AI for Logistics: Anomaly Alerts, Picking, and Routing
Published 2026-09-11 · Updated 2026-09-11 · 8 min read · AI 工作術 (FreeCo Co., Ltd.)
Most warehouses should start AI with shipment anomaly alerts, not routing. Here is the order that pays back, from a team that ships its own orders daily.
Revenue up 30 percent, warehouse headcount up 30 percent, margin exactly where it was. If you run a business that ships physical goods, you have probably lived this. We have too. Alongside our AI tools platform, we ship orders for our own brand every single day, and the uncomfortable truth is that fulfillment cost grows in a straight line with volume unless you change how the work gets done.
When people hear "AI for logistics" they picture robotic arms and lights-out warehouses. Those are real, and they pay off at a scale most of us are not at yet. This article is about the layer underneath: what you can get from software and AI without touching a single shelf, and, more importantly, the order in which to do it. Get the order wrong and you will spend on optimization while money quietly leaks out the back door as returns and reships.
The short version: start with anomaly alerts on shipments in transit, then clean up inventory data, and only then look at optimization projects like slotting or routing. Each step pays for itself, so you never need to bet the whole budget on one big system.
Start with anomaly alerts, not routing
If we could only do one logistics AI project, it would be anomaly detection on shipment status. Not route optimization. Not pick-path optimization. The reason is simple once you look at where the money actually goes. The expensive part of logistics is not the normal order that leaves on time and arrives on time. It is the order that goes wrong: a wrong address that has to be reshipped, a delay that turns into a support ticket and a refund request, a locker or pickup-point package that nobody collected and now comes back to you for a second round of handling. One bad order can eat the margin of five good ones.
Anomaly alerting works like this. You pull the tracking status feeds from every carrier you use into your own system, and you let a model watch for patterns that should not happen. A parcel stuck in the same status for more than 48 hours. A batch of shipments from one day showing a return rate double your baseline. Cash-on-delivery pickup rates dropping in one region. Before, you learned about these things when a customer emailed you. Now the system tells you first, and you fix it before the complaint exists.
One concrete win from our own operation: pickup-point parcels that are about to expire trigger an automatic reminder to the customer. That single rule cut our return rate noticeably, and every avoided return saves two legs of shipping that were pure loss. No machine learning PhD required. The "AI" part is in ranking which anomalies matter and drafting the customer message, not in the reminder itself.
This also connects directly to your support load. When the alert fires before the customer notices, the ticket never gets created. If you are building tiered support, this is the upstream work that makes the bot-handles-most-of-it model realistic. We wrote about that split in AI Customer Support Tiers.
Picking and slotting: clean data before clever algorithms
Here is a number we stand behind from experience: about 90 percent of picking efficiency is process, and maybe 10 percent is algorithm. Where the aisle starts, which row your top 50 SKUs live in, how you cut waves through the day. Classic warehouse management handles all of that without a line of AI code. If your pickers are walking the whole floor for a two-item order, no model will save you. Fix the layout first.
Where AI genuinely helps is as a decision assistant on top of a working process. Mine your order history for items that are frequently bought together and suggest putting them in adjacent bins. Forecast which SKUs will spike during your next campaign and pre-position them near the packing stations two weeks ahead instead of discovering the problem on launch day. These are real gains, but they are second-order gains.
And they only work if your data is honest. If book inventory and physical inventory disagree on a regular basis, every suggestion the model makes is garbage in, garbage out. It will tell you to move stock you do not have into a bin that is already full. We have watched too many teams try to jump straight to "AI optimization" and then discover, after a proper cycle count, that the real project was getting a basic inventory system to reflect reality. That is not a consolation prize. It is the cheapest order of operations. The same pattern shows up in most failed AI rollouts we have seen, and we listed the others in Why AI Projects Fail.
A practical test before you spend anything on slotting AI: run a blind count on 100 random bins. If more than a handful are off, stop and fix that first.
Route optimization only pays if you own the trucks
Route optimization is the textbook logistics AI problem, and we have to be blunt: most e-commerce businesses cannot use it. If your parcels go out through national carriers, regional couriers, or pickup networks, the route is their problem, not yours. You are buying a delivery outcome, and there is nothing for your model to optimize.
The businesses that actually benefit are the ones with their own fleet and a fixed set of stops. A food producer delivering to 40 restaurants every morning. A raw materials supplier running a loop of factories. A B2B distributor with weekly account visits. In those cases the savings are real and easy to measure.
What is interesting is that the biggest saving is often not fuel. In most of these companies, routing lives inside one senior driver's or one dispatcher's head. When that person takes a week off, the schedule falls apart. Once you encode the stops, time windows, vehicle capacities, and driver constraints into a system, automated scheduling removes the key-person risk and frees the dispatcher to handle exceptions. Fuel savings are the bonus, not the headline.
If you are in the outsourced camp, your equivalent lever is carrier selection: which carrier gets which order based on destination, weight, and past performance in that zone. That is a smaller, cheaper decision model, and the anomaly data from step one feeds it directly.
The right order for logistics AI is: make failures visible, make data clean, then optimize. Teams that skip ahead usually pay for the same project twice.
A three-step rollout that pays back at every step
We recommend the same sequence to almost everyone, regardless of size.
Step one: pull every carrier status feed into your own system and build anomaly alerts. This is mostly integration work. Most carriers expose tracking APIs or at least scheduled exports. Define five to ten anomaly rules, wire them to a chat channel or email, and add a model on top only to prioritize and draft responses. You should see the first prevented return within weeks.
Step two: fix inventory accuracy. Cycle counts, barcode scanning on receive and pick, and a single source of truth for stock levels. This is boring. It is also the foundation for everything after it, and it usually surfaces shrinkage and process gaps you did not know you had.
Step three: evaluate optimization based on what you actually operate. Own fleet? Look at routing. Own warehouse with clean data? Look at slotting and demand forecasting. Fully outsourced? Look at carrier selection and packaging. Only now does it make sense to compare vendors or scope a custom build. The subscribe-versus-build question is its own decision, and we laid out the framework in AI Tool Selection.
Each step has its own payback, so you are never asking leadership to fund a two-year program on faith. Every company's order mix and carrier mix is different, so validate on a slice first: one carrier, one warehouse zone, one delivery loop. Then expand.
Three mistakes that waste the budget
Buying the platform before defining the exception. Vendors will happily sell you a logistics intelligence suite. Ask them to show you the exact alert that fires when a parcel sits still for two days and who receives it. If they cannot answer in one sentence, you are buying a dashboard, not a fix.
Letting alerts fail silently. An alerting system that quietly stops receiving carrier data looks identical to a day with zero problems. Monitor the monitor: log how many shipments were checked on each run, and treat zero as an incident. We learned this the hard way on our own platform, and the same discipline applies to any model output you rely on. Our notes on LLM Evaluation cover how to test AI output before you trust it.
Optimizing a process you have not standardized. If three shifts pick three different ways, the model is learning noise. Standardize first, measure for a month, then optimize.
FAQ
Q: What is the best first AI project for a small warehouse?
Anomaly alerts on shipments in transit. It needs no hardware changes, works with the carriers you already use, and prevents the returns and reships that destroy margin. Most teams see results within weeks.
Q: Do I need machine learning for shipment anomaly detection?
Not at first. Simple rules such as "stuck in one status for 48 hours" capture most of the value. Add a model later to rank alerts, spot patterns across carriers, and draft customer messages.
Q: Is route optimization worth it for an e-commerce store?
Usually not. If third-party carriers deliver your parcels, the route is theirs to optimize. Route planning pays off for businesses with their own vehicles and a recurring set of delivery stops.
Q: How clean does inventory data need to be before using AI for slotting?
Run a blind count on a random sample of bins. If accuracy is not in the high nineties, fix counting and scanning first. Slotting suggestions built on wrong stock levels will make things worse.
Q: How much does AI for logistics cost?
It varies widely with carrier integrations, order volume, and whether you subscribe to a tool or build on an API. Check the official pricing page of any vendor you evaluate, and pilot on one slice of operations before committing.