Every SKU in the right place for how it sells this month, shorter pick walks, waves that meet carrier cut-offs, and a labour plan that sees peak coming. Supervisors approve every move, and the WMS stays in charge.
At a glance
An illustrative scenario. The numbers are typical examples, not a client case study. The cloud services and UK rules were checked on 9 November 2026.
Take a UK distribution centre serving retail stores and online orders from one building: an each-pick zone of shelving and flow rack, a case-pick zone in pallet racking, reserve storage above, and a WMS that runs locations, replenishment and pick tasks. When the WMS went live, every SKU was given a pick face based on the sales data of the time. Since then the range has changed, promotions have come and gone, and seasonal lines have arrived and stayed. Fast movers now sit at the far end of aisles, slow lines take up the best locations, and items that are usually ordered together are picked from opposite ends of the building.
Walking between locations is often the largest single part of a picker's time, so poor slotting is paid for on every order line. It also creates other problems. Popular pick faces run empty mid-wave and wait for replenishment, heavy and bulky lines end up on top or bottom shelves where they're harder to handle, and peak is met with a frantic re-slot weekend that disrupts the operation it's meant to help.
Waves and labour are planned the same way. Shift managers release waves on fixed templates, then juggle to make the carrier cut-offs. Labour is planned in a spreadsheet from last year's numbers, so agency staff are booked late, shifts are extended at short notice, and the site is either overstaffed in the morning or chasing the last vehicle at night.
Agree these, and measure the baseline, before anything is built. Example targets are typical, not promises.
Average walking distance or time per line picked, by zone. Measured from WMS task data, not per person.
ExampleDown by 20–30% in re-slotted zones
Total lines picked divided by total hours worked in the zone. A site measure, never a league table.
ExampleUp, measured against the same weeks last year
Orders picked, packed and on the right trailer before each cut-off.
ExampleAbove 99% outside peak
How often a picker reaches a location with nothing in it. A sign of slotting and replenishment out of step.
ExampleDown by half
SKUs above the agreed weight sitting above shoulder or below knee height.
ExampleZero, with exceptions agreed by health and safety
Planned against actual hours by activity, and how much notice agency and staff got of shift changes.
ExampleWithin 5%, and short-notice changes falling
Cross-sector rules are on the UK regulation map.
Employers must avoid hazardous manual handling where they reasonably can, assess what can't be avoided, and reduce the risk. Slotting is one of the best tools for that. Weight, size and height limits are hard constraints owned by the health and safety lead, and a re-slot that changes where heavy lines sit is checked against the site's risk assessments. Order picking has variable loads, which HSE's V-MAC tool is designed for.
HSE has said health and safety law applies to AI like any other technology: uses of AI that affect workplace health and safety must be risk-assessed, with controls put in place so far as is reasonably practicable, including cyber security. Pick paths, wave pace and slotting all change the physical work, so they're in the site's risk assessment.
Workers are entitled to 11 hours' rest between working days, a 20-minute break when working more than six hours, weekly rest, and an average 48-hour week unless they opt out. Wave plans include breaks, and labour suggestions are checked against these limits before a manager sees them.
The Act gives qualifying zero-hours, low-hours and agency workers rights to reasonable notice of shifts and to payment when shifts are cancelled, moved or cut short at short notice. The detail is being set in regulations and the start date isn't fixed yet. Distribution centres rely on agency labour, so the labour forecast's main job is giving managers enough notice, and every shift change is logged with when it was notified.
Scanner and task data linked to a person is personal data, and using it to monitor performance is worker monitoring. The ICO expects a clear purpose, the least intrusive means, telling workers before it starts, and a DPIA where the risk is high. This design plans from zone and activity data and doesn't score individuals. The ICO notes its monitoring guidance is under review after the Data (Use and Access) Act 2025.
Since 5 February 2026, automated decisions with significant effects need safeguards: telling people, letting them make representations and contest the decision, and a person who can step in. If the site later wants to use pick data in decisions about individual workers, those safeguards apply, along with consultation with staff representatives. In this design, people make every decision about people.
Showing AWSAzuresovereign / on-prem services. Switch with the toggle; every diagram and table on this page follows it.
01Channels
02Intake
03Intelligence
04Orchestration
05Systems of record
06People
| Layer | AWS | Azure | Sovereign / on-prem |
|---|---|---|---|
| Intake: WMS events, history, APIsTwo years of pick history is ideal, so seasonality and peak are both in the data. Most WMSs can export it. | Kinesis Data Streams, S3, API Gateway, AWS Glue | Event Hubs, Azure Data Lake Storage, API Management, Azure Data Factory | Apache Kafka; Ceph or IONOS S3 Object Storage; Kong or your existing API gateway |
| Forecasting: velocity, affinity, seasonality, labourPlain, explainable models do well here. Amazon Forecast is closed to new customers, so forecasts are built in SageMaker AI. | Amazon SageMaker AI | Azure Machine Learning | MLflow and Kubeflow on Kubernetes |
| Optimiser: slots, pick paths, wavesOptimisation, not generative AI. Handling limits, hazardous goods segregation, temperature zones and client separation are hard constraints. | An open-source solver (e.g. OR-Tools or Timefold) on AWS Batch or ECS | An open-source solver (e.g. OR-Tools or Timefold) on Azure Batch or Container Apps | An open-source solver (e.g. OR-Tools or Timefold) on Kubernetes |
| Explanations and plan summariesThe only generative step. It explains why a SKU should move and summarises the night's move list and tomorrow's labour plan in plain words. | Amazon Bedrock | Microsoft Foundry | Open-weight models served with vLLM on your own GPUs, or IONOS AI Model Hub |
| Agent runtimeThe original Bedrock Agents closed to new customers on 30 July 2026; new builds use AgentCore. | Amazon Bedrock AgentCore | Foundry Agent Service | An open-source agent framework (e.g. LangGraph) on Kubernetes |
| WorkflowRuns the nightly cycle: forecast, optimise, approve, send moves to the WMS, check they were done. | AWS Step Functions | Logic Apps or Durable Functions | Temporal, or n8n |
| Middleware and events | Amazon EventBridge, SQS | Event Grid, Service Bus | Apache Kafka or RabbitMQ |
| Link to the WMSThis decides the design. Approved moves become standard WMS move tasks. Nothing writes to WMS tables directly, so a WMS upgrade doesn't break it. | Through your WMS supplier's published APIs or interface tables | Through your WMS supplier's published APIs or interface tables | Through your WMS supplier's published APIs or interface tables |
| Human reviewSupervisors approve moves where they already work, not in a new screen. | In the WMS's own task approval screens, or the supervisors' existing dashboard | In the WMS's own task approval screens, or the supervisors' existing dashboard | In the WMS's own task approval screens, or the supervisors' existing dashboard |
| GuardrailsThe model explains plans; it can't change a constraint or release a task. | Bedrock Guardrails | Foundry guardrails (Azure AI Content Safety) | Open-source guardrails (e.g. NeMo Guardrails, Llama Guard) |
| Identity, keys, data protectionOperator IDs are removed from pick history before it reaches the models. Only planning roles see plans. | IAM Identity Center, KMS, Macie | Entra ID, Key Vault, Purview | Keycloak or your existing Active Directory; OpenBao or HashiCorp Vault; an HSM |
| Monitoring and auditLog every suggested move, every approval and rejection with its reason, and the savings measured afterwards. | CloudWatch, CloudTrail | Azure Monitor, Log Analytics | Prometheus, Grafana, OpenTelemetry |
| UK data residencyPick history is mostly commercial data. Staff and agency data stays in the UK. | eu-west-2 (London) | UK South / UK West | Your own UK data centre, or a UK or EU provider |
On a phone, the table shows the platform you pick. Sovereign / on-prem means open, portable components you can run in your own data centre or with a UK or European provider; IONOS is shown as one example, and others exist. What sovereign cloud means →
Many organisations keep core systems and sensitive data on their own servers. This is how the same design works across both.
Stays on-prem
The WMS, RF scanners, voice picking, printers and any automation controls stay where they are, on site or in your WMS supplier's hosting. The floor runs entirely on the WMS, whatever happens to the AI.
Uses cloud AI
Forecasting, the optimiser and the planning assistant can run in AWS or Azure, or on your own servers for the sovereign option. The optimiser doesn't need GPUs, and slotting is a nightly batch job, so it suits on-prem well.
How they connect
A private link (AWS Direct Connect or Azure ExpressRoute), or a site VPN, through the API layer. WMS events go out; suggested moves and plans come back for approval, and only approved moves become WMS tasks.
If the link drops
Nothing on the floor stops. SKUs stay where they are, the WMS releases waves from its standard templates, and managers plan labour as they do today. The next night's run catches up.
What each agent may do through its tools, what it may not, and where a person signs off. The limits are enforced by the platform, not the prompt.
Planning assistant
Where a person signs offAn inventory control supervisor approves every re-slot move before it reaches the WMS. A shift manager approves every wave plan and decides every labour change. The agent's permissions are enforced by the platform's identity and policy controls, not by instructions in a prompt.
The WMS stays the master of locations, stock and tasks. The AI suggests; the WMS does. Approved moves go in through the WMS supplier's published API or interface tables as ordinary move tasks, so the WMS checks stock, location capacity and conflicts exactly as it does for any other move.
Item data is the weak point in most slotting projects. Weights and dimensions in the ERP item master are often missing or wrong, and the slotting is only as good as they are. Discovery measures a sample, and the item master stays in the ERP, owned by the team that owns it today.
Events keep the AI decoupled. Task completions, receipts, replenishments and new orders become messages on a queue. Each move carries an ID, so a retry never creates a duplicate task, and moves not done by the morning are withdrawn rather than left hanging.
Orders arrive by EDI and APIs into the order management system or WMS as they do now. Carrier cut-offs come from the transport management system. The wave planner reads both and never changes either.
Extract pick and order history, measure travel and empty-face rates, check item master weights and dimensions, confirm what the WMS interface allows, and involve health and safety and worker representatives.
Gate · must be true to passBaseline agreed, WMS integration route confirmed, handling constraints written down, pick-data scope agreed, DPIA started.
One each-pick zone. Two weeks in shadow mode, then supervisors approve live moves in that zone, compared against a similar zone left as it is.
Gate · must be true to passTravel saving and empty-face rate within agreed limits, no handling constraint broken, supervisor and health and safety sign-off.
Re-slotting goes live in the pilot zone, with wave planning added, plus monitoring, support and a rollback plan.
Gate · must be true to passMonitoring, support model, rollback plan and supervisor training in place.
Add zones one at a time, then labour planning, then other sites on the same WMS.
Gate · must be true to passEach new zone passes the same pilot gate, and a pre-peak re-slot is planned each year.
| Activity | Site general manager | Operations manager | Solutions architectI can fill this | Technical PMI can fill this | EngineersI can fill this | Health & safety | HR & DPO | SecurityI can fill this |
|---|---|---|---|---|---|---|---|---|
| Business case and benefits | Accountable | Responsible | Consulted | Consulted | Informed | Informed | Informed | Informed |
| Target architecture and WMS integration | Informed | Consulted | Accountable and Responsible | Consulted | Consulted | Informed | Informed | Consulted |
| Handling constraints and risk assessment | Consulted | Consulted | Consulted | Informed | Consulted | Accountable and Responsible | Informed | Informed |
| Pick-data scope, DPIA and worker consultation | Accountable | Consulted | Consulted | Informed | Informed | Consulted | Responsible | Consulted |
| Build and integration | Informed | Consulted | Accountable | Responsible | Responsible | Informed | Informed | Consulted |
| Pilot testing in one zone | Informed | Accountable | Responsible | Responsible | Responsible | Consulted | Informed | Informed |
| Go-live decision | Accountable | Responsible | Consulted | Responsible | Informed | Consulted | Consulted | Consulted |
| Run, monitor and improve | Informed | Accountable | Consulted | Consulted | Responsible | Consulted | Consulted | Consulted |
R ResponsibleA AccountableC ConsultedI Informed· not involved
Option A
If there's no architect on the programme, I'd own the design end to end: the forecasting and optimisation, the integration with your WMS, ERP and transport systems, and the gates. I'd start with the location masters, task interfaces, replenishment rules and order flows you already have, because that's where slotting changes succeed or stall. I'd report to your programme board.
Option B
If your partner, your WMS supplier or your own team already has an architect, I'd take on the WMS integration and the optimisation engine under their lead: the part where suggestions have to turn into tasks the WMS will accept. I'd also give them a second senior opinion on data quality, handling constraints and the rollback plan.
Goes wrong
Item weights and dimensions are wrong, so the slotting puts things where they don't fit or can't be lifted safely.
How the design avoids it
Measure a sample in discovery, fix the item master at source, and treat missing data as a reason not to move a SKU.
Goes wrong
The model suggests thousands of moves, and the cost of making them eats the saving.
How the design avoids it
Rank moves by payback, cap them to the labour available each night, and measure the saving each move delivered.
Goes wrong
The optimiser puts a heavy line on the top shelf because it cuts walking.
How the design avoids it
Handling limits are hard constraints set by health and safety. The optimiser can't trade them away.
Goes wrong
Pick data quietly becomes a per-person performance tool nobody decided on.
How the design avoids it
A written scope agreed with HR and worker representatives, operator IDs removed before modelling, and a DPIA before any change of purpose.
Goes wrong
The integration writes straight into WMS tables and breaks at the next upgrade.
How the design avoids it
Use the WMS supplier's published APIs or interface tables only, and test every WMS upgrade against them.
Changes already announced or under way that affect this design. More in the UK horizon scan.
Log when every shift change and agency booking change is notified, and make the notice threshold a setting, so the rules can be applied on the day they start.
Involve worker representatives from discovery, keep per-person data out of scope, and write down what pick data is for. That fits any likely outcome.
Put the AI, and any robots or automation you add later, in the site's risk assessment now, so new guidance is a check, not a rewrite.
Keep the approval log, the pick-data scope and the DPIA in a form you could show the ICO or a staff representative.
If your WMS or AI platform is run by a managed service provider, expect new duties for them and update your contracts and supplier assurance.
Good answers are specific. Vague answers to any of these are worth following up.
Whether you need someone to lead the design or to support the architect you already have, let's talk it through.
