Ask anyone what "AI in transportation" means and they'll say self-driving cars. That's the demo. The deployment — the part actually generating returns today — is operations research wearing a new coat: routing, matching, scheduling. The unglamorous mathematics of moving more people and goods with less waste.
This matters for two reasons. It tells you where to spend. And it tells you what to measure once you've spent.
Autonomy: real, narrow, harder than it looks
Self-driving cars exist and run commercially in a handful of cities under careful constraints. The tech handles most of the driving task. The problem is the tail — freak weather, novel construction, the once-a-decade edge case. That last sliver of reliability is the whole difference between a demo and a deployment, and it's where the time and money disappear.
Highly-automated long-haul trucking is arriving faster than passenger autonomy, for a structural reason: the operating environment is more predictable and the economics are cleaner. Predictability is what makes a problem automatable. Hold that thought — it explains most of what follows.
Where the returns are
This is the part with a payback period you can put in front of a CFO without flinching.
- Delivery routing. Dynamic sequencing against traffic, weather and availability. The old UPS line — roughly ten million gallons of fuel a year saved largely by minimising left turns — is dated now; modern systems do better. The point stands: the saving is a measurable line item, not a vibe.
- Ride-hailing matching. Balancing rider wait, driver income, near-term demand and pricing, continuously. Invisible when it works. It's the reason the whole model works at all.
- Public transport. Redesigning bus routes and frequencies against real demand. Not flashy. Changes the commute for large numbers of people directly.
- Freight and ports. Unload order, container placement, truck assignment. A port handling 800 containers an hour going to 1,200 on the same physical infrastructure — that's pure software return on capital already in the ground.
- Crew and disruption management. When a storm grounds flights, the system picks the cancellations that minimise downstream damage. The value is recovering faster, not dodging the storm.
The common thread: a clear objective function and a quantity you were already losing.
Safety features are not autonomy
Worth separating these, because they get miscounted as "self-driving" constantly.
Advanced driver assistance — lane-keeping, automatic emergency braking, blind-spot monitoring — isn't autonomy. It's a set of good safety features, and the insurance claim data shows they cut incidents. Fatigue detection in trucks has prevented accidents; it's also a surveillance question, and the implementations drivers actually tolerate are the ones built with driver input. Predictive maintenance carries straight over from the factory floor: fewer breakdowns, longer life. Adaptive traffic signals cut congestion and, in some cities, emergency response times.
All real. All a different category of investment with a different justification than the optimisation work above. Don't blend the business cases — that's how a board approves the wrong thing.
The second-order effects to design around
Operations work in transport has public consequences, and they show up after launch if you didn't plan for them:
- Equity. A routing system trained on historical demand reproduces historical neglect — steering capacity away from poorer areas without anyone deciding to. That's a design choice you make by omission unless you make it on purpose.
- Climate. Transport is roughly a fifth of global emissions. Anything cutting vehicle-miles helps; anything inducing more driving — autonomy used to dodge parking costs, say — cuts the other way. Same technology, opposite sign, entirely down to the deployment.
- Workers. These systems change the work of drivers, dispatchers and schedulers. Whether that's better information or just more surveillance and less autonomy is set by how you implement it, not by the algorithm.
None of this is a reason not to build. They're line items for the plan before launch, because retrofitting them costs far more.
What I'd watch
Autonomy is arriving as a swarm of narrow deployments, not one breakthrough — campus shuttles, port trucks, field tractors. Incremental and mostly working. The bigger near-term prize is logistics optimisation at mid-sized companies. The giants already have good systems; the next wave of productivity is in the middle of the market.
If you're sizing a transport or logistics problem, the first question isn't which model. It's: what quantity are you currently losing, and can you measure it before and after? If you can't answer that, you're buying a demo.
