AI for industrial manufacturing: predictive maintenance, quality control, documentation

A plant with three production lines, hundreds of sensors on machines, quality control based on random checks by an inspector, and piles of documentation manually transcribed into the ERP system at the end of a shift. This is a scene we see regularly. The issue isn’t a lack of data—it’s the lack of a way to turn that data into timely decisions. This is where AI makes sense. But the same plant is often tempted by the promise of a "smart factory that runs itself"—and that’s a different conversation: about safety, responsibility, and the law. Below, we separate one from the other, honestly and without promises that can’t be kept.

Predictive maintenance means predicting machine failures before they happen, based on sensor data: vibrations, temperature, power consumption, pressure, sound. The model learns what "healthy" machine operation looks like and flags deviations that historically preceded failures.

It’s important not to overpromise. AI doesn’t "know" a bearing will fail in 11 days. AI detects that the vibration profile is shifting toward patterns that, in historical data, correlated with imminent failure, and raises an alert with a certain confidence level. The decision to stop the machine and replace the part is made by maintenance—because they bear the consequences of error.

The real scope of benefits depends on data quality and whether there’s a history of failures to train the model:

• Reduction of unplanned downtime: In well-instrumented plants, improvements of 10-30% are typical, rarely more. These are ranges, not guarantees.
• Extended part replacement cycles: Replacing "every X hours" with "when data indicates wear" can reduce premature replacements.
• Prerequisite: Sensors that actually collect signals at the right frequency, and time-stamped failure events. Without these, the model has nothing to learn from.

If historical data is lacking, the honest approach is to first collect and organize the signal, then model. How to structure this data is covered in preparing company data for AI.

Visual quality control is where AI has matured the most. A camera photographs the product on the line, an image classifier assesses whether it detects a defect (scratch, missing component, deformation, print error), and flags items for rejection or manual inspection.

Two key caveats here. First, the model is only as good as its training data—if it hasn’t seen a particular type of defect, it won’t detect it. Second, setting the sensitivity threshold is a business decision, not a technical one: a lower threshold catches more defects but generates more false alarms; a higher threshold lets more good items through but risks missing defects.

That’s why a well-designed system works as a preliminary filter, not the final judge. AI identifies defect candidates, and humans confirm borderline cases. This same "AI screens, humans decide" pattern is discussed in the context of ticket handling in AI classification and routing of tickets.

The least flashy but often most cost-effective area is administration. Manufacturing generates vast amounts of documentation: control cards, shift reports, quality certificates, protocols, WZ documentation. Much of this involves manually transcribing numbers from one place to another.

Here, AI excels at data extraction: reading values from forms, labels, supplier certificates, and transferring them in structured form to the system. A language model can also generate a first draft of a shift report from raw line data—with the caveat that a human signs off on the final version.

The boundary is clear: AI prepares and organizes, humans approve documents with legal or quality value. An agent can collect data and assemble a report, but the signature on a quality certificate is the responsibility of a person, not the model.

Precision is key here, because manufacturing is an area where AI can impact human and property safety.

As long as AI supports decisions—alerts for maintenance, flagging defects for review, drafting reports—and humans remain in the decision loop, legal and operational risks are limited. Problems arise when the system takes actions affecting safety on its own: stopping a machine without supervision, controlling process parameters, approving a batch for shipment without verification.

The AI Act classifies as high risk, among other things, systems that are safety components of products or machinery. If AI is a safety component (e.g., shuts down a press when it "sees" an operator’s hand) or decides on the release of a product affecting safety, obligations apply: technical documentation, risk assessment, human oversight with real intervention capability, logs, and quality monitoring over time.

Three principles we always follow:

• Human in the loop for safety decisions. AI flags, humans approve machine stops, batch rejections, or product releases.
• Logging and model quality monitoring. Models drift over time (raw materials, tools, lighting change); without monitoring, their effectiveness quietly declines. How to measure this is covered in monitoring AI agent quality.
• Audit before production deployment. Before the system goes live on the line, it’s worth going through an AI assistant security audit to check permissions, data access, and misuse scenarios.

The honest sequence for AI deployment in manufacturing is less spectacular than the slides promise.

Start with one narrowly defined problem with measurable costs—most often either predictive maintenance for one critical machine, image-based inspection for one type of defect, or automating one report. Then verify whether data for training the model even exists. Next, run a pilot alongside the current process (AI flags, humans still decide) to measure effectiveness on real data before anything operates autonomously. Only when the numbers add up—gradually expand, with oversight maintained where safety is at stake.

This isn’t a shortcut. But it’s a path that doesn’t end with an expensive system no one trusts because it once made a mistake on a batch worth more than the entire deployment.

No. Predictive maintenance detects deviations that historically preceded similar failures—it works where sensor signals actually foreshadow problems (bearing wear, imbalance, overheating). Sudden failures without prior symptoms or types of malfunctions not present in historical data remain beyond the model’s reach. That’s why we talk about reducing downtime within ranges, not eliminating it.

In practice, no—it shifts the human role. An image classifier automatically screens most items and flags defect candidates, while inspectors focus on borderline cases and decisions carrying responsibility. Full automation without humans only makes sense in very mature, well-validated processes with simple, repeatable defects.

That depends on the architecture you choose—and it’s a decision, not a necessity. Sensitive data (formulas, process parameters, line data) can stay within the plant’s infrastructure using locally run models, or with a provider guaranteeing data localization in the EU and a data processing agreement. It’s worth making this decision consciously at the start, as it shapes the entire project.

Sometimes yes. If AI is a safety component of machinery or decides on the release of a product affecting safety, it may be a high-risk system under the AI Act—with obligations for documentation, risk assessment, and human oversight. Purely administrative systems (reports, data extraction to ERP) usually don’t fall into this category, but the classification should be confirmed for each specific use case.

The honest answer is "it depends on the data," not the model. If sensors are already collecting signals and there’s a history of failures, a pilot for one machine typically takes a few weeks to a few months. If sensors need to be installed first and a failure history built, the data collection phase can take longer than the modeling itself—and no promise can shorten it.