Your personal-lines AI playbook will get you sanctioned in cyber. The auto and homeowners frameworks that carriers spent the last two years building, in response to the NAIC Model Bulletin and a wave of state adoptions, treat AI governance as a problem with a stable risk surface: well-understood lines, slow-moving loss patterns, mature actuarial baselines, and a regulator whose primary concern is consumer harm. Specialty lines satisfy none of those conditions, and the bulletin's "consumer impact" definition reads differently when applied to a $50 million cyber tower or a transit-leg cargo policy than it does to a private passenger auto rate filing.
The At-Bay 2026 InsurSec Report makes this plain on the cyber side. Akira ransomware claims jumped 364% between Q3 and Q4 2025, with one ransomware crew now driving more than 40% of all ransomware claims on At-Bay's books.1 A cyber underwriting model trained on the first three quarters of 2025 was, by January, working from a feature distribution that no longer described the world. A homeowners model that drifts that fast in a single quarter is a once-in-a-decade event. In cyber, it is the operating tempo. The governance frameworks have to match the tempo.
Specialty has three line-specific AI risk asymmetries that the personal-lines bulletin response does not capture. Each one has a regulator-facing implication that is starting to surface in financial examination requests and reinsurer questionnaires.
Cyber: Threat-Intelligence Models Drift Hourly
A cyber underwriting model is, in practice, a real-time threat intelligence model dressed in actuarial clothing. The exposure features that matter most for pricing are the ones that change fastest: which VPN appliances are exploitable this week, which ransomware affiliate is currently active, which industry is being scanned for a specific vulnerability. The At-Bay data showed VPN-initiated ransomware compromise rising from 38% of incidents two years ago to 73% in 2025, with SonicWall appliances alone appearing in 27% of ransomware claims and 86% of Akira-related attacks.2
The governance asymmetry is that the standard NAIC-aligned governance cadence is built around the assumption that model performance can be reviewed quarterly. A cyber rating model whose underlying threat distribution shifts in weeks does not survive a quarterly review cadence. By the time the review surfaces a drift signal, the model has been pricing against a stale threat surface for two to three months, and the loss ratio degradation is already booked. Cyber requires a continuous model performance evaluation regime, monitored against external threat-intelligence feeds, with the same change-management process applied to model retraining as to a rate filing.
The regulator-facing implication is specific. The Colorado AI Act requires an annual impact assessment; the NAIC Model Bulletin requires documentation of ongoing performance monitoring. For cyber, neither cadence is sufficient as a control. Examiners are starting to ask cyber-line carriers to document the trigger conditions under which a cyber rating model would be retrained mid-cycle, and to produce evidence that those triggers actually fired during the year. A cyber book that has not retrained in twelve months is a finding waiting to be written.
Professional Liability: Judgment-Laden Context Resists Explainability
A directors and officers underwriter is asked to evaluate the risk that a specific board, in a specific industry, with a specific litigation profile, will face a securities class action over the next 18 months. The features that drive the answer are intrinsically contextual: the company's strategic posture, the regulatory environment, the litigation reputation of the lead independent director's prior boards, the quality of the audit committee's prior responses to whistleblower complaints. A model can encode these features as embeddings, and embedding-based models do produce useful decile rankings. They do not produce the kind of feature-attribution explanation that a regulator examining an adverse underwriting decision will accept.
The governance asymmetry is that the explainability tools the personal-lines bulletin response relies on (LIME, SHAP, permutation importance) work cleanly on tabular models with discrete, interpretable features and they break down on contextual embedding models. A SHAP value for "narrative complexity of prior 10-K disclosures" is a number with no operational meaning. The professional liability use of AI is not pricing-dominated, it is information-extraction dominated: the model reads documents, summarizes risk posture, surfaces concerns, and the underwriter prices on top of that synthesis. The audit trail that matters is the document-to-decision trail, not the feature-to-rate trail.
The regulator-facing implication is that a professional liability carrier using AI to summarize submission documents needs to produce, on demand, the source spans the model relied on for any specific decision. This is closer to a discovery exercise than a model risk exercise. A NAIC-style model inventory does not capture it, because the relevant artifact is not the model but the per-decision retrieval and reasoning chain. Marsh, Aon, and Lockton have all begun to ask professional lines underwriters how AI-augmented submissions are documented in the underwriting file, because the broker has a cedent-protection interest in knowing whether the carrier will be able to defend a coverage decision in litigation.
Marine: Geospatial Models Do Not Fail Gracefully
A cargo or hull underwriter using a geospatial model to price transit risk is working with features that look benign on a dashboard: vessel route, port congestion, weather forecast, geopolitical risk score. The model integrates them into a transit-risk premium. The failure mode that specialty governance has not absorbed yet is that geospatial models, particularly those built on satellite-derived training data, do not degrade smoothly. They produce confident, narrowly wrong answers on inputs slightly outside the training distribution: a port that has changed its loading layout since the satellite imagery was last refreshed, a route that was opened by ice conditions the model has never seen, a piracy hotspot that emerged in the months between the model's last update and the booking.
The governance asymmetry is that the model performance metrics personal lines uses (calibration, AUC, mean absolute error against a holdout) treat all errors as roughly equivalent. In marine, an error on a $30 million single-cargo transit through a high-risk corridor is a tail event that no aggregate metric will surface in time. The specialty discipline that does surface them is scenario testing against deliberately out-of-distribution inputs, with explicit human review of any prediction whose underlying input has moved outside a defined spatial or temporal envelope. This is a different shape of supervision than personal lines builds, and most marine carriers using geospatial AI do not yet run it.
The regulator-facing implication is more about the reinsurer than the state department. Marine reinsurance treaty negotiations in 2025 and into 2026 have started to include explicit AI-use disclosures, with the cedent asked to describe whether geospatial or imagery-derived models contribute to acceptance, pricing, or aggregate management decisions. A marine cedent unable to describe its drift-monitoring approach for those models will see treaty terms tighten, regardless of what the state filing said.
There is a fourth specialty pattern worth flagging that has not yet received the same regulatory attention: the use of AI in surety underwriting on construction and infrastructure obligations. Surety models trained on a decade of pre-2022 contractor financial data are working in a market with materially different supply-chain volatility, labor cost structure, and project completion risk. The drift profile is closer to credit underwriting than to property, and the standard model risk frameworks borrowed from banking are a better starting point than the personal-lines AI bulletin response. Specialty groups that pulled their AI governance posture from the auto-and-homeowners playbook are typically applying the wrong reference framework to surety entirely.
What the Reinsurer Is Asking the Specialty Cedent
The pattern across the three lines is that the AI governance question moving fastest is not the regulator's question but the reinsurer's. Property catastrophe reinsurance has long required cedents to disclose the version of the catastrophe model used, the assumed parameters, and the exposure aggregation methodology. The same disclosure regime is now being extended to AI-driven underwriting and pricing decisions in specialty.
Three questions are appearing consistently in 2026 specialty treaty submissions:
What is the cedent's model inventory, segmented by line, and what is the documented retraining cadence and trigger condition for each model? A cyber model with a quarterly retraining cadence and no out-of-cycle trigger is, increasingly, a treaty-pricing input.
What is the cedent's process for detecting and responding to model drift, and what was the longest detection-to-response gap in the prior treaty year? This is the specialty equivalent of asking a property cedent how much of its book is unmodeled. The drift gap is the AI equivalent. The mechanics are described in our analysis of how AI model drift shows up in a loss ratio.
What is the cedent's documentation, per decision, that supports an underwriting outcome where AI contributed materially to acceptance, pricing, or limit-setting? This is the question that pulls explainability out of the policy document and into the per-bound-risk audit trail.
The carriers responding to those questions credibly are the ones that built specialty-specific AI governance infrastructure, tuned to the specific failure modes of each line, rather than retrofitting the personal-lines playbook. The retrofit answers the bulletin. It does not answer the reinsurer.
The Cost of the Wrong Frame
Specialty lines that govern AI under a personal-lines frame will not fail their state market conduct exam in the first year. They will, in the first treaty cycle, fail to defend their assumptions to the reinsurer who has started pricing the model risk into the treaty itself. That is the financial signal, and it is moving faster than the regulatory one. The specialty line that builds AI governance around continuous evaluation, document-level explainability, and out-of-distribution scenario testing is the line whose treaty rates will reflect that work. The specialty line that does not, will discover the cost in the renewal.
