When an abnormality alert appears on a production line, engineers may need to review the machine control interface, Manufacturing Execution System (MES), and maintenance records separately before deciding whether to stop the equipment. When industrial AI is introduced, managers usually ask two questions first: Once AI can read shop-floor data, could it directly alter equipment operations? And can engineers still intervene before an action is carried out?
Data access and equipment control do not have to be bundled together. Companies can allow an AI Agent to read the data required for analysis without permitting it to modify databases or directly control the production line. After the AI Agent organizes the data and presents possible causes and recommended actions, the workflow can pause for review by a designated engineer. Companies should begin by making data access read-only, requiring human approval before any operational action, and retaining complete process records. They can then use the results of the pilot to determine whether to expand the scope of automation.
If All the Data Is Connected, Why Can’t AI Understand It?
Business Next reports that many factories have been operating for decades and simultaneously use sensors, controllers, energy management systems, and process systems from different eras and vendors. Each piece of equipment comes with its own management tools, so over time, data becomes scattered across different software platforms and databases. Industrial platforms sit between these systems, collecting operational technology (OT) data from the shop floor, connecting it to information technology (IT) systems, and making the data available to AI and analytics tools.
However, the ability to connect to a system does not mean AI can understand its data. Some equipment has been in service for more than 10 years and may not have retained all the data now required for AI analysis. Even when data is available, equipment names, time formats, and process definitions may vary across systems. For example, if the same machine is identified by different codes in maintenance records and the MES, the team must first establish clear mapping rules. Otherwise, the AI Agent may be unable to connect related abnormalities across different records.
Teams should therefore avoid trying to process data from the entire factory at once. Instead, they should select one abnormality-handling workflow, identify the screens, records, and fields engineers currently review, and then verify whether each data point exists, whether timestamps can be aligned, and whether naming conventions are consistent. Managers can require the team to deliver a list of data sources and a field-mapping table. If critical records were never retained, the team should first improve equipment digitalization and recordkeeping rather than rush into model testing.
Clearly Define Which Data AI Can Access
The need to connect manufacturing data does not mean that all information must be sent to a public cloud. The report notes that companies can keep production data inside the factory for processing and analysis. For organizations that place a high priority on protecting proprietary processes, the first step is to distinguish among data stored in internal databases, files located on-site, and data that requires external services. They can then determine where each category of data may be accessed and whether it is permitted to leave the company.
During the initial pilot, the AI Agent should access only the necessary data through the company’s internal network, with database connections set to read-only by default so that no content can be added, modified, or deleted. If the AI Agent needs to search external documents or call external services, the IT team should not enable unrestricted network access. Instead, it should add only the domains required for legitimate business purposes to an egress allowlist. This separates the permission to obtain data for analysis from the permission to modify operational systems, preventing execution privileges from being granted merely for the convenience of accessing data.
EgentWrX provides a way to connect to a company’s internal network, allowing an AI Agent to read the company’s own databases. This integration permits read access only and does not allow write operations. The system denies all outbound connections by default, after which administrators can add approved domains to the egress allowlist individually. When evaluating a pilot, the IT team should verify which identity the connection will use, which database tables it can access, how far back it can query historical records, and which external domains have a genuine business need. The relevant data owners should confirm the scope before access is enabled.
Where Should the Workflow Pause After AI Recommends an Action?
Using an increase in the component-throw rate on a printed circuit board production line as an example, the report explains that industrial AI can analyze nozzle conditions, production speed, and abnormality frequency together, identify possible causes, and recommend reducing the speed first. However, an incorrect AI assessment could affect a production line worth tens of millions of dollars or even endanger personnel. AI must therefore not receive direct control of equipment after completing its analysis.
Companies can divide abnormality handling into four stages: reading data, analyzing causes, recommending actions, and executing operations. A human approval gate should be placed before execution. Once the AI Agent completes the first three stages, the workflow pauses and waits for an engineer’s approval. The approval screen should also show which data the AI Agent referenced, which possible causes it identified, and what actions it recommended. Engineers need enough information to verify the analysis; simply presenting “approve” and “reject” buttons is not sufficient.
Companies must also clearly define three aspects of human approval: who is authorized to approve an action, what information must be reviewed during approval, and which conditions require rejection. For example, a request can be rejected and referred to an engineer if timestamps do not align, critical sensor readings are missing, or the recommendation involves safety-related equipment. The pilot team should first select one abnormality-handling workflow, document the approvers, evaluation criteria, and rejection conditions in the task rules, and then run simulations using existing historical records.
Which Records Must Be Retained to Investigate Problems Later?
Even after engineers retain final decision-making authority, companies must be able to answer four questions: What data did the AI Agent query? What analysis did it produce? Which engineer approved the action? Was the workflow ever rejected? If the system retains only the final result, cybersecurity, quality assurance, and shop-floor managers will be unable to reconstruct the data used by the AI Agent or the process through which the engineer reached the decision when an incorrect assessment occurs.
Before the pilot begins, companies should therefore define the records that must be retained, including data queries, analysis results, human approvals, reasons for rejection, and subsequent execution status. They should also establish the scope of retention and procedures for spot checks. EgentWrX audit logs can be exported, and a hash chain can be used to verify the integrity of the records. The system also records when an administrator downloads audit logs. These mechanisms retain the information required for future reviews, but they do not guarantee that the AI’s analysis is correct. Cybersecurity or audit teams must still conduct regular spot checks, while shop-floor teams should continually assess whether rejection conditions need to be adjusted.
An initial industrial AI pilot can focus on consolidating essential equipment data, historical records, and operational experience so that an AI Agent can help engineers organize abnormality information, while equipment operations remain within existing control processes. Once the team confirms that data quality, approval rules, and recordkeeping methods all meet requirements, it can consider opening up the next stage of the workflow. If the organization cannot clearly explain who has approval authority or which records must be examined after an error, it should not expand the AI Agent’s execution privileges.
FAQ
Can industrial AI still be implemented if legacy equipment lacks complete data?
Companies can begin with a pilot in a workflow where the data is relatively complete; there is no need to wait until all factory equipment has been upgraded. The team can first identify missing sensor readings, maintenance records, and timestamp fields. If critical data does not exist, it should improve equipment digitalization and recordkeeping first, rather than mistakenly attributing problems caused by insufficient data to errors in the model’s judgment.
Which tasks can be assigned to an AI Agent, and which decisions must be made by engineers?
During the initial stage, an AI Agent can read data, organize the circumstances before and after an abnormality, and present possible causes and recommended actions. Actions involving shutdowns, process parameter adjustments, or the operation of safety equipment should be approved by an engineer with the appropriate authority and responsibility. Companies must also document rejection conditions clearly within the workflow instead of relying solely on verbal reminders.
Must factory data be sent to a public cloud?
Not necessarily. Companies can evaluate whether to read and analyze data directly within the factory or over the corporate intranet. Before launching a pilot, they should distinguish between internal databases and external services, use read-only access for internal connections, and approve external domains individually. They must also confirm whether data will leave the company and who will manage connection permissions.
Which workflow should be selected for the first industrial AI pilot?
Companies can begin with an abnormality-handling workflow in which engineers already follow clearly defined procedures, historical records are available, and the process can pause for confirmation before an operational action is executed. The team should first map the data sources, analysis steps, approvers, and rejection conditions, and then run simulations using historical records. It should not begin by asking an AI Agent to control equipment directly.
References
- Break Down Data Silos Before Deploying AI on the Factory Floor: How Advantech and Schneider Electric Are Pursuing IT/OT Industrial Platform Opportunities — Business Next, Kuo Tsai-hua, October 13, 2026.
