Four different things are sold under one name
The “AI” label covers functions at very different levels of maturity. Separating them makes proposals easier to read.
| What is sold | What it actually does | How to test it |
|---|---|---|
| Rule engine | Decides from pre-written thresholds (replenish when stock falls below a level). | “Which rule makes this decision, and can I change it?” If yes, this is configuration, not AI — and it is often the better answer. |
| Forecasting / optimisation | Produces workload, demand or slotting suggestions from historical data. | “How much history does it train on, what happens when it misses, do you measure accuracy?” |
| Anomaly detection | Flags unusual activity (unexpected count variance, atypical cancellations). | “What is the false-positive rate, who dismisses an alert, is the dismissal logged?” |
| Language model (assistant, voice) | Asking questions, requesting reports or creating tasks in natural language. | “Where does the model run, does warehouse data leave the premises, what happens when it misunderstands?” |
A function being a rule engine is not a flaw. Most warehouse decisions need to be explainable and repeatable, and a well-written rule is usually the best way to achieve that.
Nine questions to ask the vendor
These come from the AI section of our internal assessment catalogue; we use the same ones when we evaluate vendor profiles.
- On exactly which screen does the AI support come into play?
- Which decision does it make: does it suggest, or does it act on its own?
- Can the user see and override the result, and is the override logged?
- What data does it use: only ours, or other customers' data as well?
- Where is the data processed — in Turkey, and with which provider?
- If it forecasts demand or workload, is accuracy measured, and can you share the figure?
- Can reports be queried in natural language, and how does the user notice a wrong answer?
- If there is voice control, was it tested in a noisy warehouse, and in which languages?
- Is this feature in the standard package, an add-on module, or built per project?
Where the data is processed: the question not to skip
Warehouse data holds more personal data than it appears to: recipient names and addresses, phone numbers, sometimes order contents. If a language model runs in the cloud, that data may leave your premises to be processed, and if the provider sits abroad a cross-border transfer is in question. Treat this as a contract clause, not a technical detail: which provider, which country, is the data used for training, how long is it kept. We covered deployment and data location in <a href="/en/analysis/cloud-or-on-premise-wms">cloud or on-premise</a>.
What AI will not solve
- Wrong master data. No model produces a good forecast for an item with a wrong barcode, dimension or packaging hierarchy.
- A warehouse without locations. With no location data there is no space for an optimiser to work in.
- An undefined process. The system will not decide a matter nobody has decided; it only speeds up an existing decision.
- Unmeasured performance. If accuracy and benefit are not measured, nobody can tell whether the feature works.
With those four in place, AI can produce real benefit. Without them, the first investment belongs in data and process rather than in a model. Use the <a href="/en/wms-finder">WMS Finder</a> to select your requirements and compare source-confirmed capabilities.
Frequently asked questions
What is AI used for in a WMS?
Mainly in four areas: forecasting workload and demand from history, flagging unusual activity (anomaly detection), querying reports in natural language, and voice commands. Some of these are in fact rule engines; to tell the difference, ask “which rule makes this decision, and can I change it?”.
How do I test an “AI-powered” claim?
Four questions are enough: on which screen does it appear, which decision does it make, can the user see and override the result, and is its accuracy measured. If there is no concrete answer to all four, there is no measurable function.
Where does my data go when I use an AI feature?
If a language model runs in the cloud, warehouse data may leave your premises to be processed, and if the provider sits abroad, cross-border transfer of personal data is in question. Which provider, which country, is the data used for training, and how long is it kept — put these in the contract.
Will AI fix bad data?
No. No model produces a good forecast for an item with a wrong barcode or dimension, and with no location data there is no space for an optimiser to work in. The first investment belongs in master data and process, not in a model.
Method and corrections: editorial policy.