_Why a $900B industry running on phone calls and PDFs is finally getting rebuilt — and what the hardest problems actually look like from the inside.
If you want to see what "AI transformation" looks like outside of the SF hype cycle, look at trucking.
The US truckload market moves roughly $900 billion a year of freight, and a huge share of it is coordinated by freight brokers — middlemen who match companies that need to ship things (shippers) with trucking companies that move them (carriers). Brokers now manage roughly 30% of all truckload spend. And until very recently, the entire workflow ran on phone calls, email threads, PDF rate confirmations, and gut feel.
That's changing fast, and the interesting part isn't the "AI will replace truckers" narrative. It's that freight turns out to be a dense cluster of genuinely hard software problems: entity resolution at scale, adversarial fraud detection, real-time telemetry, lead scoring on messy public data, and workflow automation across a dozen legacy systems that were never designed to talk to each other.
Here's a tour of the space, the companies pushing it forward, and the problems that are still wide open.
The fraud problem: adversarial ML in the wild
Start with the most adversarial corner of the industry, because it's the most technically interesting.
Freight fraud — cargo theft, identity theft, double brokering — is estimated to cost the industry around $18 million per day. And it's accelerating. Highway, the Dallas-based company that's become the de facto leader in what it calls Carrier Identity®, publishes a quarterly Freight Fraud Index, and the Q1 2026 numbers are wild: over 527,000 fraudulent inbound emails blocked (up ~50% YoY), 71,800+ spoofed phone calls intercepted, and change-of-ownership fraud up nearly 170% as bad actors buy or hijack legitimate motor carrier authorities to inherit their clean records.
The most uncomfortable stat: roughly half of all theft incidents in Q1 2026 traced back to carriers with legitimate MC numbers and previously clean operating histories. Read that again. The classic vetting playbook — check the MC number, pull the safety score, verify the insurance certificate — is checking credentials that fraudsters now legitimately possess.
This is a textbook adversarial detection problem:
Entity resolution. "Chameleon carriers" get shut down by FMCSA and reopen under a new name with the same trucks, drivers, and phone numbers. Detecting them means fuzzy-matching entities across addresses, VoIP numbers, email domains, insurance agents, and equipment records — classic graph problems with intentionally poisoned data.
Behavioral signals over static credentials. Highway's approach cross-references DOT/MC data, insurance, and behavioral patterns continuously rather than at a single point in time, because a carrier that was safe at onboarding can "break bad" mid-relationship. Point-in-time verification is dead; the model has to be streaming.
Identity proofing. The stack is converging on the same tech as fintech KYC — government ID validation, liveness checks, NIST 800-63A-aligned identity proofing — because document fraud (forged insurance certs, fake licenses) got too easy with generative AI.
AI vs. AI. Organized fraud groups are now using AI themselves to spoof broker identities, manipulate records, and run social-engineering campaigns against after-hours ops teams. It's an arms race, and both sides have GPUs.
GenLogs took a completely different angle on the same problem: physical-world verification. They built a nationwide network of roadside sensors that reads actual trucks on actual highways, then uses AI to reconcile what a carrier claims (lanes, equipment, fleet size) against what their trucks are observably doing. Ground truth as a service. It's one of my favorite architectures in the space because it sidesteps the data-poisoning problem entirely — you can forge a document, but you can't easily fake your truck's physical location across 48 states.
Why vetting suddenly became existential: the legal forcing function
Here's the part most engineering-adjacent readers won't know. In May 2026, the Supreme Court decided Montgomery v. Caribe Transport II, eliminating the federal preemption defense (FAAAA) that brokers had used for years against negligent carrier selection claims. Translation: if a broker puts a load on a carrier that causes a catastrophic accident, the broker can now be sued in state court for how they picked that carrier.
Overnight, carrier vetting went from "operational best practice" to "legal survival." A broker's defense in litigation now hinges on demonstrating a documented, defensible carrier selection process — per carrier, per load. That's a data provenance and audit-trail problem, and it's exactly the kind of forcing function that turns nice-to-have software into infrastructure.
This is the environment the incumbent RMIS category (Risk Management Information Systems — the software layer that handles carrier onboarding, compliance monitoring, and insurance verification) was built for, and why it's being rebuilt with AI. Legacy RMIS platforms were essentially compliance databases: store the W-9, ping the insurance API, flag expired certs. The new generation treats onboarding and carrier vetting as a continuous, ML-driven risk assessment — because the Montgomery standard of care isn't "did you check the box once," it's "did you know what a reasonable broker should have known."
The rest of the AI stack: pricing, visibility, and agents
Fraud is the sharpest edge, but AI is eating every layer of the freight stack:
Pricing and matching. Companies like Greenscreens.ai (now part of Triumph) built ML models for spot-rate prediction that outperform the static rate benchmarks brokers used for decades. Freight Technologies (Fr8Tech) went all-in on agentic AI across its cross-border US-Mexico platform, claiming 15x productivity gains in some domestic workflows and pivoting so hard into software that it's exploring selling its own brokerage.
Visibility. FourKites and project44 turned "where's my truck?" from a phone call into a prediction problem — ETAs, disruption forecasting, and now AI-workflow automation on top of the tracking layer.
Fleet ops. Motive (the ELD/fleet management giant, fresh off a $150M raise) is layering AI onto compliance, safety coaching, and spend management for the carrier side.
Back office. The big freight conversation on Wall Street in early 2026 was literally about AI agents automating brokerage back-office tasks — the announcement of one open-source freight automation tool knocked double-digit percentages off C.H. Robinson and RXO's stock in a single day. The market believes the automation thesis, arguably ahead of the technology.
Autonomy. Waabi raised $200M and opened a purpose-built AV trucking terminal in Texas; Aurora is running driverless nighttime routes. Real, but on a longer clock.
Notice what's missing from that list? Almost all of this AI investment targets the middle and bottom of a broker's funnel — pricing a load they already have, tracking a truck they already booked, automating paperwork for a deal they already won.
The unsolved problem: the top of the funnel
Here's the asymmetry nobody outside the industry appreciates. For a freight broker, carrier supply is inbound — post a load on a load board and your phone rings within minutes. But shipper demand is pure outbound. There is no load board for shippers. There's no marketplace where demand posts itself.
So how do brokers actually find customers? Mostly by cold-calling off stale contact lists bought from data resellers — the same lists their fifty closest competitors bought. Shipper leads for freight brokers are the single most valuable and least solved data product in the industry. The signal exists — customs records, facility data, hiring patterns, permit filings, carrier movement data — but it's scattered, unstructured, and rots fast. It's a lead-scoring and data-freshness problem that the big platforms have largely ignored because they monetize the transaction, not the acquisition.
That's the gap we're building in. FleetGen (fleetgen.ai) started as exactly that: a shipper lead generation engine for freight brokers and brokerage companies — fresh, structured shipper data instead of the recycled reseller lists. Build the pipeline, work the list, book the freight.
But watching early users, we learned the lead is just the entry point into a worse problem: workflow fragmentation. The moment a broker finds a shipper, they context-switch across a CRM, a calendar, an email sequencer, a TMS, and a separate carrier vetting site. Six-plus disconnected tools to run one revenue motion. So FleetGen has been collapsing the stack around the lead:
Pipeline + outreach native. Leads, tasks, calendar, and email campaigns live in one surface, so prospecting and follow-up are the same motion instead of a tab-switching relay race.
Carrier research built in. Before tendering a load, brokers can pull deep carrier data — inspection history, authority status, risk signals — because in a post-Montgomery world, carrier vetting can't live in a separate tool you check "when you have time." The onboarding-and-vetting layer that RMIS platforms and Highway pioneered needs to sit inside the booking workflow, not beside it.
Live ELD tracking. Once the load moves, telemetry replaces check calls.
TMS integration, not replacement. Brokers have enough migration scar tissue. We plug into the existing stack.
The thesis: find the shipper, win the freight, vet the carrier, track the load — one platform. The industry has spent five years building brilliant AI for everything that happens after a broker wins a customer. We think the next wave gets built for everything that happens before.
Why this space is worth watching (even if you've never touched a bill of lading)
Freight is what AI adoption looks like when the stakes are physical. The data is adversarial, the entities are slippery, the legal environment just changed underneath everyone, and the industry's median software vintage is roughly 2009. Every hard problem in applied ML — entity resolution, streaming risk models, adversarial robustness, lead scoring on decaying public data — shows up here with a dollar sign attached.
And unlike a lot of AI application areas, the counterfactual is brutal and measurable: a missed fraud signal is a stolen truckload; a bad carrier pick is now a negligence suit; a stale lead list is a brokerage that never grows.
If you're building in this space or just curious about the data problems, I'm always happy to compare notes. The fraud side alone is wilder than you'd expect.
Building FleetGen — shipper lead generation and broker workspace for freight brokerages. fleetgen.ai
Top comments (0)