AI Now Builds the Demand Package, and Only One Side Is Regulated
CCC published casualty analytics on August 6 showing that bodily injury claims have passed auto physical damage in total dollars paid, the first time that has happened. BI accounted for 52.3% of the combined dollars across the two categories, up from 44.4% in 2022, a 7.9 point move in three years against a longer-term rate of about half a point per year. It came from both directions at once. BI frequency rose 4% over two years while APD frequency fell 14%, and the average personal auto BI payout rose 21% over the same period to more than $30,000 per feature. Average third-party medical specials, the submitted expenses argued over during evaluation and negotiation, went from roughly $24,300 in the first quarter of 2023 to $32,300 in the first quarter of 2026. That figure is what a demand package is assembled to substantiate.
Erik Bahnsen, who directs casualty industry analytics at CCC and wrote the analysis, places generative AI in that picture as “more of a force multiplier rather than the sole influence.” The mechanism he describes is clerical, not medical. AI lowers the time and effort required to turn medical records, bills, police reports, photographs, and claimant narratives into “highly organized demands,” compressing work that used to occupy paralegals for days into hours. CCC cites a Thomson Reuters survey putting law firm generative AI use at 41% in 2026, up from 28% the year before, and describes the net effect as an increase in claimant-side capacity and information velocity.
Every AI rule an insurer currently answers to points the other way. The NAIC Model Bulletin, as states issue it, sets the department’s expectations for the insurer’s AI systems. Colorado’s testing regime, New York’s Circular Letter No. 7, and the Texas bulletins all reach the insurer’s models. The AI Systems Evaluation Tool asks the insurer to inventory its systems, tier them by risk, and produce the review history. There is no counterpart inside insurance regulation. Whatever discipline applies to the software on the other side comes from the rules governing lawyers, which no insurance examiner reads and no carrier gets to inspect. A demand package assembled in four hours by a tool nobody inventoried arrives at a claims desk where the adjuster’s own assistive software is subject to Exhibit C documentation, human review records, and an examiner who can ask for both.
That changes the arithmetic of a phrase every state uses. Unfair claims settlement practices acts make failing to conduct a reasonable investigation an improper claims practice. The NAIC model they descend from, Model 900, puts no size on that investigation, which in practice means it is measured against the file in front of the adjuster. The file is now longer, faster to produce, more internally consistent, and better indexed. CCC’s own data on treatment mix shows why that matters: lower-complexity interventions such as corticosteroid injections and platelet-rich plasma therapy are appearing earlier, at an average of 122 days from date of loss against 166 days three years ago, and PRP alone averages $11,000 to $12,000 per injured party. Establishing whether a treatment trajectory is documented or scheduled takes real work, and the window to do it before the demand lands has shortened.
Read who is telling you this. CCC’s own product line is casualty AI for carriers, and the analysis closes by arguing that adjusters need “a corresponding ability to process and contextualize information at comparable speed.” The commercial interest is plain, which does not make the observation wrong, and the underlying payout data is CCC’s own transaction record rather than a survey. But the recommended remedy carries a cost the vendor pitch does not price. Answering AI-assembled demands with AI-assisted evaluation puts more of the file inside a system the insurer must inventory, tier, monitor, and defend in a market conduct examination, while the material on the other side of the negotiation stays outside anyone’s scope. Speed on the claimant side is free of governance overhead. Speed on the carrier side is not, and the carrier keeps paying that difference in records long after the claim closes.