Pylon · The dispatch board for small fleets - Vertical Agent Spec

One-line definition

An agent that assigns incoming loads to available drivers for small fleet operators and dispatchers, monitors trip status through delivery, and surfaces exceptions before they become missed loads.

The workflow it owns end-to-end

A new load arrives via load board scrape, broker email, or manual entry; the agent parses origin, destination, commodity, weight, and pickup window.
The agent checks each driver's current location (via ELD GPS feed), remaining hours of service, CDL class and endorsements, and assigned trailer against the load requirements.
It selects the optimal driver, generates a dispatch notification via SMS or in-app push, and logs the assignment with rate confirmation.
During transit, it monitors GPS check-ins at configurable intervals and alerts the dispatcher if a driver goes dark, runs behind on ETA, or approaches HOS limits mid-run.
At delivery, it captures POD confirmation, closes the load in the board, and queues an invoice draft for the dispatcher to approve before sending to the broker.

What it knows that a generic LLM doesn't

Federal HOS rules (11-hour driving limit, 14-hour on-duty window, 34-hour restart) and how they interact with split-sleeper provisions for team drivers.
CDL class and endorsement matching: Class A dry van vs. tanker endorsement vs. hazmat, and why you cannot assign a Class B driver to a 53-foot semi regardless of availability.
Load board vocabulary and broker communication norms: rate confirmations, TONU (truck ordered not used) claims, detention pay triggers, and what "lumper required" means for driver pay.
Trailer type constraints: reefer temp settings, flatbed tarping requirements, and the distinction between drop-and-hook and live-load appointments.
Dispatcher heuristics for small fleets: dead-head cost tolerance, home-time preferences for owner-operators, and the informal seniority rules that govern who gets the best paying loads.
ELD platform data formats from Samsara, Motive, and KeepTruckin, including how to interpret engine-off vs. personal-conveyance status flags.

What it explicitly declines

DOT authority paperwork, MC number filings, or any interaction with FMCSA systems; these require a human signature and legal accountability.
Driver discipline, termination recommendations, or HR documentation, even if a driver has a pattern of GPS non-compliance or late deliveries.
Insurance claims, accident documentation, or any communication with insurers following an incident.
Rate negotiation with brokers beyond a configurable floor; the agent can flag a low rate and draft a counter, but a human dispatcher approves and sends it.

Tools and integrations required

DAT One or Truckstop.com API for load board data ingestion and lane-rate benchmarking.
Samsara, Motive (KeepTruckin), or Geotab ELD APIs for real-time GPS, HOS data, and vehicle diagnostics.
Twilio for driver SMS notifications and two-way status updates.
QuickBooks Online or Wave for invoice draft generation and broker payment tracking.
Google Maps Platform or HERE Routing API for ETA calculation and dead-head cost estimates.
Gmail or Outlook API for parsing broker load tenders and sending rate confirmations.

Trust escalation: when it pings a human

Any assignment that would push a driver within one hour of their HOS limit before reaching the destination; the agent stops, flags the conflict, and waits for dispatcher override or reassignment.
A driver goes dark for more than 45 minutes during transit with no GPS update and no response to SMS; the agent does not rebook the load or contact the broker without human authorization.
An incoming load rate falls below the dispatcher-configured per-mile floor for that lane; the agent drafts a counter but holds until a human approves the outbound message.
A driver reports a breakdown, accident, or police stop via SMS keyword; the agent immediately routes the alert to the dispatcher and takes no further action on that load until instructed.

Pricing model

The SaaS resistance documented in Pylon's known risks is real and the pricing model has to account for it. A per-dispatched-load fee of $1.50 to $2.50 fits agent economics better than a monthly seat: a five-truck fleet running 40 loads a month pays $60 to $100, which is below the current $120 SaaS price point but with zero subscription anxiety because the fee only hits when the agent does work. The honest concern is that owner-operators will do the math after month two, realize they can get 80 percent of this from the Motive app they already pay for, and cancel. The agent needs to demonstrate measurably fewer missed loads or recovered detention pay to justify the incremental cost. Without that proof point built into the onboarding flow, the churn risk from the SaaS version simply migrates to the per-load version.

Differentiation from a generic LLM wrapper

A dispatcher who pastes their load board screenshot into Claude will get a reasonable answer about which driver might fit, but they will get no live HOS data, no GPS position, no ELD integration, and no audit trail the DOT can inspect. This agent's differentiation is not the language model underneath; it is the data plumbing that feeds the model current driver state at the moment of assignment, combined with decision rules built from real fleet operations that prevent legally problematic assignments from slipping through. The operational moat is shallow if a larger ELD vendor decides to add a chat layer to their existing dispatch bundle, which Samsara and Motive are both positioned to do. The window for a standalone vertical agent here is real but narrow, and it closes faster if the product does not move toward becoming a system of record rather than a dispatch assistant sitting on top of someone else's data.