Driverless trucks are now billing customers on real Texas highways, not just demoing for reporters. Here's what's actually shipped in logistics AI versus what's still a slide deck.
In April 2025, Aurora Innovation put a truck on I-45 between Dallas and Houston with no one in the driver's seat, hauling real freight for real customers. By early 2026 that had become ten routes across the Sun Belt, more than a quarter-million driverless miles, and a stated goal of 200-plus driverless trucks running by the end of the year. That's the detail worth sitting with: after a decade of demos, driverless freight is now a line item on somebody's P&L, not a press event.
That's a useful anchor for a broader question, because "AI in logistics" gets thrown around as if it's one thing. It isn't. Route optimization, demand forecasting, autonomous trucking, and the newer "agentic orchestration" layer that vendors are now pitching are four different technical problems at four different maturity levels, and conflating them is how you end up either over-hyping a forecasting model or under-crediting a genuinely hard robotics win. Here's what's actually running in each.
Dynamic route optimization is the least glamorous and most mature of the four. UPS's ORION system — which blends telematics data with route-sequencing algorithms — has been optimizing driver routes since 2013, well before "AI" was the marketing word of choice, and it's credited with cutting millions of driven miles a year through incremental stop-sequencing improvements. That's the useful baseline: this category didn't start with LLMs, and its wins are measured in single-digit-to-low-double-digit percentage improvements, not revolutions.
What's changed recently is the modeling approach. Classical route optimization used heuristics and mixed-integer programming — solvable, but brittle when constraints multiply (time windows, vehicle capacity, driver hours, real-time traffic, last-minute order changes). Newer systems apply reinforcement learning and graph neural networks, which can hold a much larger, messier constraint set and re-solve as conditions change rather than re-running a batch optimization overnight. Industry estimates put the incremental gain from RL/GNN-based routing at roughly 8-15% lower transportation cost and double-digit improvements in delivery density versus heuristic baselines — directionally credible given the added flexibility, though these numbers mostly come from vendor case studies rather than independently audited benchmarks, so treat them as "plausible upper bound," not gospel.
It's worth separating this from freight visibility platforms like project44 and FourKites, which don't optimize routes — they predict ETAs and flag exceptions (a container missing a vessel connection, a truck stuck at a border) using ML on tracking data. Visibility and optimization get bundled together in vendor pitches because they're bought by the same logistics teams, but they're solving different problems: one tells you where things are and when they'll arrive, the other decides where they should go next.
If there's a category where AI in logistics is unambiguously earning its keep, it's demand forecasting. The mechanism is straightforward: classical forecasting (moving averages, ARIMA-style models) extrapolates from historical sales patterns; ML-based forecasting adds weather, local events, promotional calendars, competitor pricing signals, and cross-SKU cannibalization effects as inputs, which matters enormously for retailers managing tens of thousands of SKUs with volatile, seasonal, or promotion-driven demand.
The reported results are genuinely strong where companies have published specifics. Saint-Gobain Sekurit, a glass manufacturer, reported a 15-percentage-point accuracy improvement in demand forecasting alongside a 9.25% inventory reduction and a 97.2% service level after deploying AI-based forecasting. PepsiCo has attributed part of an $18 billion supply-chain value creation effort to AI-driven simulation of operations. Unilever has used AI forecasting to shorten lead times and improve visibility across its global network. Blue Yonder — the SAP-owned supply chain planning vendor — has built a generative AI assistant (Luminate) on top of its existing forecasting engine, letting planners query forecast drivers in natural language rather than digging through spreadsheets.
The honest caveat: nearly every headline number in this space comes from the vendor or the customer's own press release, not a third party. A 15-30% forecasting accuracy improvement over "legacy models" is a real and repeatable pattern across multiple independent deployments, which is reassuring — but the specific double-digit ROI figures (McKinsey's oft-cited 15-20% logistics cost reduction, Accenture's 23% profitability advantage for forecasting-mature companies) are industry-analyst estimates aggregated across unnamed companies, not audited case studies. Directionally trustworthy, not precise.
This is the category that actually surprised people in 2025-2026, because the field looked dead a few years ago. Waymo and Embark both exited freight. TuSimple collapsed amid governance scandals. The survivors narrowed to a handful of companies that stopped demoing and started billing.
| Company | Model | 2026 status |
|---|---|---|
| Aurora Innovation | Sells the driving software, runs its own trucks | 10 driverless routes across the Sun Belt; 250,000+ driverless miles; targeting 200+ driverless trucks by year-end |
| Kodiak Robotics | Customers own the trucks, Kodiak licenses the driver | ~28 customer-owned driverless vehicles in the Permian Basin; 23,500+ paid driverless operating hours; targeting long-haul highway ops in H2 2026 |
| Waabi | Pure software licensing, no owned fleet | Still pre-driverless-launch; has demonstrated its stack porting from a Peterbilt to a Volvo chassis without retraining |
| Gatik | Short-haul, fixed retail routes (e.g. for Walmart) | First to run fully driverless trucks at commercial scale in North America; 60,000+ orders, no long-haul ambitions |
| Volvo Autonomous Solutions | OEM-integrated autonomous hauling | Targeting fully driverless U.S. highway ops in Q1 2027, 300+ trucks by year-end |
The pattern across all five: every driverless mile currently being driven commercially is on a small number of pre-mapped Sun Belt highway corridors (I-45, I-10, I-20) — flat terrain, minimal snow, and states with permissive regulatory environments for autonomous heavy vehicles. That's not a limitation anyone is hiding; it's the deliberate strategy. Nobody is running driverless trucks through Chicago in a snowstorm, and nobody is claiming they will soon. The FMCSA has also begun updating federal rules to exempt driverless trucks from human-specific requirements like hours-of-service limits, which removes one of the regulatory ambiguities that stalled earlier pilots — but state-by-state rules still vary, and that patchwork is exactly what's kept deployment geographically narrow.
The honest framing: autonomous trucking went from vaporware to a real, revenue-generating business in about eighteen months, which is a genuinely fast maturity curve for a hard robotics problem. It is not, yet, a general highway-driving solution — it's a small number of well-characterized corridors with heavy investment in mapping, weather tolerance, and incident response, expanding one validated lane at a time.
The newest claim in logistics AI is the "super-agent" architecture — a central AI orchestrator that coordinates specialized sub-agents across routing, demand forecasting, carrier contracts, customs documentation, and exception handling as one coherent system, rather than as siloed point tools that don't talk to each other. IBM has been the most vocal proponent of this pattern through its Sterling Supply Chain platform, and the pitch is coherent: a routing engine that doesn't know about carrier contract terms, or a forecasting model that doesn't factor in port congestion, is leaving value on the table by construction.
This is the part of the stack to be most skeptical of right now. The individual capabilities it's meant to orchestrate — routing, forecasting, exception detection — are real and separately validated, as above. The orchestration layer that unifies them into autonomous cross-functional decision-making is much newer, and the specific claim making the rounds — that roughly 15% of daily logistics decisions will soon be made autonomously — is a forward-looking analyst projection, not a measured outcome. It's plausible as a direction; it is not yet a documented result the way the forecasting or trucking numbers above are.
And practitioners closest to the problem are the ones flagging the real constraint. A Walmart last-mile delivery executive put it bluntly: AI excels at demand forecasting and route optimization in isolation, but the actual breakthrough required for cross-functional orchestration is handling the chaos of inconsistent partner data — carriers, 3PLs, customs brokers, and suppliers who don't share a schema, a timezone convention, or even consistent SKU identifiers. An orchestration agent is only as good as the data every sub-agent is fed, and that data-integration problem is decades old and not solved by adding an LLM on top of it.
| Application area | Maturity | Evidence quality | Main constraint |
|---|---|---|---|
| Route optimization | Mature | Strong (10+ year track record, RL/GNN gains directionally verified) | Diminishing returns past heuristic baselines |
| Demand forecasting | Mature-to-emerging | Strong for accuracy gains, weaker for ROI figures | Vendor-reported ROI numbers lack third-party audits |
| Autonomous trucking | Emerging, commercially real | Strong (public mileage/hours data) | Geographically confined to Sun Belt highway corridors |
| Agentic orchestration | Early/nascent | Weak (mostly vendor narrative) | Partner data quality and schema chaos across the supply chain |
The further a logistics-AI claim sits from a specific, auditable number — miles driven, forecast accuracy delta, service level achieved — the more skepticism it deserves. Route optimization and demand forecasting have earned their reputation with a decade-plus of measurable, incremental wins. Autonomous trucking earned a much faster credibility jump in 2025-2026 by shipping real freight on real highways and publishing real mileage counts, even while staying deliberately narrow in scope. The "agentic super-agent" layer being pitched on top of all of it is the part still running on projections and press releases — and the practitioners actually running these networks keep pointing at the same unglamorous blocker: messy partner data, not model capability, is what's standing between logistics AI and the next leap.