How Can AI Improve Route Planning in Logistics?
AI plans your delivery routes against the things that actually constrain them — traffic, delivery windows, vehicle weight and height limits, and driver hours. When the day changes, it re-plans the remaining stops in minutes instead of phone calls. Your dispatcher approves the plan and spends their time on exceptions, not on rebuilding the board.
Manual vs. AI-Assisted Route Planning
| Step | Manual Process | AI-Assisted Process |
|---|---|---|
| Planning the day | Dispatcher builds routes from experience and last week's plan | Every stop is solved against distance, windows and vehicle limits at once |
| Vehicle and road constraints | Held in people's heads, lost when they are on leave | Encoded once — weight limits, bridge heights, access restrictions |
| When something slips | Phone calls, and the rest of the route absorbs the delay | Remaining stops are re-planned across remaining vehicles in minutes |
| Driver hours | Tracked separately, checked after the fact | Treated as a hard constraint the plan cannot break |
| Getting better over time | Planned times stay optimistic because nobody updates them | Actual service times feed back, so tomorrow's plan is more honest |
Prove It in Shadow Mode First
Run the system alongside your current process before anyone acts on it. It plans the same day your dispatcher plans, and nobody drives to its instructions.
After a few weeks you can compare planned distance, planned hours and constraint breaches against what actually happened. That gives you the size of the saving before you have changed how a single driver works — and it surfaces the local knowledge the model is still missing.
Driver Hours Are a Hard Constraint, Not a Cost
In the United States the binding rule is FMCSA's maximum driving time for property-carrying vehicles, 49 CFR 395.3: 10 consecutive hours off duty before driving, a 14-hour window from coming on duty, 11 hours of driving inside that window, no driving once 8 hours of driving time have passed without a consecutive 30-minute interruption, and a ceiling of 60 hours in 7 days or 70 in 8, depending on whether the carrier runs vehicles every day of the week. These belong in the solver as constraints, not in the objective function as penalties.
The subtle part is the input rather than the arithmetic. The binding quantity is hours already used, and that lives in the ELD duty-status record. An optimiser seeded from planned shifts instead of actual ELD state will produce a legal-looking route the driver cannot legally finish — which is worse than no optimiser, because it looks compliant.
In the EU the numbers are different, so one global constraint model is wrong. Article 6 of Regulation (EC) No 561/2006 caps daily driving at 9 hours, extendable to 10 no more than twice a week, 56 hours in a week, and 90 hours across any two consecutive weeks. That last one is the ceiling route planners miss: this week's plan is constrained by last week's actuals, so the optimiser needs a rolling 14-day duty history per driver, pulled from the tachograph, as a first-class input.
49 CFR 395.3 applies to motor carriers operating property-carrying commercial motor vehicles in the United States. Regulation (EC) No 561/2006 applies across the EU and EEA — and, via the AETR, some neighbouring states — to goods vehicles over 3.5 tonnes and passenger vehicles carrying more than nine people including the driver, per its Article 2.
Which Model We'd Shortlist for This
The optimisation itself is a solver, not a language model. What the model does is read constraints out of contracts and emails, and write the messages that go to drivers and customers when the plan changes — so the shortlist is chosen on cost per call and on where inference can physically run.
GPT-5.4-mini — The LLM parses constraints and writes driver comms around a solver, so cost per call beats frontier capability. $0.75/$4.50 across a 400,000-token window with no long-context surcharge.
Gemini 2.5 Flash-Lite — $0.10/$0.40 for per-stop exception messages at dispatch volume, $0.05/$0.20 batched overnight for next-day planning.
Claude Sonnet 5 — A 1,000,000-token flat-rate window lets a full day's manifest and the customer contract terms sit in one prompt, rather than being chunked and re-sent.
Ministral 3 8B — Apache 2.0 at $0.15/$0.15 and small enough to run at the depot or in-cab, where connectivity is the actual constraint rather than price.
Prices are per 1M tokens, as published by each provider. Each model page carries the source and the UTC time it was checked. Rates move — Claude Sonnet 5's introductory rate ends 31 August 2026 — so confirm the current figure before you budget against it.
Where This Fits
Route planning is one part of our work in AI for logistics. It pairs naturally with predictive maintenance, since a vehicle that fails mid-route is the most expensive thing that can happen to a good plan.
Frequently Asked Questions
Our dispatchers know these routes better than any software. Why change?
They probably do, for the routes they run every week. What they cannot do is re-solve the whole board in ninety seconds when three vans are late and a customer moves a window. The aim is not to replace that judgment — it is to hand your dispatcher a strong starting plan and let them spend their expertise on the exceptions.
How is this different from the routing in our TMS?
Most built-in routing solves for distance or time on a static map. It does not know that your 12-tonne truck cannot take that bridge, that this customer only unloads before eleven, or that a driver is close to their hours limit. Those constraints are where the savings actually live, and they are why generic routing gets overridden by hand.
What does it do when the day falls apart?
That is the case worth buying for. A vehicle breaks down or a drop takes forty minutes longer than planned, and the system re-plans the remaining stops across the remaining vehicles. Your dispatcher sees what changed and why, and approves or overrides it.
How long before we know if it works?
Run it in shadow mode first. The system plans the day alongside your existing process and nobody acts on it, so you can compare planned distance, planned time and constraint breaches against what really happened. A few weeks of that tells you the size of the prize before you change how a single driver works.
Will drivers accept it?
Only if the routes make sense on the road. Plans that ignore where you can actually park, or that assume a U-turn a 40-foot trailer cannot make, get quietly ignored — and then the data says the project failed. Getting drivers into the constraint-gathering early is the cheapest thing you can do for adoption.
How does route optimisation handle FMCSA hours-of-service rules?
As hard constraints in the solver, never as a cost to trade off. 49 CFR 395.3 sets 10 consecutive hours off duty before driving, a 14-hour window from coming on duty, 11 hours of driving inside it, no driving once 8 hours of driving time have passed without a consecutive 30-minute interruption, and a 60-hour/7-day or 70-hour/8-day duty ceiling depending on whether the carrier runs vehicles every day of the week. The subtle part is the input rather than the arithmetic: the binding quantity is hours already used, which lives in the ELD duty-status record. An optimiser seeded from planned shifts instead of actual ELD state will hand a driver a legal-looking route they cannot legally finish.
Do the same driver-hours rules apply in the EU?
No, and using one set of numbers for both is a real failure mode. Article 6 of Regulation (EC) No 561/2006 caps daily driving at 9 hours, extendable to 10 not more than twice a week, 56 hours in a week, and 90 hours across any two consecutive weeks. That two-week ceiling is the one route planners miss: this week's plan is constrained by last week's actuals, so the optimiser needs a rolling 14-day duty history per driver, taken from the tachograph, as a first-class input.
Find out what better routing is worth before you change a single route.
Get a Free Proof of Concept within weeks.
Let’s talk
Tell us what you’re
trying to build
The first 45-min alignment session — and a small PoC — are free.
Or just say hello or write us an email.