Every unit checked at line speed, clear passes sent on, anything doubtful held for a person, and every defect traced to its batch. Inspectors make the calls that matter.
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.
On most lines, final inspection is a person looking at parts as they pass: scratches, cracks, missing clips, misaligned labels, solder faults, colour mismatches. People are good at spotting the unexpected, but attention fades over a shift, judgement varies between inspectors, and sampling means most units are never looked at closely. Some lines already have rule-based machine vision, which works well for measured features and struggles with variable defects like surface marks.
Escapes are expensive. A defect found by a customer costs far more than one found at the line: returns, a corrective action report, sometimes containment at the customer's site, and a quality score that affects future orders. For products in scope of safety law, a defect that creates a risk can mean corrective action or a recall, and then the first question is which batches are affected.
That question is often hard to answer quickly. Inspection results sit on paper sheets or in a spreadsheet, rework isn't always recorded against the unit, and linking a defect to a material lot, a machine and a time takes a day of digging. A good inspection system has to make that link automatically, not just spot defects.
Agree these, and measure the baseline, before anything is built. Example targets are typical, not promises.
Defects found by customers, in parts per million shipped. The measure that matters most.
ExampleDown by half in the first year
Good units sent to the hold lane. Too many and the hold lane becomes a bottleneck and inspectors stop trusting it.
ExampleUnder 2% of units
How often the model's verdict matches an inspector's on a blind audit sample, by defect type.
ExampleAbove 95% on clear cases, reviewed weekly
How long a held unit waits for a person's decision.
ExampleUnder 10 minutes on a running line
Rejected and reworked units with a complete record: image, verdict, decision, batch, lot and time.
ExampleEvery unit
Share of inspectors' time spent on borderline cases and root cause, not routine checking.
ExampleRising, measured quarterly
Cross-sector rules are on the UK regulation map.
The manufacturer stays responsible for placing safe products on the market. AI inspection is one control among several, not a substitute for design safety or testing. Batch traceability matters most here: if a corrective action or recall is needed, you can say quickly which units are affected. The 2025 Act lets government update product rules through secondary legislation, so watch for changes in your product areas.
CE marking is recognised in Great Britain indefinitely for 21 product regulations, including electrical equipment and machinery, so many manufacturers use one mark for both markets. Either way, the declaration of conformity and the production controls behind it are yours. If the AI changes how you control production, update the quality plan and technical documentation that support the mark.
The vision system is a monitoring and measuring resource: show it's fit for purpose, keep it under control, and keep records of nonconforming output and what was done with it. Model changes go through the same change control as gauges and test rigs. If you supply automotive customers under IATF 16949, expect them to ask for a measurement system analysis.
Cameras pointed at parts often catch hands, faces and badges. Do a DPIA, tell workers before cameras go live, consult their representatives, and crop, mask or angle cameras so people aren't recorded. Don't use inspection images or inspectors' decisions to judge individuals' performance unless that has been assessed and explained separately. Keep images only as long as quality records need them.
Adding cameras, lighting and a divert gate changes the line. Risk-assess the change, keep guards and interlocks intact, and keep the divert logic in the PLC, owned by the controls engineers. The model never touches a safety function or stops the line.
Showing AWSAzuresovereign / on-prem services. Switch with the toggle; every diagram and table on this page follows it.
01Shop floor (OT)
02Edge
03Intelligence
04Orchestration
05Systems of record
06People
| Layer | AWS | Azure | Sovereign / on-prem |
|---|---|---|---|
| Cameras, lighting, triggersLighting matters more than the model. Prove you can see the defect in a photo before buying anything else. | Industrial cameras and lighting chosen per defect | Industrial cameras and lighting chosen per defect | Industrial cameras and lighting chosen per defect |
| Inference at the lineVerdicts in tens of milliseconds, with no cloud dependency. Amazon Lookout for Vision (ended 31 October 2025) and AWS Panorama (ended 31 May 2026) are no longer options. | Model trained in SageMaker AI, deployed to a GPU edge box by AWS IoT Greengrass | Model trained in Azure Machine Learning, run on Arc-enabled Kubernetes with Azure IoT Operations, or Azure IoT Edge 1.6 LTS | ONNX Runtime or NVIDIA Triton on Kubernetes at the plant |
| Model trainingAzure Custom Vision retires on 25 September 2028, and Microsoft points new work to Azure Machine Learning. Amazon Rekognition Custom Labels is still offered, but it's a cloud API, so it suits offline checks better than line-speed verdicts. | Amazon SageMaker AI | Azure Machine Learning (AutoML for images) | MLflow and Kubeflow on Kubernetes |
| Image labellingSageMaker Ground Truth closed to new customers on 30 July 2026. Your inspectors are the labellers: their decisions on held units are the best training data you have. | An open-source tool such as CVAT or Label Studio, in your AWS account | Azure Machine Learning data labelling | CVAT or Label Studio on Kubernetes |
| Image and result storeKeep held, rejected and audit-sample images with their record. Most pass images can be kept briefly and then deleted. | Amazon S3 | Blob Storage | Ceph, or IONOS S3 Object Storage |
| Quality investigation and draftingThe vision model decides what it sees; the language model only groups, links and drafts. The original Bedrock Agents closed to new customers on 30 July 2026; new builds use AgentCore. | Amazon Bedrock, Bedrock AgentCore | Microsoft Foundry, Foundry Agent Service | Open-weight models served with vLLM; an open-source agent framework (e.g. LangGraph) on Kubernetes |
| Workflow and events | AWS Step Functions, EventBridge, SQS | Logic Apps or Durable Functions, Event Grid, Service Bus | Temporal or n8n; Apache Kafka or RabbitMQ |
| Link to MES, QMS and ERPThe MES owns unit and batch genealogy. Inspection results are written there, not kept in a separate AI database. | API Gateway, plus your existing integration platform | API Management, plus your existing integration platform | Kong or your existing API gateway |
| Human reviewShow the image, the highlighted area and two or three similar past decisions. Keep it to one tap per decision. | On a screen at the hold lane, recording into the MES | On a screen at the hold lane, recording into the MES | On a screen at the hold lane, recording into the MES |
| Guardrails for the agent | Bedrock Guardrails | Foundry guardrails (Azure AI Content Safety) | Open-source guardrails (e.g. NeMo Guardrails, Llama Guard) |
| Identity, keys, OT securityThe edge box sits in its own network zone. Only the PLC interface agreed with controls engineers crosses into OT. | IAM Identity Center, KMS, device certificates in AWS IoT Core | Entra ID, Key Vault, Azure Device Registry certificates | Keycloak or your existing Active Directory; OpenBao or HashiCorp Vault; an HSM |
| Monitoring and auditTrack hold rates, agreement with inspectors and image quality per camera, so drift and a dirty lens are spotted early. | CloudWatch, CloudTrail | Azure Monitor, Log Analytics | Prometheus, Grafana, OpenTelemetry |
| UK data residencyImages of your products can be commercially sensitive. Some customers' contracts limit where they can be stored. | 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
Every live verdict is made at the line, on an edge box in the plant. The PLCs, divert gates, hold lanes and the MES stay on site, and the line runs whether or not the cloud is reachable.
Uses cloud AI
Model training, labelling, the image archive and the quality investigation agent run in AWS or Azure, or on your own GPUs if product images must not leave your estate. New models are pushed back to the edge only after the quality manager approves them.
How they connect
Each plant's edge zone sends results and selected images out over an outbound-only, encrypted connection, over AWS Direct Connect, Azure ExpressRoute or the internet with mutual TLS. Model updates come down through the same managed channel, never by hand.
If the link drops
Inspection carries on at full speed with the current model. Images and results queue locally and upload when the link returns. If the edge box itself fails, every unit goes to hold and the line falls back to manual inspection.
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.
Quality investigation agent
Where a person signs offInspectors decide every held unit. A quality engineer decides containment and approves every NCR, and the quality manager approves every model change before it reaches the line. The agent's permissions are enforced by the platform's identity and policy controls, not by instructions in a prompt.
The edge box talks to the line through one interface, agreed with the controls engineers: a pass or hold signal per unit, read by the PLC's own logic. The PLC decides what moves where, and a missing or late verdict means hold. Nothing else crosses into OT.
Every verdict becomes an event carrying the unit or batch ID, so the MES, the quality system and the image store all agree. If the MES is down, events queue at the edge and are written in order when it returns, with no duplicates.
Models are versioned and deployed like any other plant software: tested on a held-back set of your own images, approved, rolled out one line at a time, with the previous version ready to roll back.
Pick one line and one defect family, prove the defect is visible in test photos under controlled lighting, measure today's escape and false-reject rates, and agree the PLC interface.
Gate · must be true to passDefect visible in images, baseline measured, DPIA started, worker representatives briefed, controls engineers signed up.
Cameras and edge inference on one line in shadow mode: the model gives verdicts, inspectors work as normal, and the two are compared on every unit.
Gate · must be true to passAgreement with inspectors and false-hold rate within agreed limits, model validation signed off by the quality manager.
The model drives pass or hold on the pilot line, with inspectors deciding every held unit, monitoring, support and a rollback plan.
Gate · must be true to passHold-lane capacity proven at full speed, traceability complete in the MES, rollback tested.
Add defect families and stations, then lines and plants, reusing the edge pattern and the hold-to-decision flow.
Gate · must be true to passEach new line and defect family passes the same pilot gate.
| Activity | Operations director | Quality manager | Controls / OT engineer | Solutions architectI can fill this | Technical PMI can fill this | EngineersI can fill this | SecurityI can fill this | DPO |
|---|---|---|---|---|---|---|---|---|
| Business case and benefits | Accountable | Responsible | Consulted | Consulted | Consulted | Informed | Informed | Informed |
| Target architecture and edge design | Informed | Consulted | Consulted | Accountable and Responsible | Consulted | Consulted | Consulted | Informed |
| PLC interface and line changes | Informed | Consulted | Accountable and Responsible | Consulted | Consulted | Consulted | Consulted | Informed |
| DPIA and worker consultation | Consulted | Consulted | Informed | Consulted | Consulted | Informed | Consulted | Accountable and Responsible |
| Build and integration with MES, QMS and ERP | Informed | Consulted | Consulted | Accountable | Responsible | Responsible | Consulted | Informed |
| Model validation on your own parts | Informed | Accountable | Consulted | Consulted | Consulted | Responsible | Informed | Informed |
| Go-live decision | Accountable | Responsible | Consulted | Consulted | Responsible | Informed | Consulted | Consulted |
| Run, monitor and improve | Informed | Accountable | Consulted | Consulted | Consulted | Responsible | Consulted | Informed |
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 edge pattern at the line, the interface with your controls engineers, the traceability into your MES and quality system, and the gates. I'd report to your programme board and work closely with your quality manager, who owns the standard the model has to meet.
Option B
If your partner, your machine vision integrator or your own team already has an architect, I'd take on the MES, QMS and ERP integration, the event flow and the quality agent under their lead, and give them a second senior opinion on OT security and on how the model is validated.
Goes wrong
The demo used perfect parts under perfect light. The line has glare, dust and vibration.
How the design avoids it
Prove the defect is visible in photos taken on the real line, at real speed, before choosing a model.
Goes wrong
Too few examples of rare defects, so the model never learns them.
How the design avoids it
Start with defect families you see often, use anomaly detection for the rare ones, and grow the training set from inspectors' decisions.
Goes wrong
The hold lane overflows and the line has to stop.
How the design avoids it
Size the hold lane and inspector cover for the false-hold rate seen in the pilot, and measure it every shift.
Goes wrong
Results live in a separate AI database, so traceability still means a day of digging.
How the design avoids it
Write every verdict and decision to the MES against the unit or batch ID.
Goes wrong
Workers hear about the cameras from a rumour, and trust is lost.
How the design avoids it
DPIA, consultation and a clear statement of purpose before any camera is fitted.
Changes already announced or under way that affect this design. More in the UK horizon scan.
If you run vision models on IoT Edge, build on 1.6 LTS or Azure IoT Operations, and plan an upgrade path for every edge box.
Keep traceability and inspection records in a form that would support a corrective action or recall under whatever rules follow.
Plan the vision system's validation records and change control so they carry straight into your transition audit.
Don't start new work on it. If an existing line uses it, move the model to Azure Machine Learning as part of this programme.
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.
