A vendor shows you how a Word work instruction can be turned into a guided assembly sequence in seventeen seconds. Eighty steps, images assigned, inspection criteria identified. The reaction in the room is always the same: Production planning staff calculate in their heads how many weeks of sequence creation this will save. What the demo doesn’t show is what happens next. Of the eighty steps, about sixty are usable, twelve need rework, and eight are incorrect—without it being obvious at first glance.
THE MOST IMPORTANT POINTS AT A GLANCE
|
This calculation is not an argument against AI in worker management. It is an argument for seeking its benefits in the right places. Deriving sixty useful steps from a document is a significant gain compared to manually recording eighty steps. That gain is lost only if no one identifies the eight incorrect ones.
This is precisely what determines whether an AI project in manufacturing will succeed. It depends not on the quality of the model, but on who reviews the proposal and where that review is documented. This article therefore addresses both: the four use cases that work reliably today, and the areas where a model has no place structurally within a managed manufacturing process.
If you’re still in the process of selecting a system, these fundamentals are essential. Criteria and an evaluation matrix can be found in the article on selection criteria for worker assistance systems. This article focuses exclusively on the AI component.
IN A NUTSHELL
|
|
FREE CHECKLIST How reliable is your production line?Seven principles of safe worker guidance as a checklist. The foundation without which no AI use case in manufacturing can succeed. PDF · Instant Download · No Spam |
Intelligent worker guidance is not a distinct system type, but rather a worker assistance system in which a model handles individual tasks. The dividing line can be precisely defined: Anything that can be formulated as a fixed rule belongs in the rule set. Anything that cannot be formulated as a rule is a candidate for a model.
An example of a rule set: After step 12, the screwdriver is enabled; otherwise, it is not. This is a condition; it does not require AI and must not have any, because it must be reproducible at all times. An example of a model: Identifying, from a sixteen-page PDF, which paragraphs describe work steps and which describe inspection criteria. There is no rule for this that works across all document templates.
| Level | What it does | Behavior in Unknown Cases | Who is responsible for the result |
|---|---|---|---|
| Rule-Based Worker Guidance | Guides through a human-defined process, interlocks, documents | Aborts or escalates; does not make anything up | The person who approved the workflow |
| AI-supported worker guidance | Generates suggestions: workflow structure, component assignment, anomaly alerts | Still generates an output that looks plausible but may be incorrect | The person who approves the suggestion |
| Automation | Performs the work step itself, without an operator | Stops if a tolerance violation occurs | Equipment responsibility, not operator supervision |
The third line is included intentionally because AI and automation are regularly discussed in the same breath. Intelligent worker guidance does not automate anything at the workplace. It supports a person who continues to perform the task. This is not modesty, but rather the reason why the question of responsibility remains solvable in the first place.
The following four use cases are currently in production at manufacturing facilities. The “Maturity Level” column represents a classification based on CSP’s project experience, not market statistics.
| Use Case | Input | What the Model Does | Who Decides | Maturity Level |
|---|---|---|---|---|
| Document Conversion | Existing instructions from Word, PDF, PowerPoint | Breaks down steps, assigns images, identifies inspection characteristics and target values | Production planning approves each step | Productive, high value |
| On-site image and component recognition | Camera view of the station | Identifies the component, variant, or assembly status and sets the process context | System is locked; operator can escalate | In production, environment-dependent |
| Anomaly alerts based on process data | History based on screwdriving, measurement, and test data | Flags trends that deviate from the established pattern | Quality assurance evaluates the alert | In production, requires historical data |
| Multilingual instruction output | Approved procedure in the source language | Translates step texts while retaining technical terms and target values | Second review by a linguistically competent person | In production; review required |
Two patterns are identical across all four cases. First, the benefit lies in the preparation or evaluation, never in the execution during the cycle. Second, every row features a person as the decision-maker. If this column were left blank, it would constitute evidence of a process that no one has verified—and in an audit, that is worse than having no evidence at all.
The most economically significant use case is the least spectacular one. Workflow creation is the most expensive component of any worker assistance system implementation, significantly more expensive than licenses and hardware. Based on CSP’s project experience, creating content for a station with 15 to 25 steps takes four to eight hours. For sixty stations, that amounts to half a person-year.
This is exactly where document conversion comes in. A model reads the existing instructions and generates a workflow structure from them. The point where expectations and results diverge is the quality of the source material. If you have well-maintained instructions with clear step numbering and labeled images, you’ll get a very good result. If you feed the system a PowerPoint created in 2013 with screenshots and handwritten additions, you’ll end up with a mess.
| Element | Output Quality | Why |
|---|---|---|
| Step Separation and Order | Good | Numbering and paragraph structures are usually consistent in documents |
| Image-to-step mapping | Good to average | Works for image captions, but becomes unreliable for standalone image galleries |
| Target values and tolerances | Average | Numbers are recognized, but the assignment to attribute and unit is prone to errors |
| Recognizing inspection characteristics as such | Average | Depends on whether the template distinguishes inspection steps linguistically from assembly steps |
| Marking safety-related steps | Poor | Relevance is rarely stated in the document; it lies in process knowledge |
| Add undocumented steps | Not possible | A model cannot generate anything that isn’t in the template. That is the most common fallacy |
The last two lines are the reason why a converted process flow must not be approved without review by someone from production. And they also serve as an argument for the order of the project: First, document the process flow at the workstation; then convert the existing documentation. Not the other way around. For information on maintaining and versioning the specifications themselves, see the article on digital work instructions.
THE MOST COSTLY MISCONCEPTIONThe assumption that AI replaces process maintenance. It replaces the initial data entry to a large extent. Ongoing maintenance remains entirely necessary, because every design change, every new variant, and every process adjustment must still be entered into the system and approved by a human. The consequence of overlooking this: The maintenance effort is not factored into the business case, no role is designated for it, and after eight months, the workflow in the system diverges from the workflow on the ward. From that point on, the system is bypassed, and the documentation becomes worthless. |
Image and component recognition solves a specific problem: Variant control requires context, and an order context from the MES isn’t always available. Components arrive in varying orders, rework parts are returned, and small production runs proceed without a consistent order reference. A camera at the workstation can bridge this gap by recognizing what is in front of it.
The limiting factor is the environment. Image recognition in manufacturing does not fail because of the model, but rather due to changing light conditions, cooling lubricant on the optics, reflections on shiny surfaces, and components that differ only in a hidden area. Training images must therefore come from the actual work environment, including the early shift in winter and the late shift in direct sunlight.
The second case does not analyze the current image but rather the process history. A classic threshold check issues an alert when a torque value exceeds the tolerance. A model based on the process history issues an alert when a trend deviates from the established pattern, even though every individual value is still within tolerance. This is the difference between an alert issued after the error occurs and one issued before it.
The practical benefit depends entirely on the quality of the alerts. A system that generates thirty alerts per shift will be ignored after a week—and after that, even the one alert that was correct will be ignored. Therefore, start with a low sensitivity setting and only increase it once every triggered alert has been verifiably evaluated. For the evaluation logic behind this, see the article on process data analysis in quality assurance.
A model that grants approval on a cyclical basis shifts responsibility to a point that cannot be questioned during an audit. That is why there is no AI there, but rather a rule.
Amadeus Lederle, Chief Technology Executive, CSP Intelligence GmbH
This section is the most important part of the article because it identifies the areas where AI structurally does not work in a guided manufacturing process. Not just not yet, but fundamentally not, because the requirements and the way it works are incompatible.
| Expectation | Reality | Consequence |
|---|---|---|
| The model grants process approval | An approval must be reproducible and justifiable. A probabilistic model is neither. | Approvals remain within the deterministic framework. AI provides insights, never decisions. |
| The model supplements missing process knowledge | What is not documented cannot be reconstructed. The model then generates plausible results rather than correct ones. | Recording procedures at the workstation remains mandatory, even with conversion. |
| The model identifies safety-critical steps | Safety relevance lies in experiential knowledge, not in the text of the instruction. | Critical steps are always marked by people. |
| Less effort required for process maintenance | Initial data entry becomes more cost-effective, but ongoing maintenance does not. | Define maintenance responsibilities and the approval workflow in writing before starting. |
| A note without a rating still has value | An unrated note increases the alert load and reduces attention to the next one. | Define a responsible role and a response time for each type of alert. |
Asking about these five points during a vendor meeting helps distinguish substance from a demo faster than any feature list. A vendor who identifies the limitations themselves has seen these scenarios in practice. One who gives a positive answer to all five has not.
AI projects in manufacturing rarely fail because of the model; they almost always fail because of the data. This isn’t just a cliché—it’s a concrete list of prerequisites that can be verified before the project begins.
| Data Type | Minimum Requirement | Typical Shortcomings in Existing Systems |
|---|---|---|
| Variant Master Data | Unique variant numbers, identical across all involved systems | Two systems, two number ranges, mapping via Excel |
| Inventory documents | A clear step-by-step structure, images linked to each step | Multiple template generations in parallel; images without captions |
| Training images | Footage from the real work environment covering all lighting and shift conditions | Studio shots or CAD renderings—not applicable in the workshop |
| Process history | A sufficiently long series with associations to components, stations, and tools | Values available, but serial number assignment is missing |
| Approval and version status | It is clear who approved which process and when | Process exists, but approval history is missing |
The fourth line is the most common roadblock. Process data is available in many companies, but it is not assigned to specific components. A model finds patterns in this data that it cannot trace back to a cause, and the analysis ends in a correlation without any possible course of action. The serial number as a unique identifier is therefore a prerequisite, not an option.
The legal landscape has shifted in 2026, and in a direction that is often misunderstood. It is the high-risk obligations that have shifted, not the EU AI Act as a whole. Several sets of obligations are already in effect and explicitly apply to operators—that is, manufacturing companies—as well as software providers.
| Scope of Regulation | Legal Reference | Applies | Implications for Worker Representation |
|---|---|---|---|
| Prohibited Practices | Art. 5 of Regulation (EU) 2024/1689 | Effective February 2, 2025 | Emotion recognition in the workplace is prohibited. Applies to camera systems in the workstation |
| Employees’ AI Competence | Art. 4 of Regulation (EU) 2024/1689 | Effective February 2, 2025 | Operator obligation: Anyone operating the system must be familiar with its functionality and limitations |
| Transparency Requirements | Art. 50 of Regulation (EU) 2024/1689 | Effective August 2, 2026 | Users must be able to recognize that an output originates from an AI system |
| High-risk, autonomous systems | Annex III, as amended by Regulation (EU) 2026/1744 | Effective December 2, 2027 | Annex III, No. 4 covers human resources management, including performance and behavioral evaluations |
| High risk in regulated products | Annex I, as amended by Regulation (EU) 2026/1744 | Effective August 2, 2028 | Relevant when AI, as a safety component, controls tool authorization |
| Employee participation | Section 87(1)(6) of the Works Constitution Act (BetrVG) | Unchanged | Time-stamped recording with operator identification is subject to co-determination, with or without AI |
Amending Regulation (EU) 2026/1744, known as the Digital Omnibus on AI, was published in the Official Journal on July 24, 2026, and entered into force on July 27, 2026. It postpones the applicability of the high-risk obligations under Annex III from August 2, 2026, to December 2, 2027, and for Annex I to August 2, 2028. The range of fines remains unchanged at up to 15 million euros or 3 percent of global annual turnover under Article 99(4), and up to 35 million euros or 7 percent for prohibited practices under Article 5. The Federal Network Agency will serve as the central supervisory authority in Germany.
WHEN YOUR SYSTEM BECOMES RELEVANT UNDER ANNEX IIIGenerally not: A system that manages the process, locks it down, and documents the execution for each component. The purpose is process reliability and documentation, not the evaluation of individuals. It likely does, however: As soon as the system generates performance evaluations based on individual employees—such as processing times or error rates per worker ID—and these evaluations are factored into decisions regarding task assignment or performance evaluation. Annex III, Section 4, specifies exactly this. Practical Implication: Specify in the works agreement that evaluations are process-based—that is, by station, variant, and tool—and not person-based. This resolves the issue of employee participation and, at the same time, keeps the system out of the high-risk category. A legal assessment of individual cases belongs in the legal department, not in a technical article. |
Before investing in an AI use case, you can determine in just one hour whether it’s viable. Five screening questions, each with a clear exclusion criterion.
| No. | Screening Question | Exclusion Criterion |
|---|---|---|
| 1 | Can the task be formulated as a fixed rule? | Yes, then it belongs in the rulebook. AI would be more expensive, slower, and harder to verify |
| 2 | Who approves the proposal, and where is the approval documented? | No designated role or location for documentation. This results in unverified evidence |
| 3 | Is the data available in the required quality—today, not after a project is completed? | Assignment to a component, station, or variant is missing. Without an assignment, there is no usable result |
| 4 | What is the cost of an incorrect suggestion that goes unnoticed? | If a safety or functional feature is affected, it requires a double human review. |
| 5 | What data leaves the company in this process, and where does it go? | No clear answer from the provider. For process and personal data, this is grounds for termination |
Anyone who has answered these five questions has practically made the investment decision. If the answer to Question 1 is “Yes,” the use case is not viable. If Questions 2 or 3 are not met, the use case is not ready. Question 4 determines the depth of the review, and Question 5 determines feasibility within the organization.
In CSP’s Manufacturing OS, the use of AI is deliberately limited to preparation and analysis. The PGX module guides the operator in a rule-based and deterministic manner, because traceability takes precedence over flexibility at this stage. The serial number serves as the consistent primary key across operator guidance, process data, and archiving—and this is precisely what makes analyses capable of identifying root causes.
A clarification is in order: This module does not replace an MES and does not make autonomous decisions on the production line. Anyone expecting AI that defines and approves the process itself will find something deliberately different here. The page on digital operator guidance with PGX provides an overview of operator guidance within the overall system. The article on AI-driven quality assurance in manufacturing discusses how AI itself functions in quality assurance—specifically in inspection and measurement data.
|
FREE DEMO See where AI comes into play and where the rules take overUsing a specific assembly process as an example: document conversion, context recognition, and the documented approval in between. No empty promises. About 30 minutes · Your process examples · No obligation |
Intelligent worker guidance refers to a worker assistance system in which an AI model handles subtasks that cannot be formulated as fixed rules. There are typically four tasks: converting existing work instructions from Word, PDF, or PowerPoint into structured workflows; recognizing components or assembly statuses via camera images; identifying anomalies in captured process data; and translating instructions into other languages. The execution of tasks in real time remains rule-based.
A rule-based operator assistance system follows a sequence defined by a human. It is deterministic: the same context, the same next step, reproducible at any time. An AI-supported system generates suggestions based on probabilities and can handle cases for which no rule exists. The trade-off is that the output is not guaranteed to be correct. That is why both approaches are combined in practice: AI for preparation and evaluation, and a rule set for execution.
To some extent. A model can structure an existing instruction, separate steps, assign images, and identify inspection criteria as such. It cannot generate a procedure that is not documented. This is precisely where the practical problem lies, because the procedure that actually works at the workstation is rarely fully documented in the work plan. Expect that a converted procedure will need to be reviewed by someone from the production floor before it can be approved. Nevertheless, the cost savings are still many times greater than those of creating a new procedure from scratch.
Technically possible, but not recommended in practice and legally risky. An approval decision must be reproducible and justifiable because it must be documented in the event of an audit or liability claim. A probabilistic model does not meet this requirement. Therefore, use AI as an alert provider, not as a decision-maker: The model flags an anomaly, a human makes the decision, and the decision is documented with a timestamp and identifier.
Not automatically. The key factor is the purpose. A system that exclusively manages and documents the process generally does not fall under Annex III of Regulation (EU) 2024/1689. Annex III, Section 4 covers employment and human resources management, including the monitoring and evaluation of employees’ performance and behavior. Therefore, as soon as the system generates personalized performance evaluations, it is likely to be classified as high-risk. These obligations apply as of December 2, 2027, pursuant to Amending Regulation (EU) 2026/1744. A legal assessment of the specific case should be handled by the legal department.
Three sets of requirements have been in effect for some time and are often overlooked by small and medium-sized enterprises. The prohibitions under Article 5 have been in effect since February 2, 2025, including the prohibition of emotion recognition in the workplace. The requirement for AI competence under Article 4 has also been in effect since February 2, 2025, and applies to operators—that is, your company—not just the software provider. The transparency requirements under Article 50 have been in effect since August 2, 2026.
Above all, unambiguous master data. A model can only distinguish between variants if variant numbers are maintained identically and unambiguously across the relevant systems. For image recognition, images are needed from the real-world work environment—not from a photo studio—including those taken under poor lighting conditions. For anomaly alerts, a sufficiently long process history is required with clear assignments to components, stations, and tools. If the assignment is missing, the model will identify patterns without a cause.
Ask three questions. First: Who approves the proposal, and where is the approval documented? Second: What happens in a case the model doesn’t recognize—does it stop, or does it make something up? Third: What data from my company leaves the premises, and where does it go? Anyone who cannot provide a concrete answer to any of these three questions has a demo, not a product.