Skip to content
DT Nova

Manufacturing and artificial intelligence: where AI works in a plant, and the rules

Manufacturing and artificial intelligence meet in two places: on the shop floor, where cameras, sensors and machine data feed models that spot defects and faults, and in the office around it, where planners, buyers, quality engineers and maintenance teams work with documents, orders and schedules. This guide explains the uses that work today, the data and systems they depend on, the EU rules on AI in machines and on worker data, and how a small or midsize manufacturer can start with one project that pays back.

It complements our guide to manufacturing predictive analytics, which covers forecasting, scrap prediction and predictive maintenance in depth. Here we look at the wider picture, including vision, scheduling, language models and the product safety rules.

Where AI is used in manufacturing

AreaWhat AI doesData it needs
Visual quality inspectionFinds scratches, cracks, missing parts, wrong labels and assembly errors in camera images, at line speedImages of good and defective parts, with the defect marked
Predictive maintenanceWarns that a machine is drifting toward a fault, so repair is planned instead of an emergencySensor data, fault and repair history
Process qualityPredicts scrap or out-of-tolerance batches from process parameters, and suggests setpointsProcess data per batch, quality results
Production schedulingProposes sequences that cut changeovers and late orders, within machine and staff limitsOrders, routings, capacities, changeover times
Demand and stockForecasts demand by product and site, which drives purchasing and production plansSales and order history, customer forecasts
Knowledge assistantsAnswers questions from manuals, work instructions, quality procedures and past fault reports, with the source shownYour documents, kept current
Supplier documentsReads order confirmations, delivery notes, certificates and invoices, and matches them to ordersDocuments, purchase orders, ERP data
EnergyShifts flexible loads, flags abnormal consumption by line or machineMeter data, production plan, energy tariffs

How far manufacturers have come

Eurostat's 2025 survey of EU enterprises with at least 10 employees found that 19.95% used at least one AI technology. In the manufacturing sector, AI software or systems were used mostly for marketing and sales (30.41%) and for the organisation of business administration processes or management (27.05%). Across all sectors, 33.46% of large enterprises using AI used it for production processes, against 19.02% of small ones (Eurostat, Use of artificial intelligence in enterprises, read on 11 November 2026).

Put simply, most manufacturers who use AI today use it in the office, not on the line. That is a reasonable order: office uses need fewer integrations and carry less risk. But it also means the largest gains, in quality, uptime and throughput, are still open for most small and midsize plants.

Visual quality inspection

Camera inspection is the most mature shop-floor use. Classic machine vision checks measurable things with fixed rules: is the hole there, is the label straight, is the dimension within tolerance. AI-based vision learns what a defect looks like from examples, which suits defects that vary in shape, such as scratches, porosity, contamination or wrinkles in fabric or film. Many good systems combine both: rules for what can be measured, a learned model for what can only be recognised.

What decides success is rarely the model. It is the lighting, the camera position, a stable presentation of the part, and enough examples of each defect type. Real defects are rare, so collecting and labelling them takes time, and the first weeks often go into building that set. Agree up front on the trade-off between two errors: a false reject throws away a good part, a missed defect reaches the customer. A threshold set for zero escapes will reject more good parts, and an operator should review the borderline cases rather than the model deciding alone.

Language models on the shop floor

Plants run on documents: machine manuals, work instructions, quality procedures, shift handover notes and years of fault reports. A knowledge assistant built on a language model can answer a technician's question from those documents in the language they speak, and point to the page it used. This works when the assistant answers only from approved, current documents, shows the source every time, and says when it does not know. It should never invent a torque value or a safety step, and it should not issue commands to machines.

The same technology reads supplier paperwork: certificates of conformity, delivery notes and invoices, checked against orders in the ERP. See document workflow automation and chatbot development for how we build answers from your own content.

Scheduling and optimisation

Production scheduling is often presented as AI, but the strongest tools here are mathematical optimisation and scheduling rules, with machine learning as a helper: predicting real cycle times, changeover times or the chance that an order runs late. The planner keeps the final say, and the system shows why it proposes a sequence. Our page on production scheduling software covers finite-capacity scheduling and changeovers, and manufacturing ERP covers the system the plan comes from.

AI inside machines: the Machinery Regulation

The new EU Machinery Regulation, Regulation (EU) 2023/1230 (read on 11 November 2026), applies from 14 January 2027 and replaces the Machinery Directive. It names AI explicitly. Annex I Part A lists safety components with fully or partially self-evolving behaviour using machine learning approaches ensuring safety functions, and machinery that has such systems embedded, as categories that need third-party conformity assessment. Recital 55 limits this to systems that learn or evolve: software programmed only to execute certain automated functions is not caught by that requirement. Software ensuring safety functions in general appears in the indicative list of safety components in Annex II.

The practical line for a plant: an AI system that advises people, such as an inspection camera flagging parts for an operator or a maintenance alert, is not a safety function. A learning system that stops a robot when a person comes close is. The second belongs with the machine manufacturer and a notified body, and should be designed that way from the start.

The AI Act and manufacturing

The EU AI Act, Regulation (EU) 2024/1689, became generally applicable on 2 August 2026. Under Article 113 as amended by the Digital Omnibus on AI, the high-risk rules apply from 2 December 2027 for the uses listed in Annex III and from 2 August 2028 for AI systems covered by the product legislation in Annex I, which includes machinery (AI Act Service Desk, Article 113, consolidated text as at 27 July 2026, read on 11 November 2026). Most plant uses, such as inspection that an operator reviews, forecasts and document reading, are not high-risk.

Two obligations apply already. Under Article 4, which has applied since 2 February 2025, providers and deployers of AI systems shall take measures to support the AI literacy of their staff and others operating AI on their behalf (AI Act Service Desk, Article 4, read on 11 November 2026). In a plant, that means operators know what the system does, where it is weak and when to overrule it. And an assistant that talks to people should say it is an AI, under Article 50(1).

Worker data: the sensitive part

Annex III point 4(b) of the AI Act lists as high-risk AI systems intended to allocate tasks based on individual behaviour or personal traits, or to monitor and evaluate the performance and behaviour of persons in work relationships (AI Act Service Desk, Annex III, read on 11 November 2026). Cameras on a line, wearables and operator logins all produce data about people, and under the GDPR (Regulation (EU) 2016/679, read on 11 November 2026) processing likely to result in a high risk to people requires a data protection impact assessment before it starts (Article 35), and decisions based solely on automated processing that significantly affect a person are restricted (Article 22).

The simplest design rule is to measure the process, not the person: point the camera at the part, aggregate line data by shift rather than by operator, and keep identifiable data out of models that do not need it. Check national rules on informing and consulting employees before any system that touches staff data goes live.

Data and integration: where the work really is

AI in a plant depends on joining data that lives in separate worlds: machine controllers and historians on the production network, the MES and quality system, and the ERP in the office. Connecting them safely, often through standard interfaces such as OPC UA, with the production network kept separate from the office network, is usually more work than the model. Two decisions shape the design early: what runs at the edge, next to the line, because it must answer in milliseconds or keep working without a connection, and what runs centrally, where models are trained and compared across lines and sites.

New machines make this easier. Under the EU Data Act, connected products placed on the market after 12 September 2026 must make their product data accessible to the user by default; the predictive analytics guide explains what to ask suppliers. For the connected-device side, see IoT software development.

How to start

  1. Pick one problem with a clear cost: defects found at the customer, a scrap rate, unplanned downtime on one line, hours spent searching manuals or matching documents.
  2. Check the data: images, sensor history or documents that exist today, and who owns them.
  3. Record the baseline: how often the problem happens and what it costs per month.
  4. Build the smallest version on one line or one document type, and run it in parallel with the current process without acting on it.
  5. Let operators judge it, measure against the baseline, then put it into daily use with a named owner for its alerts.
  6. Monitor it: a new material, supplier or product changes the data, and models are checked and retrained.

How DT Nova helps

We build manufacturing AI around one decision and the systems that already hold the data: the ERP, the MES or quality system, machine data and your document stores. Results land in the tools people already use, from an inspection screen at the line to a maintenance work order or a planning view. Data stays in your account in an EU data region, and the code and models are yours. Before any build, our AI consulting helps pick the use case worth starting with; for the cost side, see our guide to artificial intelligence cost. For the ERP that holds the production and stock data in the first place, see cloud-based ERP software.

Questions people ask

How is artificial intelligence used in manufacturing?

The uses that work today are visual quality inspection with cameras, predicting faults and scrap from machine and process data, proposing production schedules, forecasting demand, answering questions from manuals and work instructions, reading supplier documents, and adjusting energy use. Most of them advise people on the shop floor or in planning; few control a machine directly.

Is AI in a machine covered by the Machinery Regulation?

It can be. Regulation (EU) 2023/1230 applies from 14 January 2027. Safety components with fully or partially self-evolving behaviour using machine learning that ensure safety functions, and machinery with such embedded systems, are listed in Annex I Part A and need third-party conformity assessment. Software that cannot learn or evolve and only executes programmed functions is not covered by that third-party requirement.

When do the AI Act high-risk rules apply to machinery?

Under Article 113 of the AI Act as amended by the Digital Omnibus on AI, the high-risk rules for AI systems covered by the product legislation in Annex I, which includes machinery, apply from 2 August 2028. The high-risk rules for the uses listed in Annex III, such as monitoring and evaluating workers, apply from 2 December 2027. The Act's general application date was 2 August 2026.

Can AI monitor the performance of production workers?

It is the most sensitive use in a plant. AI systems intended to allocate tasks based on individual behaviour or to monitor and evaluate the performance and behaviour of workers are listed as high-risk in Annex III of the AI Act, and worker data falls under the GDPR, which usually calls for a data protection impact assessment. Measuring the process instead of the person avoids most of these questions.

Where should a small manufacturer start with AI?

With one decision that costs money when it is made late or wrong, such as a defect found at the customer, a scrapped batch or an emergency repair, and with data that already exists for it. Run the AI next to the current process for a few weeks, measure it against a baseline, and extend only what proves itself.

Sources