Quality assurance in Austrian companies is often based on experiential knowledge: a skilled worker can tell from the workpiece, a document, or a deviation right away that something is wrong. Artificial intelligence can support this inspection by making patterns visible and offering suggestions. However, it must not silently decide whether a product is shipped, an invoice is approved, or a service is assessed. The qualified person remains responsible.
Especially for small and medium-sized enterprises, AI quality control initially sounds like something for heavy industry. In fact, it often starts with simple tasks: comparing images, sorting measurements, flagging missing information, or preparing a checklist. Statistik Austria reported for 2025 that already 30 percent of companies with ten or more employees use AI. Therefore, what matters is not whether a company is "modern enough," but whether the specific use case is comprehensible, data-minimizing, and secure.
What AI can do in quality assurance
A model can pre-sort large amounts of similar information faster than a human. In a carpentry shop, an image inspection can identify scratches on a surface; in customer service it flags incomplete ticket data; in accounting it points out unusual amounts. The result is a prioritization. The skilled worker decides whether there is actually an error and which correction makes sense.
The greatest benefit is relieving routine tasks. Employees spend less time searching for obvious deviations and more time on root-cause analysis, customer contact, and improvement. For this reason the AMS describes AI in the workplace as an issue that changes not only technical but also professional and social skills. Good quality remains a joint achievement of tools and experience.
Narrow the use case
Don't start with the goal "We want to use AI," but with a clear error question. Should the system find missing mandatory fields? Should it check measurements against tolerances? Or should it sort photos by visible surface defects? The narrower the question, the easier it is to test whether the output is reliable. A vague quality assessment, by contrast, quickly leads to false expectations.
Before the pilot, define three classes: automatically acceptable, to be reviewed by a human, and clearly defective. The middle class is important. It prevents an uncertain result from being mistaken for approval. Also specify how a person corrects a misclassification and how that correction is documented. Without a feedback channel, the team learns nothing from the system's errors.
Prepare data without disclosing trade secrets
Quality data often contains more than just technical values. Photos can show faces, license plates, or customer orders; documents contain names and prices; support cases reveal internal processes. The WKO recommends considering data protection, personal and trademark rights with every use of AI. For an initial test, anonymized or synthetic examples are sufficient. Remove anything that is not necessary for the specific inspection task.
Also check where a tool stores inputs and who has access. A public online model is not automatically suitable for internal documents. Internal policies should specify permitted systems, data classes, retention and deletion. Employees need a clear explanation, not just a link to lengthy terms of use. The Chamber of Labour points out that data protection and co-determination rights also apply when using AI in the workplace.
Measurable criteria instead of gut feeling
A pilot needs a baseline. How many parts are inspected manually today? How often is an error overlooked? How long does a sample check take? Only after that can a comparison be meaningful. Measure at least hit rate, false alarms, processing time and the share of unresolved cases. A system that generates many warnings but finds hardly any real errors does not save work.
Test with data that occurs in everyday operations: different lighting conditions, various suppliers, seasonal fluctuations and rare defect types. A model trained only on ideal examples may perform well in the lab and fail in practice. Record the test date, data source and system version. This makes it traceable why a change improved or worsened quality.
Organize human sign-off properly
The final decision should be in the hands of a qualified person. This person must be able to understand what the AI warning is based on and must be able to override it. A simple interface therefore shows not only "error" but also the measured value, the comparison range and the uncertainty. For image inspections, the original image should remain visible next to the annotation, unchanged.
Avoid anonymous dismissals. If a warning is dismissed, a short reason should be selectable: false alarm, tolerance allowed, recheck to follow, or data unusable. This feedback helps improve the inspection logic. However, it must not become covert performance monitoring of individual employees. The Chamber of Labour (AK) warns against disproportionate surveillance and calls for transparency for algorithmic systems in the workplace.
Practical example: metalworks in Styria
A medium-sized metal company receives daily photos and measurements from final inspection. In the pilot, an AI system sorts images into “probably unremarkable” and “please check.” The quality supervisor checks every red and yellow alert. After four weeks it becomes clear: scratches are reliably detected, but reflective surfaces cause many false alarms. The team changes the lighting and adds more examples. Only after that is the pilot extended to a second product line.
What matters is what does not happen: the system does not automatically block any delivery and does not evaluate any person. An error is treated as a process problem, not as an assignment of blame. The findings feed into an improved work instruction and the training of new colleagues. In this way AI increases the company’s capacity to learn, rather than displace expertise.
AI Act and AI competence in the workplace
Since February 2025 the AI Act requires an adequate level of AI competence for people who work with AI systems. The WKO therefore recommends training, documented responsibilities and general rules. For a quality inspection this means: employees understand the purpose of the system, know typical error patterns and know when they must escalate. A certificate alone does not replace this practical onboarding.
Also check whether the application falls into a sensitive area. Quality control of a workpiece is different from a system that evaluates employees based on defect counts. As soon as work performance, behavior or access to opportunities are affected, the requirements for transparency, human oversight and co-determination increase. In case of doubt, the data protection officer or the works council should be involved early.
Checklist for a safe start
- Which specific deviation should be detected?
- Which data are actually required for this?
- Who reviews each uncertain result?
- Which metrics indicate benefit and false alarms?
- How are corrections and versions documented?
- Which data must not be entered into the tool?
- When will the pilot be stopped or expanded?
Quality is more than error detection
A sound inspection process looks not only at the outcome but also at the path that led to it. Ask whether the data were complete, whether the data-capture conditions were appropriate, and whether an alert reached the right place in time. A model can operate mathematically correctly and still be useless in the process if the alert is only read after delivery. Therefore define response times and escalation paths.
Human work also deserves a quality definition. The person conducting the inspection must have time for follow-up questions, be able to use appropriate measuring tools, and be allowed to make a decision without pressure. If every deviation is treated as a personal fault, errors will be concealed and training data will deteriorate. A fair error culture therefore indirectly improves the AI as well. This corresponds to the Chamber of Labour's basic principle that technical monitoring must respect the dignity and rights of employees.
Documentation for traceability and learning
Keep a short inspection log: date, product or case, system version, result, human decision, and, if applicable, correction. For sensitive procedures a pseudonymised identifier is sufficient. The log should not track every movement but should show whether the procedure is reliable. After one month the team can use a few examples to decide whether additional data, a new tolerance, or a different work instruction is needed.
If an error causes harm, clear responsibility is especially important. Secure the relevant data, stop affected outputs, and inform the responsible authority. Do not alter the original AI alert afterwards; instead, add the human assessment. That way it will be possible later to explain what information was available and why a decision was made. Such an approach supports internal audits and builds trust with customers.
Jobs change, expertise remains
With AI in quality control, work often shifts from pure visual inspection to root-cause analysis, process design, and communication. An experienced specialist will not become redundant but will review difficult cases, improve criteria, and train others. For newcomers to the profession there are learning opportunities when they get to know both the craft and the system's limitations. The AMS recommends in its labour market information combining digital skills with professional expertise.
Therefore plan training concretely. A short module can explain how a false alarm arises, what data quality is required, and how a decision should be documented. Afterwards each person should jointly work on a real, anonymised case. The learning effect is greater than from a mere tool demonstration. Record open questions and forward them to the responsible management or the provider.
When a pilot should be stopped
Not every experiment deserves an extension. End the pilot if the benefit is not measurable, false alarms remain persistently high, data cannot be processed in a legally compliant way, or employees can no longer realistically provide the human approval. A stop is not a failure but a responsible decision. Document what was learned and retain only the data for which there is a clear purpose.
If the pilot is expanded, do so step by step. Add a second product group or another site only once criteria and roles are stable. Check whether regional differences, different machines, or new lighting conditions change the results. Approval for the Vienna site is not automatically approval for an operation in Tyrol. Quality assurance always remains context-dependent.
Also include the people who work with the results every day. They know which exceptional cases occur in practice and which warning is particularly critical. A short monthly exchange is often enough: which message was helpful, which arrived too late, which rule needs to be clearer? This feedback keeps the procedure alive and prevents an old model version from quietly becoming the standard.
For customers, reliable performance is what ultimately counts. Therefore communicate only what the inspection can actually deliver. AI support is not a seal of quality and not a guarantee of error-free results. It is a tool within a responsibly managed process. Those who disclose this boundary strengthen credibility and create a good foundation for further digitization.
A final practical tip: write the system’s most important limitation directly into the work instruction. For example: "The AI pre-sorts images; final approval is granted exclusively by the quality specialist." This sentence prevents misunderstandings during stress, shift changes, or holiday cover. It reminds everyone involved that technology supports, while expertise, care, and fair accountability set the standard.
This creates a verifiable, human-accountable process that remains robust even with new tools.
Even after launch, the inspection remains a leadership task. Regularly discuss whether warnings are clear, whether staff have sufficient time for checks, and whether customer questions can be answered quickly. Update the work instruction as soon as the product, machine, or data source changes. A short quarterly review meeting prevents a once sensible pilot from quietly becoming an unexamined automatic process. The results should be accessible to all affected roles so that shift changes and cover arrangements can function without loss of information.
Conclusion: AI makes quality visible, responsibility remains human
AI quality control can ease the burden on Austrian businesses if the use case is narrowly defined and tested in real-world everyday conditions. The safe procedure is: minimize data, measure criteria, make uncertainty visible, have a human take responsibility for every approval, and learn from false alarms. Those who take this approach gain more reliable processes and simultaneously strengthen professional competence within the team.
