When a component’s dimensions change, the parameters and production arrangements for dozens of downstream machines may need to change as well. Senior engineers know where they can accelerate the process and where adjustments must not be made. New employees, however, often have only a few spreadsheets and operational records scattered across multiple locations. Even if the company introduces an AI Agent, it may be able to read those files without knowing which rule it should follow when making a decision.
Companies can begin by selecting one task and documenting how it is actually performed as a set of rules that employees can understand, the team can test, and managers can approve. They can then decide which steps to assign to an AI Agent. Processes and data do not need to be fully standardized before a project can begin, but the initial scope should be narrow. Start with a frequently performed task governed by relatively clear rules. Document how experienced employees make decisions, validate those rules against real cases, and then expand gradually.
Organize One Workflow First—There Is No Need to Wait for Company-Wide Standardization
Business Next cited Siemens’ industrial AI use cases, explaining that a Digital Twin incorporates dimensions, materials, equipment parameters, and actual production data. Before metal enters the factory for machining, engineers can simulate which cutting tool to use and how to set the spindle speed and table feed rate. If a design exceeds the machine’s operating limits, the manufacturing team can also flag the issue while the design is being revised.
This case offers a clear lesson for companies preparing to adopt AI: if AI is going to participate in a workflow, it must be given the data and rules actually used on the shop floor. PepsiCo has also tested machine locations, production-line flows, and equipment configurations in a virtual environment to determine whether proposed adjustments would create production bottlenecks. These simulations are possible because the team has already organized equipment and production data into a format the system can use.
However, companies do not need to wait until every process across the organization has been documented before starting a project. Doing so can cause the scope to grow continuously until even the first step becomes impossible to take. Instead, choose a task that recurs every week and has clearly definable inputs and outputs, such as compiling information for sales quotations, conducting a preliminary review of procurement bid comparisons, classifying quality-control anomalies, or assigning customer service cases.
In the first phase, clarify four things: what information employees receive, what conditions they use to make decisions, what exceptions they may encounter, and what output format they must deliver. If these four questions cannot be answered clearly, interview the people who actually perform the work before rushing to select a model.
Turn Veteran Experts’ Experience into Testable Rules
An SOP usually records operating steps, but it does not necessarily explain the reasoning behind a decision. Consider a sales quotation, for example. A document may simply state, “Confirm the delivery date and submit the quotation for managerial approval.” Experienced employees, however, will also examine the customer’s payment history, raw-material price fluctuations, minimum order quantities, and whether special specifications require confirmation from engineering. If these conditions are communicated only through verbal reminders, an AI Agent may still overlook them the next time it encounters a similar case.
When documenting this expertise, use real work scenarios to interview experienced employees. In addition to asking, “How do you handle this?”, follow up with questions such as, “Under what circumstances would you change your original decision?”, “What should happen if a field is missing from the data?”, and “Which cases must always be returned?” Then organize the answers into input data, decision rules, exceptions, and output formats. Prepare test cases covering correct outcomes, incorrect outcomes, and situations in which the available data is insufficient.
The team should not merely check whether the AI Agent produces an answer. It must also verify which rule the AI Agent used as the basis for that answer. If the data is insufficient, the AI Agent should stop and ask for clarification rather than make assumptions.
In EgentWrX, teams can organize this standardized approach as an AI Skill, ensuring that the AI Agent applies the same set of rules whenever it encounters the same situation. If employees want to roll out a skill they have created across the company, it must first undergo a back-end review. The process owner and experienced operational staff should review the skill together, pilot it within a single department, document any exceptions that have not yet been addressed, and then decide whether to make it available to additional business units.
Organize Data by Type, and Keep People Responsible for Confirming Critical Actions
Data standardization does not mean placing every file into a single folder. Policies, operating instructions, and Q&A content should be organized by topic. Semantic search can be used when the task requires finding similar content across a large collection of documents. Fields such as part numbers, prices, dates, and statuses should retain their original tabular structure to support subsequent filtering and verification.
Every dataset should have a designated owner, with its version, applicable business units, and update procedures clearly documented. Otherwise, the AI Agent may retrieve both old and new versions of the same rules without knowing which one to apply.
The EgentWrX knowledge base supports three data types: Wiki, vector, and table. Teams can also configure the permitted scope of data usage, similarity thresholds, and the minimum number of citations required. If the system cannot find sufficient supporting evidence, it will refuse to answer. The implementation team can test search results using real work-related questions—for example, by asking which guideline should be consulted for a particular quality-inspection anomaly—and then verify that the response cites the correct version and provisions. If the necessary information cannot be found, add the missing sources or reorganize the categories rather than asking the AI Agent to guess.
Even after the data and rules have been properly organized, any work involving process parameters, scheduling, or changes to on-site equipment should still be reviewed by a designated engineer. Companies can divide the workflow into data preparation, solution recommendations, engineering validation, and on-site deployment. After completing the preliminary work, the AI Agent should pause and wait for an engineer’s approval. Only after confirmation should the results proceed to the next stage.
The Siemens report also noted that some industrial AI systems currently generate recommendations, after which engineers modify the programs and deploy them. Dynamic scheduling that combines an AI Agent with a manufacturing execution system (MES), however, still requires further validation. When accepting the first project, companies should therefore evaluate not only whether the system can find the required information and apply rules consistently, but also whether it stops when it encounters an exception and whether responsibility for each step has been clearly assigned.
FAQ
Can we launch an AI Agent pilot before our processes are standardized?
Yes. Start with a frequently performed task that follows relatively clear rules and allows errors to be intercepted promptly. Interview the people who actually perform the work, document the input data, decision criteria, exceptions, and output format, and then test the workflow using real scenarios. There is no need to wait until every process across the company has been fully documented.
What should we do if veteran experts can provide the answer but cannot clearly explain the rules?
Use specific cases during the interview. Ask them to compare two cases that appear similar but were handled differently, and explain which piece of information changed their decision. Then ask how they would respond to missing data, special customers, or equipment anomalies. Finally, turn their answers into testable rules and exceptions.
Do all paper records, Excel files, and documents need to be moved into the same system first?
No. Begin by identifying the information actually used in the target workflow. Policies and instructional documents, content that requires semantic search, and tables with fixed fields should be organized separately according to their characteristics. Assign an owner, version, and scope of application to each dataset, then use real questions to verify that the correct supporting information can be found.
Which steps should not be assigned directly to an AI Agent?
Any step involving process parameters, dynamic scheduling, equipment operation, or formal external commitments should first be approved by a designated person. An AI Agent can help organize information and propose solutions, but when the workflow reaches engineering validation or managerial approval, it must pause. Deployment or formal external delivery should proceed only after the results have been confirmed.
References
- Breaking Down Siemens’ Digital Twin Success Stories: From an Aircraft Pump to PepsiCo, Virtual Trial and Error Is Possible—but Three Things Must Come First — Business Next’s overview of Siemens’ industrial AI and Digital Twin use cases and the prerequisites for implementation.
