AI in Logistics: How Machine Learning Unclogged a Distribution Centre in Five Months

Series post 4 of the AI in Lean Six Sigma series

A regional distribution centre processes 120,000 orders a month. Its on-time delivery rate is 88.4%, average order cycle time is 2.3 days, and dock-to-stock time is 6.5 hours. Pickers walk an average of 9.4 kilometres per shift.

The team has already completed three consecutive value stream mapping (VSM) projects. Each produced some improvement, but the gains were marginal. The obvious response might be to ask pickers to move faster or to add labour. The more useful question is: what is creating all that travel and delay in the first place?

In this illustrative case, a machine-learning analysis of order history changes the team’s view. The primary cost driver is not picking speed. It is slotting logic: high-frequency stock-keeping units (SKUs) are positioned in distant aisles, while products often ordered together are spread across the warehouse.

That distinction matters. Lean Six Sigma helps teams define the performance gap, measure the current process, and test its causes. AI can add a layer of visibility by identifying patterns across large, complex datasets. But the method still depends on sound data, practical constraints and frontline knowledge.

AI in logistics warehouse optimization: start with the process

Warehouse optimization is not a standalone modelling exercise. It connects the flow of orders, inventory, labour, equipment and information from receipt through dispatch. A local improvement can create a new constraint elsewhere: moving fast-picking SKUs nearer to dispatch, for example, may increase replenishment traffic or congestion.

That is why AI in logistics warehouse optimization works best within a structured improvement approach. The team should first define customer-critical requirements and operational measures, then use analysis to guide experiments, not to bypass process knowledge.

The figures in this article describe an illustrative worked example, using the scenario’s supplied baseline and outcomes. They are not presented as an independently verified client case or as a guaranteed result. The research references below provide separate examples of warehouse optimisation methods and reported outcomes.

The logistics AI toolkit, mapped to DMAIC

DMAIC (Define, Measure, Analyse, Improve and Control) provides a practical structure for an AI-assisted improvement project. It helps keep the work grounded in the problem that matters, rather than the model that happens to be available.

Define: agree what good performance means

The distribution centre should set its objective around customer and business needs. That might mean improving on-time delivery without increasing mispicks, safety exposure or inventory costs. The team can translate those needs into operational measures such as:

• On-time delivery rate and order cycle time
• Dock-to-stock time and picks per labour hour
• Walking distance per shift
• Mispick rate, replenishment touches and labour cost per order

The project charter should also establish a baseline period, scope, data owners and decision rights. For example, does the project cover one site, one process area or all outbound waves? Who can approve a slot move? Which constraints (such as product compatibility, rack dimensions and replenishment frequency) must not be compromised?

Clear definitions make the later analysis more useful. “Order cycle time” needs agreed start and end events. “On-time” needs a defined commitment date and rules for cancelled or split orders. Without those operational definitions, a dashboard can look precise while comparing inconsistent measures.

Measure: capture the actual flow

The Measure phase builds a reliable picture of what happens, where and when. In a digitally enabled warehouse, that can combine:

• RFID and barcode scans to timestamp receipt, put-away, pick and dispatch events.
• Warehouse management system (WMS) telemetry to show order releases, task completion, inventory locations and replenishment activity.
• Forklift geolocation to understand travel, waiting and congestion patterns.
• Labour, calendar and promotion data to provide context for daily and seasonal variation.

The team should check that clocks are aligned across systems and that event definitions are consistent. It should also verify whether location data represents actual movement or only task assignments. Missing scans, shared devices and offline activity can create misleading gaps.

A VSM remains valuable here. It makes material and information flow visible, helps the team locate queues and hand-offs, and provides an operational frame for the data. AI does not make that map redundant; it can help investigate patterns the initial map cannot quantify on its own.

Analyse: find the drivers behind the delay

In the scenario, the team uses 24 months of order data to examine pick frequency, SKU location, order combinations, replenishment and travel. At a steady 120,000 orders per month, that history would contain approximately 2.88 million orders: enough to analyse recurring patterns, provided the data is reliable and the model accounts for changes over time.

The analysis identifies 61% of pick-path waste as attributable to slotting: high-frequency SKUs are disproportionately located in far aisles. Association-rule analysis also detects products that frequently appear in the same orders. If two items regularly travel together, placing them in convenient positions relative to each other may reduce unnecessary movement.

These are not automatic instructions to move every popular SKU. The team should test the finding against factors such as:

• SKU dimensions, weight, fragility and storage requirements
• Replenishment frequency and the capacity of pick faces
• Rack and aisle constraints, fire and safety rules
• Congestion, equipment access and the needs of other workflows
• Whether demand patterns vary by season, promotion or customer group

A model can identify a relationship; process analysis helps determine whether it is actionable and what else might change as a result.

Improve: simulate and test the new design

Once the team understands the likely causes, it can evaluate potential changes. A slotting optimisation model might use demand forecasts, co-occurrence, travel distance, replenishment needs and physical constraints to propose alternative locations. Depending on the problem, the method might use mathematical optimisation, a genetic algorithm or reinforcement learning.

Reinforcement learning is one option when decisions are sequential and conditions change over time. It can learn policies for assigning or replenishing locations under specified objectives and constraints. It is not necessarily the right choice for every warehouse: simpler, transparent optimisation may be easier to validate and maintain for a periodic slotting review.

Before physically moving stock, the team can use a digital twin, a simulation of warehouse operations, to compare candidate layouts. The simulation can replay historical order waves and test scenarios such as peak demand, labour shortages, equipment downtime and congestion. This helps expose trade-offs before they disrupt live operations.

The team can also optimise outbound routes and wave sequencing. That work should be evaluated alongside slotting: a better path plan may produce limited value if the underlying locations make every route unnecessarily long.

Control: keep the improvement working

A successful pilot needs an ongoing control plan. In the scenario, the team uses predictive analytics to forecast dock congestion 90 minutes ahead, anomaly scoring to flag unusual cycle times by zone, and agreed thresholds to prompt review.

A practical control plan specifies:

1. What to monitor: service, cycle time, travel, productivity, mispicks and replenishment.
2. Who responds: the shift leader, inventory controller, process owner or data analyst.
3. What action follows: investigate the zone, verify the data, adjust the wave or review slotting.
4. When to escalate: for example, if a service measure crosses a pre-agreed threshold or a trend persists across shifts.

The objective is not to let an algorithm make every operational decision. It is to alert the right people early, give them credible evidence and make corrective action clear.

Worked example: from long pick paths to controlled flow

The scenario’s illustrative project runs for five months, beginning with measurement and analysis, followed by simulation, a phased pilot and control planning. Its performance comparison is:

If the centre processes 120,000 orders in a month and applies the on-time percentages to the same order denominator, the change corresponds to about 16,080 additional on-time orders per month. That is a useful way to translate a percentage-point improvement into an operational scale, but it is not, by itself, proof that slotting caused the entire change.

A sound evaluation would check for other influences, including volume, staffing, service definitions, carrier performance and seasonal demand. It would compare like-for-like periods or use a carefully controlled pilot, and it would watch for unintended effects such as increased replenishment work or congestion in popular aisles.

The implementation also has real costs. The project requires WMS integration, RFID hardware, a five-month machine-learning project, and retraining for affected teams. The business case should account for these costs, the time needed to maintain models and layouts, and any temporary disruption while stock is moved. A credible financial assessment compares the total implementation and operating costs with validated benefits over an agreed period.

What to watch for when applying AI

Models are useful only within the conditions they have learned to represent. Several practical risks deserve attention from the outset.

Seasonal demand can change the pattern

A model trained mainly on pre-holiday data may perform poorly during peak season, when order mix, volume and staffing differ. Use time-based validation, test peak scenarios and monitor forecast error. Where the operating pattern changes materially, retrain or reassess the model rather than assuming historical relationships still apply.

The best mathematical location may not fit the rack

A slotting recommendation can conflict with rack dimensions, storage rules, weight limits, safety requirements or replenishment access. Treat these as hard constraints, not details to resolve after optimisation. Warehouse operators and safety leads should validate proposed changes before a pilot.

Teams need a voice in the decision

If algorithmic recommendations arrive without explanation or operator input, people may reasonably question whether the system understands the work. Involve pickers, replenishment teams and shift leaders in interpreting the findings. Explain the evidence, invite feedback and make it easy to report exceptions. Their experience can identify conditions that the data does not capture.

Where Ci Flow fits in an AI continuous improvement logistics programme

The warehouse’s models belong in its warehouse analytics stack: the systems that collect operational data, forecast conditions and evaluate or recommend decisions. Ci Flow is a separate product from SigmaFlow and has a different role. It is an operational excellence platform for managing improvement work, not a WMS, a warehouse analytics engine or a home for training machine-learning models.

For a logistics team, Ci Flow can provide a managed workspace for the improvement effort around those models. The team can structure a data-led project using its RDMAICS methodology, document evidence and decisions, track actions and risks, and maintain visibility across the improvement portfolio. Its toolset includes process mapping, value stream mapping, spaghetti diagrams and tollgate presentations, helping connect analysis to decisions and follow-through.

A practical arrangement is to:

• Keep model development, WMS data and operational predictions in the warehouse analytics stack.
• Use Ci Flow to track the DMAIC project, owners, actions, decisions and benefits.
• Bring agreed operational measures into KPI reviews and shift handovers through the organisation’s chosen data and reporting arrangements.
• Review project evidence at phase gates, then capture lessons and control ownership as part of sustainment.

Do not assume a native WMS integration or automatic data pipeline without confirming that it is available for your systems and use case. The value of Ci Flow is in managing improvement work and its governance; the warehouse’s technical architecture remains a separate design decision. Ci Flow also provides project collaboration, portfolio reporting and role-based access, with security details published at ciflow.app/security.

Start with a clear problem, then build the capability

The broader lesson is simple: AI does not replace Lean Six Sigma discipline. It can help teams see patterns across large datasets, predict operational pressures and test alternative designs. DMAIC makes sure those capabilities are applied to a defined customer or business problem, validated with sound measures and sustained through clear controls.

For logistics leaders, a practical starting point is to select one process and one measurable gap. Confirm the definitions and data quality, involve the operators who know the work, and test a limited change before scaling it. That sequence creates stronger evidence than launching a large technology programme around an unverified assumption.

Explore Ci Flow at https://ciflow.app to manage improvement projects, actions and portfolio governance. For the skills to lead data-driven logistics projects, explore Lean 6 Sigma Hub’s CSSC-accredited Green Belt and Black Belt training, which includes practical case studies and project methods.

Choose the next step: build your improvement workspace with Ci Flow, and develop your capability with Lean Six Sigma certification.

Further reading

• Dynamic Storage Location Assignment in Warehouses Using Deep Reinforcement Learning
• Smart Slotting: A Hybrid Machine Learning and Optimization Approach
• DHL: Transforming decision-making with a simulation-powered digital twin
• A DMAIC approach for warehouse storage and order picking process improvement

Kaizen. Kai-Care. Kai-Done. Lean Six Sigma