When factories prepare to implement AI, discussions often begin with equipment, sensors, models, or platforms. In practice, however, the real bottleneck is often the process itself. Different shifts may handle the same anomaly differently; machines may generate data, but part numbers, timestamps, and error codes do not align; and experienced technicians may know how to adjust parameters but struggle to explain the basis for their decisions. If these issues remain unresolved, allowing AI Agent to participate in decision-making simply transfers an ambiguous process into a new tool.
We believe manufacturers should implement AI in the order of “process, data, AI.” Process documentation must also go beyond producing a polished flowchart. Companies first need to clarify who is responsible for each step, what evidence is required, which exceptions should cause the process to stop, and who provides final approval. They should then organize the data needed for the selected process rather than trying to connect all factory data from the outset.
How Should You Choose the First Process?
An article originally published by NAIP News and republished by Business Next argues that companies should first eliminate process bottlenecks, then build data collection and integration capabilities, and only afterward use AI to support decision-making. It also reminds companies that purchasing hardware, packaged software, or AI algorithms does not mean the transformation is complete. If legacy equipment lacks standardized data interfaces and intelligent monitoring, companies may struggle to maintain their place in global supply chains.
When executing a project, the first question should not be which type of AI is the most powerful. Instead, identify which process is most valuable to standardize first. Companies can begin with tasks that are performed frequently, governed by relatively clear rules, and produce results that someone can validate. Examples include the initial classification of quality-control anomalies, the organization of equipment maintenance data, or the compilation of documents before procurement price comparisons. If a manager must decide the approach on an ad hoc basis every time the process is performed—and even the input data has no consistent source—the project team should first standardize the operating procedure rather than rushing to have AI Agent handle subsequent tasks.
After selecting a process, document the current approach step by step as a repeatable workflow. At a minimum, specify the input data, expected output, exceptions that should cause the process to stop, the person responsible, and the steps requiring approval. For example, the process can be paused when an amount exceeds a specified limit, conditions are not clearly defined, the result will affect machine settings, or content is about to be sent to a customer. It can then resume after the appropriate personnel have reviewed and approved it. EgentWrX provides task and workflow mechanisms and supports human approval before work is handed off between tasks. Companies can first turn one process into a work card, verify that someone is accountable for every stage, and then evaluate whether to expand its use.
How Organized Must Data Be Before AI Can Read It?
Legacy factory equipment may come from different brands and generations and may lack open communication protocols such as OPC UA. As a result, operational technology (OT) data can be difficult to integrate across equipment and facilities. Collecting “more data” alone is not enough to address this fragmentation. Unless the project team first standardizes fields, time references, and equipment identification methods, the same term may refer to different states, and values that appear identical may not be directly comparable.
However, companies do not need to wait until all their data has been organized before getting started. A more practical approach is to identify the minimum data required for the first process, including where the data comes from, its format, the department responsible for it, how often it is updated, its sensitivity level, and whether AI Agent is permitted to access it. The team should then spot-check several real records to confirm that timestamps, part numbers, equipment, and anomaly statuses correspond to events that actually occurred on the factory floor. If personnel cannot reconstruct what happened from the data, AI Agent will not have a reliable basis for making decisions either.
Data access should also be granted in tiers. Based on content type, companies can place organized documents, spreadsheets, and standards in a knowledge base and instruct AI Agent not to answer when it cannot find sufficient supporting evidence. Internal databases can provide read-only access through the company intranet, allowing AI Agent to query but not modify data. External connections should be enabled only after the company reviews and approves them individually. EgentWrX provides these knowledge retrieval, intranet read-only access, and external connection control mechanisms. During the first phase of a project, companies should grant access only to the data genuinely required by a single process and for which ownership and accountability are clearly defined. During the trial period, they should review every cited source and output. Once the results are proven reliable, they can gradually expand the scope of access.
How Can Experienced Technicians’ Expertise Be Captured Without Ending Up as Just a Stack of Documents?
The original article also notes that equipment calibration, troubleshooting, and machining parameter optimization may rely heavily on the decades of experience, intuition, and tacit knowledge accumulated by senior technicians. If an interview asks only, “How do you respond to this alarm?” the resulting documentation will usually capture the procedural steps but omit the evidence technicians actually use to make decisions—for example, which sound they listen for first, which segment of historical records they compare, and which conditions mean they should not continue following the standard procedure.
Companies can begin by selecting one common fault or calibration scenario and asking an experienced technician to recall the most recent real case. They can then document, in sequence, the decision criteria involved, the evidence reviewed, the actions taken, the exceptions encountered, and the stages at which another person must be consulted. Once the content has been documented, it should not immediately be published company-wide. Instead, a team in the same workspace can test it using real records. If different people still interpret the same content differently, the documentation should be supplemented with photos, field definitions, counterexamples, or explicit stop conditions.
Project teams can encode decision-making methods that need to be applied repeatedly as an AI Skill, enabling AI Agent to follow consistent rules whenever the specified conditions are met. If the AI Skill is to be deployed company-wide, it should first undergo a formal review. The project owner should also confirm whether new employees can use the same evidence to make an initial assessment and whether they know to stop and whom to consult when an exception arises.
FAQ
When implementing AI in manufacturing, should companies organize their processes first or collect data first?
First organize the selected process, and then collect the minimum data required for that process. If a process lacks consistent inputs, outputs, and accountable owners, even a large volume of data will not make it easy to determine whether the results produced by AI Agent are correct. Clarifying the steps and exceptions first reveals which data genuinely needs to be organized.
Can a company begin an AI project if its legacy equipment lacks standard communication interfaces?
Yes, but it should begin with a process that has clearly identified data sources and produces results that someone can validate. Connecting every machine in the factory should not be the first project. Companies can first inventory equipment interfaces, existing reports, and manual records, determine which data corresponds to actual events on the factory floor, and then decide what additional data to collect.
Is documenting experienced technicians’ expertise enough?
No. In addition to procedural steps, the documentation must clearly specify decision criteria, required evidence, exceptions, and stop points. Start with a recently resolved fault and ask an experienced technician to reconstruct the process step by step. Then have a team test the documentation using real records. If different people still interpret it differently, add photos, counterexamples, or field definitions.
How can a company determine whether the first phase is ready to expand to other processes?
First confirm that the data corresponds to events on the factory floor, that the basis for each output can be traced, and that the process stops for human approval when high-risk situations arise. Once these conditions can be maintained consistently, add more data sources or incorporate adjacent processes. Do not expand the scope simply because the solution performed well during a demonstration.
References
- Why Can’t Traditional Equipment Manufacturers Implement AI by Simply Buying Equipment? Four Stages of Transformation, from Process Improvement to Data Services — An article originally published by NAIP News and republished by Business Next, outlining a four-stage transformation path for traditional equipment manufacturers, from process improvement and data integration to AI applications.
