// Commercial auto and fleet
59 quarters of rate have not fixed this line.
Rate has been rising since before telematics was a submission requirement, and the line is still unprofitable. That gap is not a pricing-execution problem. Severity is being set somewhere your submission data cannot see, and the half of the line that is broken is not the half everyone talks about.
This is the line where we are most explicit about what AI cannot do, because on this line that is the useful part.
// The situation
The longest-running underwriting failure in the industry.
59 consecutive quarters of rate increases through Q1 2026, against 14 consecutive unprofitable years.
Rate: CIAB Commercial P/C Market Index, Q1 2026. Loss years: AM Best · 2026 — Conning counts 13 on a different basis, and the disagreement is worth printing.
Two sources counting the same failure differently is not a reason to round it off. It is the reason the first thing we instrument is not price: loss runs get aligned to a common valuation date and the development assumption gets stated out loud, because a five-year loss run priced at face value understates ultimate cost.
// The stack you actually run
The artifacts, the registries and the telematics.
Two things that matter operationally, and both get misread.
Rating territory is driven by garaging ZIP rather than domicile, which makes misstated garaging a classic premium-leakage and rescission issue. And CSA violations carry time weights of 3 for the most recent 6 months, 2 from 6 to 12 months and 1 from 12 to 24 months, so a carrier’s score improves mechanically with time even with no change in behavior. Scores are also percentile ranks against a peer group, which means a fleet’s score can worsen because its peers improved. A broker who understands both facts can time a submission. An underwriter who does not will read a decaying score as a safety improvement.
FMCSA CSA Safety Measurement System methodology, current at 2026.
// The build
Four builds, and what each one is measured on.
MVR normalization across 51 jurisdictions
Violation and conviction codes are not standardized across states. Normalizing them into a comparable severity model is the most mechanical and most valuable extraction task in the line. Annual pulls also miss violations occurring eleven months out, which is the economic argument for continuous monitoring.
Surface: MVR vendor feeds into the driver schedule, with the code mapping held as versioned reference data you own.
Loss run structuring with development awareness
Parse across carrier formats, then handle what breaks comparability: valuation-date alignment, ALAE inclusion, open reserves that will develop, and claimant-versus-claim-level ambiguity. The development assumption is a stated input rather than a number buried in a total.
Surface: submission documents into the pricing workbook and the data warehouse, with the assumption recorded per file.
FMCSA and CSA monitoring with threshold alerting
Pull MCS-150 and SAFER, track the seven BASICs against intervention thresholds — 65% for Unsafe Driving, Hours of Service and Crash Indicator, 80% for Driver Fitness, Controlled Substances, Vehicle Maintenance and hazardous materials, and 50% across most BASICs for hazmat and passenger carriers — and alert on movement rather than level. Percentile movement and peer-group shift report separately, or the alert is noise.
Surface: FMCSA public data into the underwriting workstation and the renewal queue.
Schedule reconciliation against the fleet that exists
Fleet composition changes weekly, so the schedule you underwrote is never the schedule at loss. Bound schedule reconciles against current telematics and MCS-150 powered-unit counts, and drift is reported as an exposure change rather than found later as an audit finding.
Surface: telematics API and FMCSA counts against the bound schedule in policy administration, written back as an endorsement candidate.
What we measure
Units and drivers added mid-term without an endorsement. Share of MVR codes mapped to a comparable severity band versus left unmapped. Days between a violation occurring and it becoming visible in your file. Share of loss runs aligned to a common valuation date with the development assumption stated. Definitions and baselines, set in the first week against your own book.
// The one number on this page
What happened when we ran it ourselves.
Our own operating result from the carrier our founders built and ran, on personal auto. Not a client outcome, and not a forecast of yours. It sits on this page because liability determination is the decision this line turns on.
Clearcover · verified July 2026.
// We ran one of these
We ran auto claims at a carrier where liability determination was a live production system with a cycle-time clock attached rather than a research project — the same decision, on personal lines, with the same evidence problem and the same regulator asking who decided.
The operating record →// What's hard about this
Two limits, and what we do about each.
AI cannot price social inflation.
Severity is being set in courtrooms by noneconomic-damage anchoring, jurisdiction and jury composition, and none of that is in your submission data. In over 80% of trucking verdicts above $1M, pain and suffering ran up to 10 times the actual medical bills, and forum decides the rest — state-court median award above $1M is $3.6M against $2.5M in federal court.
American Transportation Research Institute, 2025.
The results say the same thing from the other side: in 2024 auto liability lost $6.4B at an 87.6 loss ratio while physical damage earned $1.5B, its best ever, a gap of 24.6 points.
AM Best, 2024 statutory results.
Anyone blaming repair-cost inflation has it backward. Severity has been running around 8% a year against roughly 3% economic inflation, and that five-point wedge is the working definition of social inflation. Every ultimate-severity model here is fitted to a series that is structurally non-stationary, which is why 59 quarters of rate have not closed the gap: the industry keeps pricing to a loss trend that keeps moving. What we build instead is fleet-level frequency prediction from telematics, because frequency is observable, sits in the data you already collect, and holds.
The telematics liability paradox.
Collecting behavioral data without acting on it creates litigation exposure, because plaintiffs argue the fleet knew about unsafe driving and ignored it. Telematics is simultaneously an underwriting requirement and a discoverable liability artifact. Underwriters have caught up: installation alone no longer earns a credit, and the question is how operators coach drivers off the data. So the system has to produce the evidence of action taken — which event triggered which coaching, when it was delivered, who delivered it, and what changed in that driver’s later scores — a record that holds up in a deposition rather than a dashboard proving only that data was collected.
// What ships with it
The governance file, plus the record of action taken.
Model inventory entry, data lineage from telematics and MVR vendor to the rating input, pre-deployment testing results, drift thresholds with remediation triggers, and a named human decision-maker specification for every determination the system informs. Two artifacts carry extra weight here: the coaching-action record above, and the separation of percentile movement from peer-group shift in every CSA alert, so a score change is auditable back to its cause. Connecticut requires an annual AI compliance certification attested by a named officer, and Virginia’s bulletin says eliminate the risk rather than mitigate it.
NAIC and state bulletins · current at September 2026.
Bring us one loss run.
We will tell you where the liability evidence goes missing, and what the file needs to survive litigation.