For years, AI in logistics lived on conference slides. Clever demos, a robot or two, and plenty of "imagine if". That phase is over.
In January, UK search interest in "AI in logistics" hit its highest level on record, up 156% on early November, and operators across retail, manufacturing and third-party logistics are moving from pilots to live deployments. Demand forecasting, transport planning, inventory management and automated exception handling are now everyday use cases, alongside autonomous mobile robots, automated storage and retrieval systems and mixed human-robot operations (IT Brief UK).
As one industry executive put it: "conversations have moved from exploration to execution."
The more interesting part of that report is what comes next. The next phase, it suggests, is optimising across multiple sites rather than single facilities. That shift sounds incremental. It is not.
Why one site is the easy part
Optimising a single warehouse is hard engineering, but it is a contained problem. One building, one warehouse management system, one team, one set of rules. You can model it, simulate it and tune it.
A network is a different animal. Stock sits in several sites, often running different systems bought at different times. Transport is planned in one tool, labour in another, customer promises in a third. Each site has its own habits and its own spreadsheets. Optimise one site in isolation and you can easily make the network worse, for example by pulling stock into a fast site that then becomes the bottleneck for everyone else.
The prize for getting this right is large: fewer empty miles, less safety stock sitting in the wrong place, better delivery promises and more resilience when one site has a bad day. But you only get it if the AI can see the whole picture.
The data problem hiding under the AI problem
Multi-site optimisation is, first and foremost, an integration and data project. Before any model can decide where stock should sit or which site should fulfil an order, it needs answers to questions most networks struggle with:
- Is a SKU the same SKU in every site and every system?
- How fresh is the stock position, really: seconds, hours or yesterday?
- Do sites record exceptions, delays and capacity in comparable ways?
If the answer to any of these is "it depends", the model will be optimising against a fiction. That is why the most successful logistics AI programmes spend their first months on master data, event streams from warehouse and transport systems, and a shared model of capacity. It is not glamorous. It is what makes everything after it work.
Exceptions are where the value is
There is a tendency to focus AI on the planned flow: the forecast, the route, the pick path. But in logistics, the money is often lost in the unplanned. The late inbound, the short pick, the vehicle breakdown, the customer who changes their address after dispatch.
Automated exception handling is one of the most valuable uses of AI precisely because it is where human attention is most stretched. At network level it becomes even more powerful: a delay in one site can trigger a re-plan across others before anyone picks up the phone. That needs systems that can talk to each other in near real time, not overnight batch files.
People still run the network
Labour shortages are one of the main drivers behind the current wave of automation. But the operators getting most from AI are not those trying to remove people. They are the ones redesigning roles so that people handle judgement and exceptions while automation handles volume and repetition.
That means involving site teams early, explaining what the system is doing and why, and giving them the ability to override it. A recommendation that nobody trusts is a recommendation that nobody follows.
Where to start
If you have successful single-site automation and are wondering what comes next, a sensible first step is a network data assessment. Map every system that holds stock, order, capacity and transport data. Measure how current and consistent it is. Identify the two or three decisions that, if made at network level, would save the most money.
That gives you a business case built on your data rather than a vendor's brochure. It also tells you, without spin, whether you are ready for network AI, or whether the integration work has to come first.
At Vitalos, Operations, Industry & Logistics is one of our core sectors. We help operators connect the systems they already have and build the data foundations network-level AI depends on. Our article on what UK AI adoption research tells us offers some wider context on where AI is delivering value.
Frequently asked questions
How is AI used in UK logistics?
Common uses include demand forecasting, transport planning, inventory management, automated exception handling, autonomous mobile robots and automated storage and retrieval systems.
What is multi-site optimisation in logistics?
It means making decisions such as stock placement, order fulfilment and transport planning across a whole network of sites rather than optimising each warehouse in isolation.
What do you need before using AI across a logistics network?
Consistent master data, near real-time stock and event data from each site, and integration between warehouse, transport and order systems.