what it did. Records assembled after the fact from application logs and vendor invoices can answer the second part imperfectly and the first part rarely. Records produced where policy is enforced can answer both, because the same event captures the policy applied and the outcome.

Trussed enforces policy in the execution path across models, agents, tools and MCP servers, at the level of individual interactions and tool calls. Each decision to allow, redact, block, rate limit or escalate for human approval is enforced at runtime, not reviewed after the fact, and generates audit-ready evidence. That evidence is mapped to regimes including NAIC and NYDFS. For expense attribution, this means an insurer can retain a time-stamped record of which policy governed each interaction and what action followed. Use-case tagging depends on how systems and applications are mapped to functions in a given deployment, so that mapping should be designed with finance and actuarial.

The records support attribution. They do not decide classification. The category still follows from the authoritative text and the insurer's approved allocation method.

Decisions to resolve with finance, actuarial and compliance

Settle these points in writing before an examiner or auditor asks, and record the approvals and owners.

• Retrieve and apply the current SSAP No. 55 and Annual Statement instruction text to each AI cost component, and record which text supports each position
• Inventory AI cost components and decide for each whether it is direct to a use case or shared
• Define the allocation method: basis, drivers, frequency, and which usage records feed each driver
• Test whether usage can be tied to claim-specific versus general activity, and document the limitation and the alternative basis where it cannot
• Require vendor contracts and reporting to provide usage data sufficient to reconcile invoices to internal attribution
• Assemble the examination evidence set: the method, the usage and enforcement records behind it, and the link to the AIS Program documentation