Service · AI development for logistics

AI logistics optimisation. Practical projects, measurable gains.

We design, build and operationalise AI models that optimise your supply chain, from demand forecasting and inventory policy to pick paths, last-mile routing and customs workflows. We don't replace your whole stack; we focus on the flows where most of the gains are.

OptimisationDemand & inventory

OptimisationWarehouse

OptimisationRouting & last mile

OptimisationDocuments & customs

Book a supply chain call See a logistics project

The Dutch logistics sector in figures.

~9,4%

Share of logistics in the Dutch GDP

~770.000

Jobs in transport and logistics

~5.500

Logistics service providers

61%

Cite predictability as the biggest pain point

Source: CBS Transport & Logistics 2025, TLN Sector Monitor, Logistics Digital Index.

Excel forecasts and gut-feel buffers no longer hold up.

Most logistics organisations run on a combination of a good WMS, a sound TMS and, behind the scenes, a handful of Excel files where the real decisions get made. Forecasts are pulled into a spreadsheet, safety stock percentages sit in a forgotten cell, and routes are adjusted by hand by a dispatcher with thirty years of experience.

That works, until volume doubles, the SKU mix becomes more complex, a second DC is added or purchase prices fluctuate again. Then the forecast errors start to grow, inventory costs stay structurally high, pickers lose half-days to walking routes, and last-mile costs rise faster than revenue.

AI models solve this differently than an even thicker spreadsheet. A good model learns from six seasons of historical data, factors in promotions, weather and macro signals, and produces a specific forecast per SKU per DC, not one average line. Our role as an enterprise AI implementation partner is to build that model, validate it and integrate it with your existing WMS, TMS and ERP, so the output reaches the shop floor instead of disappearing into yet another dashboard.

What we don't do: deliver a platform you can never get out of. We work in your cloud tenant, with your data warehouse, and with model artefacts you own. If you no longer need us in a year's time, the model still runs and your own team can take it over. That is a different starting point from a large suite vendor who would rather keep everything bolted together.

We focus on organisations that want optimisation as a standalone service: 3PL and 4PL providers, retailers and e-commerce companies with their own DC, manufacturers with supply chain complexity, fulfilment partners, cold chain logistics and importers/exporters with customs workflows. Our pragmatic approach suits organisations that want concrete gains on one or two flows, not an ambitious multi-year project that tackles everything at once.

Optimisations we deliver as a service.

Each optimisation type has its own model family, data requirements and integration pattern. Below, per discipline, is what we typically build. Most clients start with a single programme and expand once the value has been proven.

Demand Forecasting and Inventory Optimisation

The largest and most rewarding optimisation for most clients. We build ML models for demand forecasting per SKU, per DC and per channel, using seasonal patterns, promotions, and weather and macroeconomic features as inputs. The forecast then passes through a second layer that calculates safety stock, replenishment and recommended reorder points based on your service-level targets.

The output feeds directly into your ERP or procurement tool, with confidence intervals so planners know where the model is certain and where manual adjustment makes sense. We add multi-echelon optimisation if you have several stock layers (central DC, regional hubs and stores).

For the warehouse layer we build slotting models that assign items to locations based on rotation, weight, multi-item orders and seasonal effects, rather than one fixed layout. Pick paths are calculated dynamically for each order and displayed on the handheld. For routing and last-mile delivery we connect an optimisation engine to your transport management system, and can set cost, time and CO₂ as shared objectives. For customs and freight flows we apply document AI: waybills, invoices, T1 documents and commercial invoices are read automatically, checked and passed on to your customs declaration.

Finally, anomaly detection is an underrated but powerful application: a model that continuously scans the event stream from your WMS, TMS and ERP and raises an alert when something deviates, such as a shipment not arriving where expected, a receipt that differs from the purchase order, or a stock count that doesn't match the pick profile. It is not a forecast, but it provides direct signal value for the control tower.

SKU-level forecastPer item, per DC, per channel: no ABC averages.
Dynamic safety stockBased on actual forecast error per SKU, not a single rule.
Multi-echelon allocationCentral DC, regional hubs and stores in one optimisation.
Promotion and weather effectsExternal signals are weighted in, not averaged away.

Built within the frameworks that matter.

We build compliance, data ownership and governance in from sprint one. An AI model that does not pass legal review by your DPO, security officer or accountant is not an option for logistics.

EU AI Act

Risk Classification and Governance

For AI features in planning, labour allocation or picker performance models, we map the risk classification under the AI Act, document the governance and ensure human oversight of decisions where this is mandatory. Pure flow optimisation (routing, slotting, forecasting) usually falls under a lighter regime, but we determine that distinction upfront rather than discovering it later.

GDPR & Data Ownership

Your data remains your data

Models run in your cloud tenant or in our hosted environment under a data processing agreement. No training on client data without explicit consent, and no vendor lock-in on the model artefacts. When an engagement ends, you receive the model code, weights and pipelines so your own team can carry on without us.

ISO 27001 & NIS2

Security in the Pipeline

Data pipelines, model registries and inference endpoints are set up following ISO-compliant processes. For clients in NIS2 essential sectors, we build monitoring and audit logging accordingly. Access to production models runs through your SSO, with an audit trail per inference call if desired.

CSRD & CO₂ Reporting

Optimisation with Sustainability as a KPI

Routing and network models can treat cost and CO₂ as a shared objective. The output is CSRD-reportable, with visibility into the trade-offs between the cost optimum and emissions reduction. For shippers with scope 3 reporting obligations, this is increasingly a hard requirement from their carriers.

MLOps discipline

Models that don't quietly degrade

Drift detection, automatic retraining, A/B tests when replacing a model, and a retraining window that suits your seasonal pattern. A model in production has a lifecycle, not a go-live date. We build dashboards so your data team can continuously see how the model performs compared with previous periods.

Alongside your existing supply chain stack.

We don't replace your WMS, TMS or ERP. We connect our models to what you already run and make sure the output lands where your teams work.

SAP

ERP & IBP

Microsoft

Dynamics & Fabric

Boltrics

WMS for 3PL

Manhattan

WMS and TMS

Transporeon

Carrier platform

PortBase

Rotterdam port community

Snowflake

Data warehouse

Databricks

Lakehouse & MLflow

The model doesn't live in a dashboard, it lives on your shop floor.

A forecast that's only visible in a BI tool is rarely used. We integrate our models directly into the system where the action happens: replenishment proposals in your ERP, optimal pick paths on your WMS handheld, route suggestions in the driver app, and anomaly alerts in the control tower's Teams channel.

Underneath that, we often rely on a data engineering platform that we can build separately or connect to what you already have in place. The combination of a data layer and models is what separates a pilot from a production project. We work equally well with clients who run everything on Azure or AWS and with clients on Google Cloud, and we have patterns for on-premise inference for data that must not leave the organisation.

For shippers and carriers, the broader context of AI in the supply chain is also relevant. Our analysis of AI in transport covers the transport and network perspective, while this page focuses specifically on the service form of logistics optimisation: what you can buy from us as a concrete engagement rather than as a platform licence.

From audit to continuous improvement.

Every optimisation engagement follows the same five phases. No lengthy preliminary reports, just sharp scoping and a working model in your hands quickly.

01 · Audit

Supply chain mapping

A day on site with your operations and data team. Data quality, KPIs, bottlenecks and current tooling mapped out. Outcome: a bottleneck report and an initial optimisation shortlist.

02 · Prioritisation

Use case business case

For each candidate optimisation: data readiness, expected impact and integration effort. We advise which one to start with, not four at once. A one-pager per use case for your steering committee.

03 · Pilot

Greenfield model + benchmark

One flow, one DC, one product line. The model runs alongside your current way of working (shadow mode) and is compared on hard KPIs over several sprints.

04 · Rollout

Production integration

The proven pilot is integrated into your WMS, TMS or ERP. Rolled out in phases per DC or region, with training for planners and operators and a fallback path if something doesn't work.

05 · MLOps

Ongoing improvement

Drift monitoring, retraining, quarterly reviews of KPI development, and expansion to the next use case on the roadmap. Knowledge transfer to your data team where desired.

Work for logistics organisations.

Three types of engagements that show the optimisation approach in practice. Each case starts with the same audit phase and ends with integration into the existing supply chain stack; the optimisation itself differs per type of flow.

E-commerce DC · Demand forecast

SKU-level forecasting for a multi-channel retailer

Replacing an Excel-based flow. Forecasting at SKU, distribution centre and channel level, connected to the existing replenishment engine in the ERP.

Lower

forecast error on A-SKUs

Fewer

expedited orders and stock-outs

3PL · Warehouse optimisation

Slotting and pick-path optimisation

AI-driven slotting by customer segment, plus dynamic pick paths in the WMS handheld for a multi-shipper 3PL site.

Shorter

walking routes per order

Higher

picks per hour

Last mile · Route optimisation

Routing and CO₂-aware planning

A routing model that treats cost and CO₂ as joint objectives, with deviation alerts for planners when actual execution drifts too far from the plan output.

Fewer

empty kilometres

CSRD-

reportable CO₂ output

How the sector sees it.

Logistiek.nl Trend Report 2025

"AI-driven forecasting and multi-echelon optimisation are moving from pilot to production. Specialist builders are gaining ground on classic APS suites among organisations that want to address specific workstreams rather than replace the full stack."

EVO/Fenedex Sector Monitor

"Predictability remains the biggest pain point for shippers; companies that structurally reduce their forecast error gain margin without having to lower purchase prices."

E-commerce DC customer survey

"Appfront's approach is very pragmatic: first one stream, one model, a hard benchmark. Only then a rollout. No hundred-page report before anything is running."

Answers for the supply chain director who is optimising.

The questions we hear most often in a first conversation.

What is the difference from enterprise APS suites such as Blue Yonder, o9 or SAP IBP?

Enterprise suites often replace your entire planning stack and require a multi-year implementation with a large consultancy team. We do the opposite: one concrete optimisation as a service, alongside your existing WMS/TMS/ERP, with a dedicated model that fits your data. For organisations that only want to improve the demand, slotting, routing or customs side rather than the complete S&OP stack, that is cheaper, faster and more flexible. If you genuinely want to replace everything, a suite is still the right choice, and we will be honest and say so. For the broader context, see our analysis of AI in logistics.

How much historical data do we need for demand forecasting?

For a serious seasonally sensitive model, two to three years of transaction data is a sound starting point, as it gives us at least two full seasonal cycles plus a comparison year. Less can work, particularly if external features (weather, macro, promotions) are readily available. For new SKUs we use cold-start techniques that draw on comparable products. The audit phase always begins with a data-readiness check so you know in advance whether the foundations are sound.

Does this work alongside our existing WMS or TMS?

Yes, that is very much the starting point. We connect via the APIs and messaging of your WMS or TMS (Boltrics, Manhattan, Centric, custom middleware, it makes no difference to us) and make sure the model output lands in the right place: replenishment proposals in the ERP, pick paths on the handheld, route suggestions in the driver app. For clients who also want to modernise or build their WMS, we refer to our WMS warehouse software page, but that is a separate project.

What are the quick wins that can go live within a few sprints?

In practice we see three common quick wins: a document AI flow for customs documents or consignment notes (quicker to set up because the data requirements are more limited), anomaly detection in supply chain events (no transaction-data model training needed, only log streams), and pick-path optimisation within a single site where the WMS data is already clean. These are pragmatic pilots that deliver a working result within a few sprints. Demand forecasting and multi-echelon optimisation are more valuable but require a longer, multi-sprint process, particularly on the data side.

What does the EU AI Act mean for workforce planning or performance models?

Models that directly influence work allocation, staff scheduling or performance assessment fall into a higher risk category under the AI Act. That means mandatory risk classification, documentation, human review of critical decisions and transparency towards those affected. We map this out from the design stage, deliver the governance documentation and build in human-in-the-loop where it is required. For pure optimisation of physical flows (routing, slotting, forecasting), the risk category is usually lower, and lighter requirements apply.

What does such a project cost?

An initial audit and use-case prioritisation is a defined preliminary phase with a fixed budget. The pilot that follows is scope-driven: one flow, one model, a hard benchmark. Only once the pilot results are in do we decide together whether and how the rollout goes ahead. We work with fixed sprint budgets and provide a concrete quote once the scope is clear. A non-binding conversation about your situation is always free; costs only begin once we actually start the work.

How does the scope differ between a pilot and a full rollout?

In the pilot we deliberately keep the scope small: one DC, one product line or one route segment, with a clear yardstick. The aim is to prove that the model delivers value in your context, not just on a clean test dataset. The rollout phase then covers integration with production systems, change management for planners and operators, and deployment to the other DCs or flows. Many clients choose to run a second pilot in another part of the chain after a successful first pilot before rolling out more broadly. We are happy to follow that sequence.

What is your approach to data ownership and exit?

All models, weights, code, pipelines and documentation belong to you. When an engagement ends, you receive the artefacts in a form that your own team or another party can continue with. We work in your cloud tenant or in a hosted environment under a data processing agreement, not in a platform-as-a-service set-up where you cannot continue without us. That may sound obvious, but in this market it is more the exception than the rule.

Do you also offer general AI strategy, or only build work?

Both. For organisations still early in their AI journey, it often starts with strategy, prioritisation and a roadmap, which is a shorter and more conceptual engagement. Our enterprise AI implementation page goes into more depth on that side. For logistics organisations, we often find the strategy phase can be more compact than in other sectors: the valuable use cases (forecasting, slotting, routing) are fairly well known, so we can move quickly on to prioritisation and the first pilot.

Talk to us about your AI logistics optimisation?

A half-hour conversation with your supply chain director or operations manager. We listen, ask questions about your flows and data, and give an initial direction on where the value lies. Non-binding and without a sales pitch.

Schedule an introductory call Email fabian.vandijk@appfront.nl