How to use AI in manufacturing: what to automate first
In a factory, AI helps fastest around the machines, not at them: an assistant that finds the procedure for a fault code, structured shift reports, quality documentation prepared from records, supplier documents checked automatically, and visual inspection. Start with documents you already have, on a single line, with no new sensors.
In a factory, AI is most likely to help not at the machines but around them: the manuals and maintenance logs nobody searches, shift reports, quality documentation, supplier orders and certificates. In the EU, only 17.3% of manufacturing enterprises used AI in 2025, and in Romania 3.2% (Eurostat). This guide is for the plant or operations director who wants to know where to start.
How do you use AI in manufacturing?
You use AI on knowledge and documents: an assistant that finds the procedure for a fault code, structured shift reports from notes, non-conformance reports prepared from records, supplier documents checked against the order, and visual inspection on the line. You do not use AI to control machines or safety functions. The table sums up the common problems.
| Problem | What AI can do today | What it does not do well |
|---|---|---|
| Manuals, fault codes, maintenance logs | Finds the procedure and past interventions, citing the document | Does not replace the technician's diagnosis |
| Shift handover and production reports | Builds the structured report from notes, messages or dictation | Does not know what nobody wrote down |
| Quality documentation | Prepares non-conformance reports and audit packs from records | Does not sign off or decide the root cause |
| Visual inspection | Flags visible defects at line speed | Cannot see internal defects |
| Supplier orders, delivery notes, certificates | Extracts the data and checks it against the purchase order | Cannot reliably read a poor scan |
1. Maintenance knowledge
When a machine stops, the technician searches manuals hundreds of pages long, the intervention log and the memory of the colleague with 20 years on the floor. An AI assistant built on manuals, procedures and the intervention history answers "what does fault code E-214 on press 3 mean, and what was done last time?" with a reference to the exact document and intervention. It works on machines older than any API, because it works with the documents, not the machine.
What it cannot do: diagnose the fault on its own or control the machine. The assistant shortens the search; the technician decides.
2. Shift handover and production reports
Shift handover often happens on paper, on WhatsApp or from memory. AI can turn notes, messages or a two-minute dictation into a structured report: what was produced, which stoppages occurred, what is still open. The report lands in the same place every shift and can be searched a month later.
3. Quality documentation
Non-conformance reports, 8D reports, customer audit packs and certificates of conformity are written from records that already exist: measurements, batches, complaints. AI gathers the data and prepares the draft in the format the customer asks for. The quality engineer checks it, completes the root-cause analysis and signs it off.
4. Visual inspection
On visible, repetitive defects (scratches, missing components, wrong labels), a computer-vision system checks every part, not just a sample, and sends uncertain cases to a person. The details, including why lighting matters more than the model, are in how to use computer vision to automate inspection.
5. Supplier documents and predictive maintenance
Orders, confirmations, delivery notes and material certificates arrive in a different format from every supplier. AI extracts the data and checks it against the purchase order and the goods receipt; differences go to purchasing. The pipeline is in how to automate document processing with AI.
Predictive maintenance is the most promised and least delivered case. It needs sensor data over long periods and a well-recorded failure history. If the maintenance history is on paper, the first step is to make it searchable, not to predict failures. The impressive downtime-reduction figures in circulation usually have no public study behind them.
What should a manufacturer automate first?
Start with a case that works on documents you already have and where a person checks the result anyway: the maintenance assistant on a single line, or supplier documents. Both work without touching the machines and without new sensors. If you want the advisory view on automating operations first, Vlad Tudor covers it on his page on AI process automation for operations teams.
A hypothetical example: a component plant with three lines would start with the manuals and intervention log for the line with the most stoppages. It measures today's time from stoppage to finding the procedure, runs the assistant for a month with the technicians, then compares. That is the only way to learn the real gain: measured in your own pilot. The steps are in how to run an AI pilot.
One obstacle to know about in advance: among EU manufacturers that considered AI and did not proceed, 44.7% named incompatibility with existing equipment and software (Eurostat, KS-01-26-009). That is why the first projects work with exports and documents, and integration with the ERP or MES comes once the case is proven.
Data and compliance: the AI Act, GDPR, production data
A few things to know. This is not legal advice; for your case, talk to your company's lawyer.
- AI Act: assistants for documentation, reports and supplier documents are not, as a rule, high-risk systems. AI that acts as a safety component of a machine falls under Annex I of Regulation (EU) 2024/1689, with obligations from 2 August 2028 under Regulation (EU) 2026/1744. We do not build systems that control machines or safety functions.
- GDPR: shift reports and dictations contain names and sometimes employees' voices, so they are personal data. Tell people what is recorded and why, and do not use the reports to evaluate individuals without a separate assessment.
- Production data: recipes, drawings and process parameters are the company's know-how. If you do not want them to leave the plant, the system can run on-premise or in a cloud you control.
- Since February 2025, Article 4 of the AI Act requires measures that support the AI literacy of the people operating the systems. For a team of technicians, that usually means a short training session on the specific case.
How do I start?
- Tech Call, free: a short conversation where you show us the problem and we tell you honestly whether AI is the answer. Book it here.
- Tech Audit: a short, paid audit at the plant, on real documents and exports. You get in writing the use case, the data it needs, where the system runs and what "working" means in numbers.
- Pilot: one line or one document flow, with your people checking every result.
- Production: integration with the ERP or MES where it makes sense, monitoring, training. The code and documentation stay with you.
What we can show: computer vision, speech (including transcription models adapted to a client's terminology) and assistants that answer from documents with the source cited, like ai-aflat.ro, built on 220,000+ legislative acts. What we build is on the AI automation and AI agents pages, and our published projects are on the projects page.
Sources
- Eurostat, enterprises using AI by economic activity (isoc_eb_ain2), reference year 2025
- Eurostat, statistical report KS-01-26-009, published 26 March 2026 (Table 8, reasons for not using AI)
- Regulation (EU) 2024/1689 (AI Act), Article 4 and Annex I
- Regulation (EU) 2026/1744, the new dates for high-risk systems
In 2025, 17.3% of EU manufacturing enterprises used AI (Romania: 3.2%), and 44.7% of those that considered AI and did not proceed named incompatibility with existing equipment and software (Eurostat).
Frequently asked questions
Does AI work on old machines with no connectivity?
Yes, if it works with the documents rather than the machine. A maintenance assistant reads manuals, procedures and the intervention log, and a document system reads orders and delivery notes. Machine data only becomes necessary for predictive maintenance or real-time monitoring.
Should I start with predictive maintenance?
Usually not. It needs sensor data over long periods and a well-recorded failure history, which few plants have. If the history is on paper, the first step is to make it searchable. The impressive downtime-reduction figures in circulation usually have no public study behind them.
Does production data leave the plant if I use AI?
It does not have to. The system can run on-premise or in a cloud the company controls, and if you use a model through an API, the contract has to state where data is processed and rule out training on it. The decision is made in the audit, before the pilot.
Is an AI system in a factory high-risk under the AI Act?
Assistants for documentation, reports and supplier documents usually are not. AI that acts as a safety component of a machine falls under Annex I, with obligations from 2 August 2028. The exact classification depends on the specific system. This is not legal advice.
What should a factory automate first?
A case that works on existing documents and where a person checks the result anyway: the maintenance assistant on the line with the most stoppages, or supplier documents. Measure today's time, run the pilot for a month and compare. Visual inspection is a good second step if the defects are visible.
Want to discuss a project?
Book a free discovery call with the Sapio team.