The conversation around legal AI usually follows a predictable script about BigLaw billable hours, the democratization of small firms, or whether an LLM can pass the bar. These debates focus on the visible front lines—the lawyers and the courtrooms. But focusing there means you're watching the wrong game. The real transformation is happening deeper within the economic engine of civil litigation, driven by the insurance carriers. And on the other side, the plaintiff bar is arming up just as fast.
McKinsey estimates generative AI could unlock $50–70 billion in insurance industry revenue. Bain found that 78% of P&C insurers are already using generative AI—though only 4% have scaled it meaningfully. The arms race is underway. Most litigators just haven't noticed the battlefield has moved.
From Colossus to LLMs
Carriers have used algorithmic case valuation for 30 years. The best-known tool, Colossus, is a rules-based system with over 10,000 decision rules, relying on structured inputs, including ICD codes, CPT codes, and severity ratings. If something wasn't in a form field, the algorithm was blind to it. What many practitioners don't realize is that Colossus is reportedly still used by over 70% of insurers. If you've negotiated a bodily injury claim in the last decade, your demand was likely run through it or a similar tool on the other side of the table.
Colossus generated over $293 million in class action settlements and sustained NAIC scrutiny over the "black box" problem of algorithmic valuation. That history matters, because the next generation of these tools is far more powerful and far less transparent.
LLMs are the leap. They don't need structured fields. They ingest the entire case file, medical records, deposition transcripts, and police reports, and spot nuance a rules engine never could. The gap between what a carrier knows about a case and what a plaintiff's attorney knows has always been a matter of leverage. That gap is narrowing fast. Vendors such as Shift Technology, CLARA Analytics, DigitalOwl, and Wisedocs are deploying LLM-driven analysis at scale across the carrier ecosystem. Meanwhile, carriers from Allstate to Chubb are building proprietary tools internally.
The Data Moat Is Eroding
The carrier's deepest advantage isn't computing, it's context. Carriers sit on millions of closed claims, private settlements, and internal outcomes that never see a public docket. A carrier AI doesn't just know what a jury in Cook County did last week; it knows what the carrier paid to settle 10,000 similar cases over the last decade without a trial. That training data is unique.
But the moat is narrowing. CLARA Analytics operates a contributory database trained on millions of closed claims across its carrier clients. On the plaintiff side, EvenUp, now valued at over $2 billion, has crowdsourced actual settlement data from over 2,000 plaintiff firms processing roughly 10,000 cases per week. The information asymmetry that defined carrier leverage for decades is real, but both sides are now building proprietary data assets. The gap is closing.
When Models Argue With Models
This isn't hypothetical any more. In January 2026, a startup called Mighty launched a platform that acts as an AI agent negotiating personal injury settlements against carrier AI on behalf of consumers. Its CEO stated plainly: the company gives consumers AI to negotiate with the insurance company's AI. This builds on decades of automated dispute resolution. Cybersettle alone has facilitated roughly 200,000 claims totaling $1.4 billion using algorithmic double-blind settlement since the late 1990s.
Now imagine the next step. A plaintiff firm's AI evaluates a case at $850,000 based on crowdsourced settlement data. The carrier's AI, trained on 40 years of internal claims history, pegs it at $320,000. Does a shared analytical baseline strip away posturing and accelerate resolution? Or does it entrench positions because each side treats its own model as truth? We risk moving from a world of legal judgment to a world of model drift, where outcomes depend less on case facts and more on whose training data runs deeper.
Regulators Are Already on the Case
When a carrier's AI determines a claim is worth zero, how does a plaintiff challenge that logic? Regulators have been working on this since at least 2021. As of early 2026, at least 25 states plus D.C. have adopted the NAIC's Model Bulletin on AI, requiring written governance programs, consumer notice when AI affects decisions, and bias testing. Colorado has gone further, SB 21-169 requires quantitative bias testing for AI used in claims handling, with enforcement tools including civil penalties and license revocation. The black box problem is real, but it's an active regulatory battleground, not an open question. Practitioners who don't understand the compliance landscape their opponents operate under are leaving leverage on the table.
Nuisance Value
If carrier AI gets better at early case triage, the economics of "nuisance value" - paying $5,000 to make a weak claim go away rather than litigating - could shift. Claims that used to settle for small sums may face an automated "no." But let's be honest: there is no published empirical evidence that AI triage is currently eroding nuisance settlement patterns. This is a plausible hypothesis, not an observed trend. And the counter-argument has merit, if AI reduces per-claim evaluation costs, carriers might become more willing to pay small amounts quickly, not less. Conversely, if a model flags a case as high-exposure early, carriers have every incentive to settle fast rather than lowball a claim they're likely to lose at trial.
A New Equilibrium?
Law is an adversarial system. When one side upgrades, the other responds. Carriers are deploying AI across claims processing, litigation prediction, and settlement valuation. The plaintiff bar is responding in kind. Contributory databases are eroding data monopolies. Regulators are imposing transparency requirements that may force carriers to show their work in ways they never have before.
The question for litigators isn't whether AI will change how cases are valued. It already is. The question is whether you understand what's in the black box on the other side of the table—and whether you have your own.
