Each segment has a materially different duty cycle, telemetry density, and regulatory surface — long-haul trucks are bound by FMCSA Hours-of-Service rules and ELD mandates, transit carries passenger-safety obligations that field-service fleets do not, and yard vehicles operate on private property outside most NHTSA FMVSS scope. One-sizing a single model architecture across segments wastes the highest-signal features in each segment. The usual failure mode is a platform built for long-haul corridors being retrofitted for urban last-mile, where the corridor-density assumption no longer holds.

Different fleet workloads have different latency physics — a collision-avoidance cue at 65 mph is useless past ~45ms, while predictive maintenance can tolerate minute-scale latency. A single SLA across all three models either over-provisions infrastructure for maintenance or under-provisions it for safety. NHTSA 2024 ADAS guidance treats sub-45ms as the ceiling for actionable highway-speed collision cues; any architecture that cannot hit that budget should not advertise collision avoidance as a feature.

SAE J3016 is the governing taxonomy for driving automation and is cited directly in NHTSA and UNECE regulations. The documentation and safety-case burden scales steeply from L2 to L3 — L3 is where ISO 21448 SOTIF and ISO 26262 functional-safety evidence stops being optional. A model that merely informs a driver at L0-L2 has a fundamentally different regulatory posture than one that actuates control at L3+.

ISO 26262 covers functional safety of E/E systems and assigns ASIL levels (A through D) that drive the entire V&V evidence chain. ISO 21448 SOTIF covers the safety of the intended function — specifically the unknown-unsafe space that ML-driven perception creates, which 26262 alone does not address. Fleet intelligence that feeds any advisory or actuation path typically needs both; the combined evidence package is a major work item that must be scoped before model development, not after.

UN Regulation No. 155 mandates a Cyber Security Management System across the vehicle lifecycle, and UN R156 mandates a Software Update Management System — both are type-approval prerequisites in UNECE 1958 Agreement markets including the EU, UK, Japan, and Korea. A connected-vehicle fleet intelligence platform that pushes model updates over the air is squarely in R156 scope. US-only fleets are not directly bound by WP.29 but global OEMs frequently impose the same controls contractually.

3GPP Release 18 and 19 are the 5G-Advanced stream that materially extended C-V2X sidelink capabilities, and C-V2X PC5 (3GPP TS 23.285) is the direct-mode channel that safety-critical V2V/V2I messaging depends on — it does not require the MNO uplink. CBRS Part 96 is the FCC framework under which private LTE/5G networks can be stood up for depot and yard operations without carrier integration. Picking the wrong layer for a given workload is how programs end up trying to deliver safety messaging over best-effort public MNO paths.

Cross-border freight corridors between the US, Canada, and Mexico generate data that straddles privacy regimes in a way many fleet platforms have not reckoned with — PIPEDA in Canada and evolving Mexican data-protection rules do not automatically accept US retention defaults. Public-sector fleets (state DOT, municipal transit, federal GSA) also frequently add sovereignty requirements that exceed commercial baselines. Residency decisions made late are extraordinarily expensive to unwind.

Fleet operators rarely rip out their existing telematics platform to adopt a carrier AI service — integration with incumbents is the minimum viable product. Geotab and Samsara together account for a very large share of North American commercial fleet telematics, and both expose documented APIs. Picking integrations up front clarifies which event schemas, auth models, and retention policies the platform must honor, and which partnerships need to be papered before launch.