Routes, stock and the warehouse: where AI actually moves the needle in logistics

Logistics is the sector where AI gets talked about the most and where it is easiest to spend money without changing anything. The reason is that most of what gets sold as AI in logistics is really three very different things bundled into one slide deck: mathematical optimisation that has existed for thirty years, statistical forecasting that depends entirely on the history you hold, and models that read documents or images. All three work, but they need different data, return different value and take different amounts of time to deploy.

This article separates those fronts: what AI contributes to routing, to stock and to the warehouse, what data you need before you start in each case, and how to choose a first project so that a second one ever happens. No invented case studies and no percentage promises that nobody can make without seeing your data.

What problems does AI solve in logistics, and which ones does it not?

Before looking at technology, classify the problem. In logistics almost everything falls into one of four families, and each family has its natural tool:

• Deciding under future uncertainty. How much will I sell, how much freight will I move, how many drivers do I need on Tuesday. This is statistical forecasting and machine learning over time series.
• Picking the best combination out of millions. Which orders go on which truck, in what order, with which driver. This is combinatorial optimisation — not AI in the strict sense, but where a large share of the real savings sits.
• Reading something a person reads today. Delivery notes, CMRs, labels, photos of goods on unloading, incident reports. Document extraction models and computer vision.
• Responding and coordinating. "Where is my order?", delay notifications, rescheduling a failed delivery. Conversational automation wired into the management system.

And what AI does not solve: an item master with wrong weights and volumes, delivery windows a salesperson promises off their own bat, or a warehouse with no defined locations. No model compensates for master data that does not exist. It is the same pattern we describe in why AI pilots fail: the project does not die because of the algorithm, it dies three layers further down.

If correcting the system's proposal takes the planner longer than building it by hand, the system will not get used. However good the model is.

How does route optimisation work, and when is it worth it?

Route optimisation answers a precise question: given N deliveries, M vehicles with their capacity, time windows, service times and driving restrictions, which assignment minimises kilometres or hours? It is a classic, solved problem — today's engines return very good solutions in minutes for hundreds of stops.

Here is the catch: solution quality depends entirely on the quality of the constraints. An engine that does not know a customer on an industrial estate only accepts rigid trucks, that unloading at a shopping centre takes 40 minutes rather than 10, or that a driver cannot run that route because of tachograph rules, produces beautiful plans the traffic manager discards on day one.

It is worth considering when several of these hold:

• More than 15-20 stops per vehicle per day, or an own fleet of meaningful size.
• Routes change daily; if they are fixed weekly, the margin is small.
• There are real time windows or restrictions handled today "in someone's head".
• Someone spends hours every morning piecing the delivery plan together.

It is not worth it with five vans and stable routes, or when the dominant cost is not the kilometre but the failed delivery. In that second case the right project is not routing: it is reducing failed deliveries with prior confirmation and realistic windows.

Where AI proper comes in: predicting the times that feed the optimiser. How long unloading at each customer really takes, by hour, day and order type. An optimiser with eyeballed service times plans 20% over or under and creates deviations nobody can explain. With times learned from GPS traces and delivery notes, the plan starts to resemble reality.

What does AI contribute to stock and replenishment?

This is the most measurable return in the sector, because it touches two things that sit on the balance sheet: stockouts and tied-up inventory. The chain has three pieces and you should not skip any of them:

1. Demand forecasting by item and location, at the real decision horizon (if your supplier takes six weeks, forecasting next week is useless). 2. Stock policy: target service level by family, safety stock calculated from real demand and lead-time variability, not from an inherited number. 3. Purchase proposal that crosses the above with minimum order quantities, pallet multiples and warehouse capacity.

Most projects stop at point 1 and present a nice chart nobody uses. Value appears when the proposal reaches the buyer inside their own tool and they can accept or change it in two clicks. We go into detail in demand forecasting.

One nuance that matters in logistics: not every item deserves a model. Segment the catalogue by volume and variability. High-rotation items with a stable pattern are solved by simple, well-tuned methods; intermittent demand needs specific approaches; and the long tail is managed with rules, because nobody will pay to maintain a model per item.

And inside the warehouse?

The warehouse is where it shows most clearly that useful AI tends to be unspectacular:

• Smart slotting. Relocating items by rotation and order affinity to shorten the picker's walk. It is analytics, not magic, and it pays for itself in warehouses with intensive manual picking.
• Order batching and pick sequencing. Deciding which orders are prepared together and in what order. Pure optimisation, with a direct return in staff hours.
• Computer vision for counting, pallet build checks, reading damaged labels or verifying that what was loaded matches the delivery note. It works well on bounded tasks with controlled lighting; we set out the criteria in computer vision.
• Document extraction from delivery notes, CMRs and carrier invoices, including automatic reconciliation against the contracted rate. Usually the least friction of any project, and we cover it applied in document data extraction.
• Workload forecasting to size shifts: how many units will arrive on Thursday and how many people are needed. It depends on having activity history by hour, not just by day.

The practical rule: inside the warehouse, start with what you already measure. If you have no traceability on how long an order takes to pick, no model will tell you how to improve it.

What data does each use case require?

This is the table we use to work out whether a case is viable today or needs groundwork first. The order matters: a company can be ready for the last row and not for the first.

| Use case | Essential data | Minimum history | Sign you are not ready yet | |---|---|---|---| | Route optimisation | Geocoded addresses, capacities, time windows, service times | No long history needed, but well-loaded constraints | Addresses are fixed by hand every week | | Delivery time prediction | GPS traces or arrival/departure records with real timestamps | 6-12 months | You only have the delivery note date, no time | | Demand forecasting | Sales by item, week and location; stockouts and promotions flagged | 2 years with seasonality | "Did not sell" is indistinguishable from "was not in stock" | | Dynamic safety stock | Real supplier lead times, not the theoretical ones | 12 months of receipts | The master lead time has not been reviewed in years | | Slotting and picking | Order lines with location and timestamp | 6 months | The warehouse has no coded locations | | Computer vision at the dock | Labelled images of the specific case | Hundreds of examples per class | No cameras and no way to label | | Reading delivery notes and CMRs | Digitised documents and a clean supplier master | Weeks | They arrive on paper and get filed unscanned | | "Where is my order?" support | Shipment status available by API and carrier traceability | Immediate | Only the person who calls the carrier knows the status |

If several of the right-hand cells sound familiar, your first project is not an AI project: it is data engineering. That is not a failure, it is the correct sequence.

How do you choose the first project?

Our criterion is not "which is most innovative" but which meets four conditions at once: data available, a concrete decision that changes, an identified owner and a result measurable in weeks. On that basis, a sensible order for most operators and companies running their own logistics:

1. Reading and reconciling transport documents. Low risk, immediate administrative saving, does not touch operations. It also reveals whether you are billed what you contracted. 2. Demand forecasting on the family with the most stockouts or the most tied-up capital. Data is usually available, impact lands on the balance sheet. 3. Route or picking optimisation, depending on where your cost sits: kilometres or warehouse hours. 4. Computer vision or support agents, once the above works and internal confidence exists.

One calendar warning: do not launch a forecasting project in peak season, and do not touch the routing algorithm in your highest-volume week. In logistics the window for change is narrow and the operation never stops.

Realistic timescales: a document project is running in 4-8 weeks; demand forecasting takes 10-14 until the proposal reaches purchasing; routing, 8 to 16 depending on how dirty the addresses and constraints are. Anyone promising all three in a quarter is selling the demo, not the project.

Frequently asked questions

Do I need a TMS or WMS before applying AI in logistics?

Not always, but you do need the data those systems generate. Without a structured record of deliveries, locations and times, the model has nothing to learn from. If you manage on spreadsheets, the first step is usually capturing those events consistently, even with a simple system.

How much history do I need to forecast demand in logistics?

Two full years by item and week if seasonality is pronounced, which in distribution it usually is. One year is enough to start, but the model will not tell an annual pattern from an anomaly. Quantity matters less than having stockouts and promotions flagged: otherwise it learns from censored demand.

Is route optimisation really artificial intelligence?

Strictly speaking it is combinatorial optimisation, a discipline that predates the current AI wave. The learning part appears when you estimate service and transit times from real data. The label matters little; what matters is that the saving is among the most tangible in the sector.

Can I start with a small pilot without touching operations?

Yes, and that is the sensible route. Document projects and forecasts can run in parallel for weeks, comparing what the system proposes with what the person decided, without any of it reaching a truck. When the proposal wins consistently, you switch it on.

AI in logistics is not a single project, it is a sequence: first the data you already generate and do not use, then the repetitive decisions, and only then the eye-catching part. If you want to know where your operation sits in that sequence, in the audit we review your delivery, stock and warehouse data and tell you which case is viable today, which needs groundwork and which does not pay off at your size. And if you would rather talk it through in half an hour before moving anything, let's talk.