Industry · Transport & logistics

AI in logistics. From demand forecasting to dock-to-stock.

We build AI applications that fit the breadth of logistics: warehouse, distribution, supply chain, last-mile and planning. Custom where the enterprise suite is too heavy and off-the-shelf software falls short, with the WMS, TMS and ERP integrations that keep your operation running.

DomainWarehouse & DC
DomainSupply chain planning
DomainLast-mile & fulfilment
DomainDocuments & customs

The Dutch logistics sector in numbers.

~14.000
Logistics service providers in the Netherlands
~770.000
Jobs in transport & logistics
~9%
Contribution to Dutch GDP
63%
Of DC managers see AI as a strategic priority

Source: CBS / TLN sector monitor, Logistiek.nl benchmarks.

The enterprise suite doesn't fit every DC.

Anyone serious about deploying AI in logistics soon runs into the same three vendors: Blue Yonder, o9 Solutions and SAP IBP. They are powerful packages, but implementation runs over a long horizon and the price tag suits an operation moving hundreds of millions in goods. For a mid-market 3PL, a specialised DC or a cold-chain operator, that means a suite that never gets fully used, and one that demands more data maturity than an operations team still in build-up can deliver.

Off-the-shelf packages in the WMS and TMS segment do the opposite: they provide an operational foundation, but demand forecasting, slotting optimisation, computer vision at the dock and document AI for customs remain separate modules that nobody integrates well. The result is that planners fall back on Excel, an experienced team leader decides by hand which SKU goes to which dock, and the value of your own order history gathers dust in a data warehouse.

Meanwhile, the operation changes every month. An e-fulfilment peak, a new shipper assortment, a changed customs regime, a switch of carrier for the last mile: each of these is a moment when a static set of rules causes hiccups. This is exactly where AI has a role: not as a replacement for the people who know the operation, but as a boost to the signals they would otherwise only see a week later in a KPI dashboard.

Custom AI solves this differently: a model trained on your own order volume, connected to the WMS feed you already have, with an interface your planners recognise. No second system alongside it, no second version of the truth in the data: a layer that sharpens the existing stack and can be maintained by your own IT team once we step out after the build.

AI applications per logistics domain.

The use cases vary greatly by domain. A DC has different questions from a planner at a 4PL, and cold chain has different requirements again. For each area, we show what we typically build on.

Warehouse & DC

The centre of gravity of many logistics operations. Here, AI delivers direct gains in pick productivity, slotting quality and damage control. A well-trained model on your own order history saves a picker metres per order and measurably reduces the number of handling errors. For 3PLs serving multiple shippers in a single DC, the extra gain lies in cross-dock matching: linking inbound freight to outbound shipments before the pallets ever touch the warehouse floor.

Our warehouse module set includes SKU-level slotting optimisation, pick-path sequencing suggestions that take the physical layout into account, computer vision for damage inspection and volumetric measurement at inbound, and a labour planning forecast based on expected incoming volume. For reverse logistics, we also build classification models that automatically route incoming returns to back-to-stock, refurbishment or destruction — a use case that barely scales with traditional rules once your product range grows.

Without exception, the models run alongside existing WMS systems such as Manhattan, Körber, Boltrics or Centric. We don't write a replacement picking engine; we deliver decision suggestions that your operation can either act on directly or review via a dashboard, depending on your risk appetite and what regulations allow.

  • Slotting optimisationHigh-runners closer to the dock, dead stock moved out of hot zones.
  • Pick-path suggestionsSequencing models that take layout and weight into account.
  • Camera-based damage inspectionComputer vision at the inbound dock, with a photo archive per pallet.
  • Volume forecast for planningIncoming volume predicted, labour planning aligned per shift.
  • Stock anomaly detectionDiscrepancies between WMS stock and counts automatically categorised.

Built on the standards that apply to logistics.

AI in a distribution centre touches labour planning, personal data and, in cold chain or pharma, strictly regulated product flows. From the first sprint, we work within the relevant frameworks.

GDPR

Privacy by design

Camera footage at the dock, pick statistics per employee and labour planning models all touch on personal data. We deliver a DPIA, separate training and production data, anonymise wherever possible and document retention periods per data flow. Camera footage is subject to a strict regime: short retention, detection only and no biometric identification without a valid legal basis.

EU AI Act

Risk classification for workforce AI

AI in labour planning and performance measurement may fall under the high-risk category. We map the classification of each model, deliver a model card covering data preparation, training method and known limitations, and ensure human review of decisions that affect individual employees. We build in the governance layer (versions, datasets, audit logs) from the first sprint, so a later audit doesn't turn into a major catch-up exercise.

CSRD & Scope 3

CO2 reporting from logistics data

For shippers and 3PLs covered by the CSRD, we link the TMS feed to Scope 3 calculations, per shipment, per lane and per mode of transport. Auditable by your external accountant. See also our page on CSRD reporting software.

NIS2

Cybersecurity for critical infrastructure

Logistics service providers in critical chains fall under NIS2. We work with segmentation, logging, vulnerability management and an incident plan that fits your existing security organisation.

EU MDR & GDP

Cold chain and pharmaceutical logistics

For medical device logistics or pharma distribution, we support temperature monitoring, batch tracing and the associated GDP requirements. AI-based anomaly detection complements your existing loggers.

HACCP & food safety

Food-safe flows

For food logistics, we ensure that AI decisions (slotting, pick sequence) are compatible with HACCP zones and allergen segregation. No automatic optimisation that compromises food safety.

Seamlessly connected to your logistics stack.

We integrate with the systems your operation already runs. Our AI layer complements them rather than replacing them. WMS, TMS, ERP and EDI flows remain leading.

SAP / Oracle
ERP feed
Manhattan / Körber
WMS integration
Boltrics / Centric
3PL WMS
Transwide / Transporeon
TMS & carrier
FourKites / project44
Real-time visibility
EDIFACT & EANCOM
Shipper messages
Customs / NCTS
T1, AGS, customs API
AWS / Azure ML
Model hosting

Not ad-hoc integrations, but a repeatable integration layer.

We have standardised the integrations mentioned above across multiple logistics projects. For a new project, we set up the feed using the same abstraction layer, which means fewer bugs at the interfaces, a shorter build phase and a data architecture your own IT team can maintain.

For the heavier cases, such as a specific WMS installation with many customisations or a legacy integration at AS/400 level, we build an adapter case by case. That takes extra time in the planning phase, but prevents a fragile point solution later on. Also read how we approach this for a new custom WMS implementation or a TMS replacement.

On the visibility side, we increasingly see FourKites or project44 used as a source for real-time ETAs and exception signals. Our AI layer picks up those signals and links them to operational actions: rescheduling a dock slot before the truck arrives, moving an outbound shipment to the front of the queue, or proactively informing a customer via the portal. No more standalone alerts, but actions with context.

From data audit to production AI in clear steps.

An AI project in a logistics operation runs differently from a traditional software build. Five phases that will be familiar for most projects.

01 · Data audit

Data volume and quality

A day on site and a week with your data dump. The result: which models are feasible with your own data and where external sources (weather, price elasticity) add extra value.

02 · Use-case scope

Highest leverage first

We choose one use case with a clear business case, usually slotting, demand forecasting or document AI, and define it tightly. No scope creep.

03 · Sprint build

Working model in production

One working component per sprint: model, integration, interface. Planners test alongside us right away, and models are trained on real historical data.

04 · Cutover

Alongside the existing process

First the model runs alongside the existing process and the predictions are compared. Only once quality is proven do we go live as a decision layer.

05 · MLOps

Continuous monitoring

Models go stale. We monitor drift, retrain on new data, adjust for seasonal changes and keep model versions tracked via MLflow. Fixed monthly price for maintenance and further development.

Work in the logistics sector.

DC operation · Brabant

SKU-level slotting AI

Monthly reallocation of pick locations based on demand forecasting per SKU, integrated with an existing Körber WMS. The model takes weight, turnover rate and the physical layout of the pick paths into account.

significant
reduction in pick distance
12.000+
SKUs under management
3PL · Randstad

Document AI for customs

OCR and LLM pipeline for CMR, T1, AWB and certificates. Automatic extraction into the TMS, with human review of exceptions and a feedback loop that lets the model learn from operator corrections.

major
time saved per shipment
4 languages
automatically recognised
Cold chain · Nationwide

Temperature anomaly detection

Real-time model that recognises abnormal temperature patterns before the threshold is exceeded. Logged in line with GDP and linked to an escalation flow to the qualified person in the chain.

earlier
alert than a classic logger
3 parties
in the same chain

What the industry says about AI in logistics.

Logistiek.nl, autumn 2025

"The Dutch logistics sector no longer sees AI as pilot territory. Slotting, demand forecasting and document AI are use cases that mid-market 3PLs are now actively investing in."

TLN sector monitor

"The biggest brake on AI adoption in logistics is not the technology, but the integration with existing WMS and TMS installations. Whoever solves that integration reaps the benefits."

Client conversation, 3PL Randstad

"We rebuilt the document pipeline with a custom solution where Blue Yonder proved too heavy for us. The payback period was within the first financial year."

Answers for the logistics manager getting started with AI.

The questions we hear most from COOs, supply chain managers and DC directors, answered from the experience of our own projects in the logistics sector.

What is the difference between AI in logistics and AI in transport?
AI in transport is mainly about vehicles: route optimisation, telematics, predictive maintenance on trucks and ETA models for trips. AI in logistics is broader and touches the entire goods flow: warehouse optimisation, demand forecasting, inventory allocation, cross-docking, reverse logistics, document AI for customs and labour planning. If you only handle transport, your question is usually TMS-oriented; if you handle logistics, it is typically WMS- and planning-oriented. We have also written a separate page on AI in transport.
Which quick wins can run within a few sprints?
Three use cases keep coming back as a starting point: document AI on incoming freight documents (CMR, AWB, T1), because the business case is clear and the data is already digital; a first demand forecast on your top 100 SKUs, because the historical data there is usually sufficient; and a slotting analysis that shows which SKUs are in the wrong place. None of these three requires major infrastructure changes or extensive data preparation. We integrate directly with the WMS or ERP feed you already have, deliver a first working output within a few sprints and only then decide whether the use case should be rolled out more widely.
Do you replace Blue Yonder, o9 or SAP IBP?
No — for enterprise demand and supply planning at the scale those suites are built for, we refer clients to the right partner. We build custom software for mid-market organisations that find the enterprise suite too heavy, for industry-specific challenges (cold chain, pharma, food with HACCP rules) and for deep integrations between five or more systems that no suite covers out of the box. White-label visibility platforms for 3PLs and customer portals for shippers — with their own branding and their own flow — also fall into this custom segment. We build these from the first sprint.
How do we handle both GDPR and the AI Act in labour planning?
AI in labour planning falls under two regimes at once. For GDPR, we work with data separation, a DPIA and clear retention periods for pick statistics per employee. For the AI Act, we map out the risk classification of the model (labour planning can fall under high-risk) and ensure human review of individual decisions, a model card and transparent explanations for the works council. We do this from sprint 1, not as an afterthought.
How much data do you need for demand forecasting?
It depends on the type of product. For stable SKUs with a seasonal pattern, we work with several years of order history at weekly level. For fast-moving e-commerce assortments we need less history, because we factor in external signals (weather, price elasticity, marketing spend) through demand sensing. For new SKUs we use hierarchical forecasting at product group level until there is enough history on the individual SKU. We work with both classic time-series models (Prophet, NeuralProphet, ARIMA variants) and neural architectures (DeepAR, Temporal Fusion Transformer), depending on data density and the complexity of the assortment.
Computer vision in the warehouse: what is realistic?
Damage inspection at the inbound dock, volumetric measurement for shipping costs, licence plate recognition on site, pallet condition checks and empty location detection. All of these work well in a DC environment with an industrial camera and a trained model. We usually work with PyTorch-based models hosted on AWS or Azure, and feed the results directly into the WMS so the operator gets immediate feedback. What we advise against: vision systems that make production decisions without a human fallback, because the operational cost of an error is too high there. We leave warehouse robot coordination (Locus, Geek+, AutoStore) to the robot vendors themselves; we mainly integrate at the data exchange level.
What does an AI project in logistics cost?
A first, tightly scoped use case (for example document AI on a single type of consignment note, or a demand forecast for your top SKUs) is a compact project with a fixed sprint budget. A DC-wide AI layer with multiple models, MLOps and integrations with WMS, TMS and ERP is a larger project. We always start with a fixed planning phase in which we get the scope clear — only then do we give a concrete price for the complete build, including support in the first period after go-live. Also read how we approach this in custom LLM integrations, which much of our document AI grows out of.
How do you get started in a mid-market DC without a data team?
With one use case and existing data. We carry out the data audit, deliver the first pipeline and set up a dashboard that lets your operational team see the model outputs for themselves. Only once the use case is proven do we broaden the scope. You don't need your own data team to get started — though it helps to have one within a year or two to help shape the expansion. We hand over the models, monitoring and documentation in such a way that a mid-market IT department can take them over without having to build a complete ML stack itself. Also see our page on enterprise AI implementation for the broader context.
Which models and frameworks do you use as standard?
For time-series demand forecasting we work with Prophet, NeuralProphet and — where there is enough data — DeepAR or a Temporal Fusion Transformer. For document AI we use OCR engines combined with an LLM layer for extraction, validation and exception detection. For computer vision we work with PyTorch models on an industrial camera feed. The infrastructure runs on AWS or Azure, with MLflow for model versioning and our own monitoring layer for drift detection. No vendor lock-in: you can always move to your own environment.

Talk to us about AI in your logistics operation?

A half-hour introduction with your supply chain manager or DC director. We listen, ask questions about your stack and flows, and give an initial direction on where AI delivers the most value. No obligation.

Edit content