The reach into third-party arrangements is the part that reorganizes the work. The NAIC bulletin obliges insurers to govern the AI they procure as well as the AI they build [3], and Holland & Knight's May 2025 analysis reads that as a requirement to design the program so it prevents violations in systems the insurer buys rather than builds [6]. A vendor contract stops being a procurement document and becomes part of the compliance record, which means the operative terms are drafted by people who do not currently sit in the model evaluation meeting.
The footprint arithmetic is simple and incomplete. Twenty-five jurisdictions have adopted the bulletin [1] and four more regulate insurer AI under their own frameworks [2], which is twenty-nine places where a carrier's AI faces a named supervisory expectation [5]. The article does not identify them [24], so no carrier can read its own exposure off that figure; it can only read the direction.
The privacy question resolves along the same line. Nowak argues the constraint is not that personal data must never reach a public model but that consumer accounts generally lack the contractual protections enterprise offerings carry [7], among them commitments that customer data will not be used to train foundation models, that the customer retains ownership, and, depending on product and region, controls over residency and processing [8]. Self-hosting answers the question differently, by putting the runtime and the access list inside the insurer [9]. Either path can satisfy the standard; they differ in how much of the evidence the carrier holds in its own hands.
The trade is worth naming plainly. Self-hosting is available only where weights are licensed for it, and frontier weights are not [10], while frontier models are the ones with a real edge on open-ended work [11]. Nowak's case for accepting that trade rests on the shape of insurance work, where intake, document extraction, coverage checks and claims correspondence recur in recognizable patterns [12]. The harness, the software layer that decides which questions get asked, in what order, with what data and under what rules [13], is what allows a smaller, more precisely defined task to go to a smaller model [14], with a golden set per task type measuring performance on the carrier's own work instead of a public benchmark [15]. He reports having seen the harness matter more than the model it wraps [25], and says a well-designed one makes cheaper models viable for workflows that would otherwise call for a frontier system [26].
A skeptic in underwriting would say that rules around a weaker model still leave you with a weaker model. The design answer in the article is that low confidence is treated as a routing decision: answers are checked against rules and confidence thresholds, and anything below the bar goes to a human adjuster [16]. That is one practitioner's account rather than a measured result, and Nowak is a board member at Decerto, a software supplier to the industry [21], which belongs in the read.
The board-deck version of this quarter is a model selection with evaluation scores attached, and it is incomplete because those scores are calibrated to a version the provider controls. Public services can update model versions, change serving infrastructure or retire older versions, so a model that behaved well on Tuesday can behave differently on Thursday [17], and prompts, decomposition and validation thresholds tuned to one model have to be redone when it moves [19]. Deloitte's 2026 Global Insurance Outlook puts the advantage with insurers that act on their models, tools and strategies rather than those that adopt the newest technology fastest [20]. The decision taken this quarter is which model to license; the invoice that arrives next quarter is recalibration, and it lands on the same engineers who built the harness.
