AI for Manufacturing and Industrial
Where AI meets machinery that has mass. Thirteen verticals, written for the person who signs off a change to a running line rather than the person who demonstrates one in a lab.
The physics does not care how good your model is.
Manufacturing is the sector where AI meets consequences that are mechanical rather than statistical. A model that misclassifies a product recommendation costs a sale. A model that misclassifies a weld, a torque value or a proximity signal costs a recall, a line stoppage or a person. That changes what counts as sufficient evidence, and it changes who has to sign.
It also changes where the difficulty sits. Almost every manufacturing AI project we assess is bottlenecked not on modelling but on data that was never collected for this purpose: sensor histories at the wrong resolution, quality records kept for audit rather than analysis, maintenance logs written as free text by whoever was on shift, and machine data locked in a control system nobody has an interface to. The engineering that matters is usually the unglamorous half.
Where AI actually earns its place on a factory floor
Ranked by evidence rather than by trade-show frequency. The maturity column is our own read; the catch column is what the vendor demonstration leaves out.
| Where | What it does | Maturity | The catch |
|---|---|---|---|
| Visual inspection and defect detection | Classifies surface, assembly and dimensional defects from imagery at line speed. | Proven | The most reliable value in the sector by a distance — provided you have labelled defect examples, which is exactly what a good process does not generate. |
| Process parameter optimisation | Tunes setpoints against yield, energy and quality outcomes. | Proven | Strong where the process is instrumented and the operating window is genuinely explorable. Frequently it is not, for good safety reasons. |
| Production scheduling and sequencing | Sequences jobs against changeovers, due dates and constraints. | Strong | This is operations research more than machine learning, and it works. The constraint is whether the plan survives contact with the shop floor. |
| Demand and supply planning | Forecasts demand and drives material requirements. | Strong | Ordinary forecasting over unusually long lead times, so errors are expensive and slow to correct. |
| Predictive maintenance | Predicts equipment failure before it happens. | Mixed | The most over-sold application in manufacturing. Needs failure examples that well-maintained plants do not produce, and most pilots never scale. |
| Energy and utilities optimisation | Reduces consumption against production and tariff conditions. | Strong | Underrated, measurable, and it pays without touching product quality — which makes it the easiest first project to get approved. |
| Generative design and process simulation | Explores design or process spaces computationally. | Early | Real in narrow, well-simulated domains. Frequently sold far beyond where the physics is actually captured. |
| Autonomous safety functions | AI deciding whether machinery moves or stops. | Not ready for most | Technically feasible and now a notified-body conformity assessment under the Machinery Regulation. The engineering is the easy part. |
The number that should temper every predictive maintenance pitch
Around two-thirds of maintenance teams report planning to adopt AI, and roughly a third have partially or fully implemented it. That gap is the story. The reason is rarely modelling: predictive maintenance needs examples of the failure you are trying to predict, and a well-maintained plant does not produce many — you may have thirty years of operation and eleven relevant failures, which is not a training set. Budget six to twelve months of baseline data collection before predictions become reliable, expect most of the work to be data cleaning and failure-mode mapping, and treat any vendor's downtime-reduction figure as a claim to be tested on your own equipment. See predictive analytics.
Five regimes converging on the same machine
Industrial AI sits under machinery safety law, functional safety standards, the AI Act, cybersecurity requirements and now data access rights — and for a connected machine with an AI safety function, all of them apply at once. This is our reading as at September 2026 and we work alongside your regulatory, safety and legal functions rather than in place of them.
Applies from 20 January 2027
Regulation (EU) 2023/1230 replaces the Machinery Directive for machinery placed on the market from 20 January 2027; machinery placed before that date stays under the old directive. It addresses AI directly, covering safety components with fully or partially self-evolving behaviour using machine learning, and requiring that such components do not cause machinery to act beyond their defined task and movement space. Annex I is split into a part requiring third-party assessment and a part allowing self-certification against harmonised standards.
Annex I, from 2 August 2028
The AI Act lists the Machinery Regulation as Union harmonisation legislation, which means an AI system that is a safety component of machinery requiring third-party conformity assessment is automatically high-risk. Obligations apply from 2 August 2028 following the Digital Omnibus extension. A pre-trained model frozen before deployment still counts; a purely rule-based control system generally does not. Classification follows function rather than what the product is called.
Four requirements that shape the build
Declared capability and bounded autonomy, stated explicitly rather than left implicit in model weights. Safe stop and interlock logic that validates commands against a declared envelope and withstands foreseeable malicious attempts, with emergency stop outside the system's reach. Cybersecurity for model files, configuration and tool access. And documented intended evolution — training data, evaluation results and every update — as an auditable record that triggers re-certification.
Reporting now, full obligations December 2027
For products with digital elements, vulnerability reporting obligations began on 11 September 2026 and the main cybersecurity requirements apply from 11 December 2027, with third-party assessment for high-risk products. For a manufacturer shipping connected machinery this lands on the same product as the Machinery Regulation and the AI Act, and the three documentation sets have to agree with each other.
Applying since 12 September 2025
Users of connected products have the right to access the data their use generates, and to share it with third parties including independent repairers. Connected devices sold in the EU must be designed to allow that sharing. For equipment makers this is a commercial question as much as a compliance one, because service revenue built on exclusive access to machine data is the model the regulation was written to open up.
The standards that were already there
IEC 61508, ISO 13849 and ISO 12100 did not go away, and for most machinery they remain the operative framework. A machine learning component inside a safety function has to be reconciled with a standards regime built around deterministic behaviour and quantified failure rates, which is genuinely hard and is the reason most credible industrial deployments keep AI advisory rather than actuating.
What this means for a build
The decisive question is whether your AI performs a safety function or informs a person who does. An advisory system that flags a likely defect for an operator to confirm sits outside almost all of the above. The same model wired to stop the line automatically is a safety component, with notified body assessment, AI Act high-risk obligations and functional safety reconciliation attached. That is not an argument against automation — it is an argument for deciding deliberately in week one, because the two architectures are different from the first sprint and converting one into the other is a rebuild. See EU AI Act compliance readiness and AI risk assessment.
Who we write for
Each page starts from that organisation's own problems, names the regulatory exposure it carries, and routes into the engineering. Depth varies and is stated on each page.
Discrete manufacturing
Visual inspection, scheduling and quality root cause — with defect data treated as the real constraint.
Read more →Process manufacturing
Soft sensors, yield optimisation and batch analytics — inside the operating envelope, not against it.
Read more →Automotive & OEM
Inline inspection, weld and paint quality, warranty analytics and supplier quality — at takt time.
Read more →Aerospace & defence
Inspection support, composite quality, configuration data and MRO planning — under airworthiness constraints.
Read more →Electronics & semiconductors
Defect classification, wafer map analysis, yield analytics and test optimisation — the densest data in industry.
Read more →Chemicals
Soft sensors, yield and energy optimisation, formulation analytics and regulatory documentation.
Read more →Pharmaceuticals manufacturing
Batch record review, deviation analytics and inspection — under GxP, with validation designed in.
Read more →Heavy machinery & equipment
Fleet analytics, service prediction and warranty — with Machinery Regulation and Data Act obligations named.
Read more →Metals & mining
Grade and recovery optimisation, heavy asset health and energy — where a percentage point moves millions.
Read more →Textiles
Fabric defect detection, colour consistency and planning — plus the supply chain data DPP will need.
Read more →Packaging
Print and web inspection, registration control, and setup waste — where speed makes inspection the only option.
Read more →Industrial IoT
Edge inference, sensor pipelines and condition monitoring — with CRA and Data Act obligations designed in.
Read more →Contract manufacturing
Quoting, scheduling across a mixed portfolio and customer-specific quality — with data segregation as a constraint.
Read more →FAQ
Marked up with FAQPage schema so these answers can surface in search results and inside AI assistant responses.
Do you have manufacturing experience?
Yes in visual inspection, process optimisation, scheduling, quality analytics and industrial data engineering. Less in functional safety certification, where we work alongside your safety engineers and notified body rather than in place of them, and none in control system design. Each of the thirteen vertical pages states our depth in that area rather than implying uniform expertise across the sector.
Why do our predictive maintenance pilots never scale?
Usually because the failure data is not there. Predicting a failure mode requires examples of it, and a well-maintained plant produces very few — thirty years of operation might yield eleven relevant events, which will not support a model however it is framed. The pilots that scale are the ones that started with a specific high-consequence failure mode, enough historical examples to learn from, and a maintenance process that can actually act on a warning. The ones that stall usually skipped the first question.
Is our AI system a machinery safety component?
It depends on function, not on what it is called. If it decides whether machinery moves or stops, whether a person is too close, or whether an unsafe condition exists, it is a safety component — with third-party conformity assessment under the Machinery Regulation from 20 January 2027 and automatic high-risk classification under the AI Act. If it flags a condition for a person to act on, it generally is not. Settle this in week one with your safety team, because the two architectures diverge immediately.
Where should a manufacturer start?
Visual inspection or energy optimisation, usually. Inspection has the strongest evidence base in the sector and a clear measurable outcome, provided you have labelled defect examples. Energy optimisation is underrated: it is measurable, it does not touch product quality, and it therefore clears internal approval far faster than anything near the process or the safety envelope. Both build the data infrastructure that harder projects need.
What is the most common way manufacturing AI projects fail?
The data was never collected for the purpose. Sensor histories at the wrong resolution, quality records structured for audit rather than analysis, maintenance logs as free text, and machine data trapped in a control system with no interface. Teams discover this in month three, having scoped a modelling project and found a data engineering one. We front-load that assessment, and it has ended engagements before anyone spent money on a model.
Start with the problem, not the technology.
Thirty minutes, no charge, no deck. Tell us what is going wrong in your organisation and we will tell you whether AI is the right instrument — including when it plainly is not.