Legal History · Bad Faith Analysis

From Colossus to Modern LLMs: Why Automated Claim Deflation Triggers Bad Faith

Thirty years after landmark multi-million-dollar class action settlements and regulatory consent decrees restricted computerized claims software, a new generation of generative AI models is repeating the exact same legal mistakes.

The Institutional Bad Faith Standard

In insurance bad-faith litigation, plaintiff counsel does not simply argue that an individual adjuster made a mistake. Instead, they introduce software manuals, algorithmic tuning directives, and vendor marketing decks to prove an institutional scheme of claim suppression. Juries and courts view predetermined algorithmic haircuts as evidence of intentional bad faith, opening carriers to massive punitive damages.

The Precedent: The Colossus Era and Consent Decrees

In the 1990s, McKinsey & Company famously advised major property and casualty insurers to implement automated bodily injury valuation software (most notably Computer Sciences Corporation's Colossus) to achieve systemic indemnity savings. The strategy was straightforward: tune the software's baseline rules and multiplier tables so that the system consistently recommended settlement values 10% to 25% lower than human adjusters would have historically paid.

The regulatory and litigation fallout was catastrophic:

The Modern Resurgence: How Generative AI Re-creates the Colossus Trap

Today, early-stage insurtech startups are repackaging the exact same concept using modern terminology: “AI settlement prediction,” “algorithmic bill auditing,” “dynamic loss-cost optimization,” and “generative negotiation playbooks.”

When evaluated under state insurance laws and bad-faith jurisprudence, modern AI valuation tools commit the same fatal legal errors:

The Colossus Playbook (1995–2005) The Modern Claims AI Playbook (2024–2026) Statutory Violation (UCSPA / Bad Faith)
Hard-coded multiplier tables designed to cap pain and suffering damages. Neural network embeddings trained on historical lowball settlements to generate “predicted value bands.” Failing to conduct an individualized investigation of the specific claimant’s injury (NAIC Model #900 § 4(D)).
Proprietary geographic benchmarks that slashed medical provider charges without clinical review. Synthetic “regional customary averages” derived from opaque third-party black-box datasets. Failing to provide a prompt, written factual explanation for claim compromise offers (Cal. Ins. Code § 790.03(h)(13)).
Adjuster performance evaluations tied to adherence to software valuation ranges. Automated dashboards tracking “adjuster leakage” when examiners offer settlements above the AI recommendation. Breach of the covenant of good faith and fair dealing; institutional bad-faith scheme (*Egan v. Mutual of Omaha*).

The Deposition Disaster: Why Black-Box AI Collapses in Court

When bad-faith litigation reaches the discovery phase, plaintiff trial lawyers immediately subpoena the technology platform. With a black-box AI model, the defense adjuster is completely defenceless:

Plaintiff Counsel: “Mr. Adjuster, you offered my client $18,000 when her itemized medical bills were $42,000. Exactly which line items did you disallow, and what was the clinical justification?”
Adjuster: “Our AI system processed the records and generated a recommended settlement range of $16,000 to $19,000 based on comparable historical claims.”
Plaintiff Counsel: “Can you show the jury the specific algorithm, the source code, or the training dataset that produced that number? Did you personally verify whether the orthopedic surgeon’s bill was fair and reasonable?”
Adjuster: “No, the vendor's model is proprietary.”

In that single exchange, the insurer concedes that it failed to conduct a reasonable, individualized investigation based upon all available information—the definitive statutory definition of bad faith.

The Defensible Standard: Transparent Provenance and Human Authority

The lesson of thirty years of claims technology litigation is clear: software must never be used to dictate claim valuation.

Compliant claims technology serves as an administrative accelerator: organizing incoming medical packets, extracting itemized lines with clickable source links to original Bates-stamped PDF pages, and highlighting factual discrepancies (such as billing duplicates or preexisting conditions), while leaving 100% of valuation, coverage, and negotiation decisions in the hands of qualified, licensed human claims professionals.