How AI Prices Commercial General Liability Policies More Accurately
Commercial general liability pricing has worked roughly the same way for decades. Take the insured's SIC or NAICS code, look up the base rate, multiply by revenue or payroll, apply experience modification if available, and add a judgmental adjustment based on the underwriter's assessment of the risk. The result is a price that is often right in the aggregate but wrong for individual accounts, sometimes dramatically so.
The problem is that SIC codes and revenue are crude proxies for the actual risk characteristics that drive GL losses. Two restaurants with identical revenue can have vastly different risk profiles based on factors like their liquor-to-food sales ratio, the condition of their premises, their history of health code violations, and their employee training practices. Traditional pricing treats them the same. AI-based pricing does not.
The Data Advantage
Modern GL pricing models ingest far more data than the traditional rating algorithm. Beyond the standard inputs of class code, revenue, and experience modification, AI models can incorporate business reviews and ratings (which correlate with premises condition and management quality), health and safety inspection results (public records in many jurisdictions), years in business and ownership stability, employee count and turnover rates, website content analysis (which reveals the nature and scope of operations), social media presence and sentiment, litigation history from public court records, and geographic risk factors at the street-address level.
None of these data sources is individually decisive. A restaurant with poor Yelp reviews is not necessarily a bad GL risk. But in combination, these signals create a risk picture that is far more granular than SIC code plus revenue. A model that considers 50 variables will stratify risk more accurately than one that considers 5, assuming the variables are relevant and the model is properly validated.
How the Models Work
Most production GL pricing models use gradient-boosted trees (XGBoost or LightGBM) because these algorithms handle the mix of categorical and continuous variables common in commercial underwriting, and they produce interpretable results that actuaries and regulators can review. The models are trained on historical policy and claims data, learning which combinations of risk characteristics predict higher or lower loss frequency and severity.
The model output is typically a loss cost multiplier that adjusts the base rate up or down for each individual risk. A restaurant with strong safety practices, good reviews, long operating history, and no prior claims might receive a multiplier of 0.7, reducing its rate by 30% from the class average. A similar restaurant with poor reviews, recent health code violations, and a prior slip-and-fall claim might receive a multiplier of 1.4, increasing its rate by 40%.
This individual risk adjustment is where the value lies. Traditional pricing charges both restaurants the same rate (or close to it), meaning the carrier subsidizes the bad risk with premium from the good risk. AI pricing removes this cross-subsidy, making the carrier more competitive on good risks (which they are more likely to win) and more appropriately priced on poor risks (which they are less likely to write at inadequate rates).
Impact on Loss Ratios
Carriers that have implemented AI-based GL pricing report improvements in loss ratio accuracy of 15-20% at the individual account level. This does not necessarily change the aggregate loss ratio in the short term, but it changes the mix of business the carrier writes. Better risk selection means the carrier attracts more of the profitable accounts and fewer of the unprofitable ones.
One mid-market carrier tracked the performance of accounts priced with their AI model versus accounts priced with traditional methods over a three-year period. The AI-priced book had a loss ratio 8 points lower than the traditionally priced book, with the difference driven primarily by fewer large losses. The AI model had identified subtle risk characteristics that correlated with catastrophic GL events (major slip-and-fall injuries, product liability incidents) and priced those accounts accordingly.
Underwriter Adoption
The practical challenge with AI pricing is underwriter adoption. Experienced commercial underwriters have developed intuitions about risk over years of practice, and they are often skeptical of model recommendations that contradict their judgment. An underwriter who has written a particular type of business for 20 years may resist a model that says a risk they consider good is actually overpriced.
Successful implementations address this by showing underwriters the data behind the model's recommendation rather than presenting it as a black box. When the model suggests a higher price for a particular account, the system shows the specific risk factors driving the recommendation: this account has higher-than-average employee turnover, recent negative reviews mentioning cleanliness issues, and is located in a high-litigation-frequency zip code.
Over time, most underwriters develop trust in the model as they see its predictions validated by actual loss experience. The transition period, typically 12-18 months, requires patience and a feedback loop where underwriters can flag concerns that the modeling team investigates and addresses.
Regulatory Considerations
Commercial lines pricing has more regulatory flexibility than personal lines, but there are still constraints. Some states require filed rates that must be actuarially justified. Others allow more pricing flexibility but require that rates not be unfairly discriminatory. Carriers using AI pricing models need to ensure that the variables used in the model are actuarially justified and do not serve as proxies for protected characteristics.
The regulatory landscape is evolving as more carriers adopt AI pricing. Several state departments of insurance have issued guidance or bulletins addressing the use of AI and machine learning in insurance pricing. Insurance carriers navigating this landscape benefit from building model documentation and validation processes that anticipate regulatory scrutiny, rather than treating compliance as an afterthought.