AI in logistics: what it does well today, and where people stay in charge

How is AI used in logistics?

UPS and Amazon have published figures on their route planning and warehouse AI. For a carrier, broker or warehouse of ordinary size, the useful parts are narrower: reading what customers send, planning what drivers do, and warning you about what is about to go wrong. Each use reads something, does something with it and leaves something for a person:

Route planning and forecasting are older than the current wave of AI: they have run on statistics and planning math for years. What is newer is software that can read loosely written text, such as an email asking for a pickup tomorrow morning, well enough to act on it.

What is an example of AI in logistics?

UPS’s route planning system, ORION, is a well-documented example. It plans each delivery driver’s route, and at its investor day in June 2021, UPS reported:

• The first version: 8 fewer miles per driver per day.
• Dynamic ORION, the newer version: routes re-planned as conditions change during the day, with turn-by-turn directions. Initial results pointed to a further 2 to 4 miles per driver per day, and the rollout was 97% complete, with the rest expected by the end of July 2021.
• The yearly total: more than 130 million fewer miles driven and 10 million gallons of fuel saved each year.
• What one mile is worth: in the question session, UPS said that each mile saved per driver, across all its US drivers over a year, is worth more than $60 million a year to it.

Two things sit behind those numbers. One is scale: a mile per route is multiplied by every UPS driver in the US, so a company with 20 trucks saves the same mile 20 times. The other is data: a route planner works from the addresses, stops and delivery windows a company already records, and it is only as good as those records.

How is AI used in last-mile logistics?

The last mile is the trip from the final depot or warehouse to the customer’s door. AI is used there in four ways:

• Ordering the stops: which stop comes first, given time windows, vehicle size and where the driver starts.
• Re-planning during the day: a stop is added, a customer cancels or a road closes, and the route changes. That is what UPS’s Dynamic ORION added, in a rollout that was 97% complete in June 2021.
• Predicting arrival times: using the driver’s position and past trips to give the customer a delivery window, and to warn dispatch about a stop that will run late.
• Checking proof of delivery: reading a delivery photo or signature against the order, and flagging one that doesn’t match.

The limits show up at the door. A route planner doesn’t know that a dock closes early on Fridays, or that a customer only takes deliveries at the back, until someone records it. Drivers know these things, and a plan that ignores them gets ignored in turn. A planner worth keeping lets drivers and dispatchers correct it, and feeds those corrections back into the records.

What does warehouse AI actually do?

Warehouse AI covers two different things: robots that move goods, and software that decides where goods are stored and in what order they are picked.

Amazon is the big example of the first. On June 30, 2025, Amazon said it had deployed its one millionth robot, in a network of more than 300 facilities worldwide, and launched DeepFleet, a generative AI model built to coordinate how its robots move. Amazon said DeepFleet would improve its robot fleet’s travel time by 10%.

For a regional warehouse, the AI within reach is usually software:

• Slotting: placing fast-moving items near the dock, based on order history.
• Pick paths: ordering a picker’s list so the walk through the aisles is shorter.
• Counts and checks: cameras or scanners that count cartons and flag a short shipment before it leaves.
• Labor planning: forecasting tomorrow’s order volume so the shift has the right number of people.

Each one depends on accurate locations and stock counts in the warehouse management system. If the system says a pallet is in aisle 12 and it is in aisle 14, no model fixes that.

How is AI used in supply chain management?

AI in supply chain work is mostly forecasting and planning, at a scale a spreadsheet can’t handle. Planners use it to:

• Forecast demand: predict orders by product, customer and week, from history and seasons.
• Set inventory: decide how much stock to hold, and where, against that forecast.
• Plan the network and the loads: which lane, which carrier and which truck, and how to move freight around a problem.
• Watch suppliers: read updates on lead times and late shipments, and flag the customer orders they affect.

UPS described its own network planning at the same June 2021 event. Its Network Planning Tools cover demand planning, package flow planning, load planning and movement scheduling, and in February 2021 they helped it route volume around severe weather in the southeast. UPS called this advanced analytics, and said the tools had helped it avoid approximately $250 million of cost in its transportation network from 2019 to mid-2021.

AI for supply chain planning has one limit worth knowing: a forecast is built from the past. When a customer changes its plans or a port closes, the model learns about it after your planner does. The setups that hold up show the forecast to a planner who can question it and override it, and they record why.

Where does AI fit in a logistics office?

Away from the trucks and the warehouse floor, a logistics business runs on documents. Order requests arrive by email, often with a PDF or a photo of a handwritten bill of lading. Rate confirmations, delivery receipts and carrier invoices follow. Someone reads each one and types it into the dispatch or transportation management system.

That reading and retyping is the office work AI for logistics can now take on. A language model can read an email written in a customer’s own words and pull out the pickup address, the delivery window and the freight details. The work around the model is ordinary software: finding the customer’s account, checking the address, looking for a duplicate and posting the order through the system’s API. Three things decide whether it works:

• How varied the orders are: a few customers sending the same form is easy to handle, while many shippers writing in their own words take longer to get right.
• Whether your system has an API: the order has to land in your dispatch system. An API makes that reliable, and a robot that types into the screen breaks when the screen changes.
• What happens to the unclear ones: the system has to know when it isn’t sure, and hand those orders to a person.

What does AI order intake look like in practice?

We run one. AlvyOne runs the operations of Alvy Logistics: we built it with Alvy and run it with them, and DSI Transportation works from the same order system. Alvy is owned by our co-founder Noah Alvarez, and DSI is an Alvy company. Both companies take much of their courier work by email, and from June 30 to November 1, 2026, the intake built 13,554 jobs from those emails in the live dispatch system. Its measured share: 91.4% of new-job emails became orders with no person stepping in, October 3 to November 2, 2026; the other 8.6% were held for a dispatcher. DSI’s monthly overhead went from $22,500 to $10,000: $12,500 a month saved since July 2026. The AlvyOne case study has the full account.

On each new email, the intake works through the same steps:

• It sets aside what isn’t a new order: replies about jobs it has already built, mail from blocked senders, and requests meant for somewhere else.
• It reads attachments as well as the message: pickup and delivery details and the time window may sit in the email itself, an attached PDF or an image.
• It checks before it creates anything: it matches the sender to a customer account, confirms the address and makes sure the job isn’t already in the system, and only then posts the order to the dispatch system through its API.
• It holds what it can’t settle: an unclear order goes to the team’s chat as a card listing what is missing, and a dispatcher answers it or keys the order by hand.
• It leaves the driver to dispatch: its job ends once the order exists, and a dispatcher picks who drives it.

Our page on order entry automation covers this kind of work in more detail.

What are the advantages of AI in logistics?

When it works, the advantages are concrete:

• Fewer miles: better routes cut driving and fuel, as UPS’s figures show at its scale.
• Orders in sooner: an emailed order reaches the system without waiting for someone to get to the inbox.
• Fewer typing errors: the details are copied straight from what the customer wrote.
• Staff time on the exceptions: the hours that went to data entry go to problem shipments and customers.
• Earlier warnings: a late load flagged at noon can still be fixed, and one found at five o’clock often can’t.

It also costs money every month. A system that reads documents pays for each one it reads, it needs hosting and monitoring, and someone has to look at the orders it holds and update it when a customer changes its form. Weigh those costs against the hours it saves before anything is built. The market’s figures for this kind of work are on our AI consulting cost page.

What can go wrong with AI in logistics?

Most failures are ordinary, and each has a plain fix:

• It misreads a document: a blurry photo, a date in an odd format or two addresses on one page. Check what it reads against your own records, and send anything that fails to a person.
• It sounds sure when it’s wrong: language models write fluent answers whether or not they are right. The system should check the facts it can check, such as whether the account exists and the address is real, before it acts on what the model says.
• The data is wrong at the source: a bad address in a customer record gets copied into every order. Clean the records first.
• Something upstream changes: a customer redesigns its order form, a provider retires a model or the dispatch system changes its API. Name the person who notices and fixes it.
• Costs creep up: usage grows with volume, and a fault that re-reads the same email can run up a bill. Set spending limits and alerts from the start.

Each is a reason to keep a person on the exceptions and to watch the system after it goes live.

What is the future of AI in logistics?

Adoption is still early: most US businesses don’t use AI yet. The US Census Bureau’s business survey asks whether a business used AI in any business function in the past two weeks, and from December 2025 to May 2026 between 17% and 20% said yes, with 19.8% as of May 3, 2026. Asked about the next six months, 20% to 23% said they expected to be using it. The bureau’s May 2026 story gives no figure for transportation and warehousing.

The direction is clear enough from the examples above:

• More of the paperwork read by software: orders, rate confirmations, delivery receipts and invoices.
• Plans that change during the day: routes re-planned as stops change, as UPS did with Dynamic ORION.
• People on the judgment calls: fewer hours typing, and more on the shipments that need a decision and the customers who need an answer.

Where should a logistics company start?

Start where the volume and the repetition are, and measure before you build:

1. Count the paperwork: for two weeks, count the order emails, documents and status requests your team handles each day, and how long each one takes.
2. Pick one flow: the one with the most volume and the clearest rules, often order intake from the customers who send the most orders.
3. Check your systems: confirm your dispatch or transportation management system accepts orders through an API, and that your customer and address records are clean.
4. Decide what a person checks: write down which orders the system may post on its own and which go to a person.
5. Run it beside your team first: compare what the system would have entered with what your team entered, before it posts anything live.
6. Watch it every week: track errors, held orders and running cost, and name who fixes it when a customer’s format changes.

The AI readiness checklist goes deeper on the data and systems questions, and what to automate first shows how to rank one flow against another.

If you want help deciding where AI fits in your operation, our page for logistics companies is the place to start, and the first conversation is a 15-minute call.