California Legislative Tech News
AI settlement apps may shortchange crash victims - what insurance teams need to know
California logged roughly 384,246 car crashes in 2025 - more than 1,000 a day. As claim volume stays high, carriers are using AI-driven apps to generate instant settlement offers. The speed is real, but so is the risk of undervaluing injuries, wage loss, and future care. With AB 1107 raising minimum BI limits to $30,000/$60,000, the exposure on underpaid claims gets bigger.
"The offers may sound significant at first, often a few thousand dollars, but they probably do not reflect the full value of injuries like neck and back pain that develop over time," said personal injury lawyer Lem Garcia in an interview. He added, "You're not just negotiating with an adjuster anymore - you're up against a program trained to save the insurance company money."
California just moved to regulate AI in health insurance under SB 1120, but auto-injury claimants don't have comparable protections. SB 1120 sets clear guardrails: no final coverage decisions by AI without licensed clinician review, no denials based solely on AI, individual-medical-record grounding, and transparent, auditable tools. AB 3030 also requires disclosures when generative AI is used for patient communication. While these laws target health plans and disability insurers, the principles map cleanly to P&C claims risk management.
Insurers say AI shortens claim cycles. That's true - and it can also push quick settlements that miss late-appearing injuries and ongoing treatment needs. Fast money feels helpful to consumers, but it can increase dispute risk, reopens, and complaints if the offer proves inadequate once symptoms evolve.
Where the risk shows up
Garcia reports some claimants get offers within hours or days of app submission. He often sees people sign away rights after a program labels a crash "minor," even when pain builds later. At the same time, major carriers face lawsuits alleging total-loss software undervalues vehicles by several percentage points - small per claim, material at scale.
Operational guardrails for claims teams
- Human-in-the-loop by design: no AI-generated settlement becomes final without adjuster review and signoff.
- Cooling-off windows on early offers (e.g., 7-14 days) with clear instructions on how to pause, reopen, or supplement a claim as symptoms develop.
- Disclose that AI assists the valuation and that a human can review any decision on request. Offer a simple opt-out to a human-led path.
- Don't pressure for signatures before treatment plans stabilize. If you make an early offer, document assumptions and carve out pathways for additional medicals.
- Flag "minor-impact" classifications for secondary review when there are soft-tissue complaints, prior injuries, or occupation-related wage risk.
- Run fairness and adequacy tests by cohort (age, vehicle type, crash configuration, represented vs. unrepresented) and monitor drift over time.
- Sample and re-score settled files monthly to compare AI valuations vs. medical outcomes and indemnity paid; feed gaps back into training.
- Total loss: audit vendor models, document comparable selections, and keep an appraisal-clause playbook to resolve disputes quickly.
- Keep full model versioning, prompts/inputs used, and adjuster notes to maintain an auditable trail for DOI inquiries or litigation.
- Balance speed KPIs with adequacy KPIs: track reopen rates, supplemental medical payments, complaint ratios, and attorney representation rates post-offer.
Product and compliance checklist
- Publish internal "model cards" that define data sources, intended use, limits, and known failure modes for each claims AI.
- Prohibit sole reliance on AI for denials or final settlement amounts; require documented human rationale for any final decision.
- Give clear consumer disclosures in apps and texts; allow a one-click route to a human adjuster.
- Retain audit logs for inputs, outputs, overrides, and handoffs; set retention aligned to claims statutes.
- Vendor oversight: verify training data provenance, bias testing, and error bounds; require transparency and periodic third-party audits.
- Account for AB 1107's higher BI limits in reserving assumptions and offer authorities, especially on early-settlement workflows.
Field signals an offer may be undervalued (from Garcia)
- Offers arrive immediately after a crash, before diagnostics or treatment plans are complete.
- Pressure to sign quickly, with little room to add medical records or wage documentation.
- Settlements that ignore ongoing pain, therapy needs, injections, surgery risk, or time off work.
What Garcia says about the legal backdrop
"Insurance companies have relied on computer programs to evaluate and settle injury cases for decades. AI simply speeds up that process," Garcia said. "The at-fault driver's insurer does not owe a legal duty to the injured person. If someone believes their injuries are being undervalued, they do not have to accept the offer. Injured people still have the constitutional right to put their case in front of a jury."
Practical next steps
- Stand up a cross-functional review board (claims, SIU, legal, compliance, data science) to vet settlement models before release.
- Throttle instant-offer eligibility to lower-risk profiles; route uncertain files to human-led evaluation.
- Give consumers plain-language explanations of how the offer was calculated and what evidence might change it.
- Measure what matters: fewer reopens and complaints, not just cycle time.
For deeper implementation playbooks on AI deployment in claims and underwriting, see AI for Insurance.
Feature image: Courtesy of photographer Paul Walker.
Your membership also unlocks:
