When operators in an older factory still need USB drives to operate machinery, maintenance personnel can only call the shop floor for status updates, and shift handover information is scattered across paper records, Excel files, and messaging apps, managers may instinctively think they need to replace equipment and install sensors before adopting AI. However, factories do not have to wait until all equipment has been replaced to address immediate challenges. They can begin by identifying tasks that are repeated every day, supported by accessible data, and can still be stopped by personnel before an error causes damage.
The first step in bringing AI into an older factory does not need to be fully autonomous operations or a factory-wide transformation. Start with either exception reporting or shift handovers, and clearly define what data is required, what AI should produce, who is responsible, and which situations must be escalated to a manager. Any action involving maintenance dispatch, production shutdowns, or the assignment of responsibility must be confirmed by authorized personnel before the process can continue.
Start with a Repetitive Process That Has Accessible Data
According to TechOrange, U.S. startup Harmoni focuses on existing factories, placing tablets next to machinery and using AI software to organize shop-floor work. Operators can use the same tablet to clock in, view work instructions, track work orders, and report problems. Harmoni’s AI product, HAL, uses data collected through the tablets to help personnel investigate why machinery is slowing down, submit maintenance requests, and hand over production issues between shifts. Harmoni’s goal is to reduce repetitive administrative work for shop-floor personnel, not replace them.
The lesson for Taiwan’s traditional manufacturers is not to purchase the same equipment and copy the approach exactly. Instead, they should first identify a process that shop-floor personnel handle every day and that generates a data trail. In exception reporting, for example, operators must describe the situation, consult the SOP, verify the exception code, and then notify maintenance personnel or a manager. Shift handovers require personnel to summarize unfinished work orders, equipment conditions, and items the next shift needs to monitor. Both types of work have clear starting points, and shop-floor personnel know who should ultimately handle the issue, making it easier to verify whether the information organized by AI is complete.
When selecting the first process, assess four factors individually: whether the task recurs every week, whether the necessary data is available, whether personnel can review the AI-generated output, and whether the process can be stopped before an error causes an impact. If a process spans multiple systems but no one can clearly explain where the data is stored, organize the process and clarify responsibilities before rushing to connect it to AI.
Organize Searchable Data Before Expecting AI to Provide Answers
Whether AI can provide useful answers depends on whether sufficient shop-floor data is available for retrieval. If data about machinery, personnel, and operating processes remains scattered across different locations, AI will only have a partial view of the situation. Even if an answer reads smoothly, it may overlook newly updated operating procedures, exceptions left over from the previous shift, or restrictions already documented in maintenance records.
The implementation team does not need to organize data for the entire factory at the outset. Instead, it should first inventory the SOPs, maintenance records, exception codes, work-order fields, and handover forms required by the target process. Each document should specify its version, responsible department, and update frequency. Older versions should also be clearly marked to prevent AI from retrieving conflicting procedures. The team should then test the system using questions that personnel actually ask on the shop floor, such as which procedure applies to a specific exception code or which fields are still missing from a handover. If the available data is insufficient to answer a question, AI must explain what information is missing rather than constructing an answer on its own.
The acceptance process can also be straightforward. The team can select a set of common questions and review which sources AI retrieves, which versions it cites, and how it responds when information is insufficient. When problems are identified, the team should correct the documents and fields before adjusting the prompts. If even shop-floor personnel cannot find the current version, the first step is to assign a document owner—not ask the model to guess its way around a document management problem.
Break Exception Reporting into Work Steps and Leave Critical Decisions to People
Once a process has been selected, document each step as a work card. For exception reporting, for example, the operator first provides the machine number, exception code, current conditions, and supplementary details. The AI Agent then prepares a summary, retrieves the relevant procedures, and identifies any missing information. Once all required information is available, the shop-floor manager confirms the exception category and course of action. If the case involves maintenance dispatch, a production shutdown, or the assignment of responsibility, the process must stop until an authorized person approves it.
This design clearly separates the responsibilities of the AI Agent from those of shop-floor personnel. The AI Agent can organize information, validate fields, and prepare the materials required for the next step. However, managers remain responsible for any decision that could affect production. If the AI Agent finds two procedures that conflict with each other, the workflow must not select one and proceed on its own. Instead, it should send the conflicting content to the document owner for confirmation.
EgentWrX can connect recurring tasks into workflows and add human approval before moving to the next step. After the AI Agent completes the initial data preparation, the process pauses until personnel review and approve it. The implementation team must clearly define the input data, output, and person responsible for each step. It should also specify the situations that require escalation to a manager, such as insufficient data, conflicting procedures, or high-risk decisions, rather than relying on a vague instruction such as “contact the manager if there is a problem.”
Let AI Retrieve Information First, Then Decide Which Systems to Integrate
Modernizing an existing factory does not mean replacing every legacy system at once. However, the implementation team still needs to map how data flows, including which data is stored in on-premises databases, which remains in files on the shop floor, and which services require external connections. The team must also confirm which data AI is allowed to read, which systems are read-only, and who has the authority to approve expanded access.
EgentWrX allows companies to place documents, spreadsheets, and procedures into separate knowledge bases for retrieval by AI Agent. If the AI Agent cannot find sufficient supporting evidence, it will not provide an answer. The platform can also read on-premises databases through the company’s internal network. By default, this access is read-only, so the platform cannot modify the data. If a task requires access to an external service, administrators can approve connection domains individually. These mechanisms can manage the scope of access to knowledge and internal data. The implementation team must still assess the most appropriate integration method for shop-floor equipment and existing systems based on each factory’s environment.
During the first phase, keep the pilot small in scope, on the internal network, and read-only. Record which fields are frequently missing, how often AI cites the wrong version, why personnel revise the content, and at which approval step the process most often stalls. Once the team has confirmed that it can manage data quality and has clarified responsibilities, it can decide whether to add data sources or expand the scope of integration. AI should not be allowed to modify production systems simply because it produced the correct answer once during a demonstration.
FAQ
Must older machinery be connected to the network before AI can be adopted?
There is no need to connect every older machine to the network first. Begin by organizing the SOPs, work orders, exception records, and handover forms required by the target process into searchable data. The AI Agent can then help personnel retrieve and organize the information. Once the process is stable, assess which equipment data is worth integrating.
Which should come first: exception reporting or shift handovers?
Start with the process that has more complete data, clearly defined ownership, and errors that can be stopped in time. If exception codes and maintenance records already follow a standardized format, exception reporting may be the better starting point. If information is still passed along verbally, begin by defining the fields required for shift handovers.
Can we pilot an AI Agent if our SOPs are still incomplete?
A small-scale pilot is possible, but the AI Agent must be restricted to answering solely on the basis of existing data. When it cannot find supporting evidence, it must stop and explain what information is missing. During the pilot, the team must also assign document owners and remove outdated content. The model must not be allowed to fill in missing procedures on its own.
Which steps must retain human approval?
Maintenance dispatch, production shutdowns, the assignment of responsibility, and other decisions that affect production must be approved by authorized personnel. An AI Agent can organize information and make recommendations, but whenever data is insufficient, procedures conflict, or a high-risk situation arises, the process must stop and wait for human confirmation.
How should we determine whether the first phase is worth continuing?
Review missing data, cited document versions, the reasons personnel revise the content, and the steps where the process frequently stalls. Do not judge the pilot solely by whether the answers sound fluent during a demonstration. If the team can identify recurring problems, correct the data sources, and clearly assign approval responsibilities, it can then consider expanding the scope of data and processes.
References
- Modernizing Old Factories Instead of Building New Ones: How Startup Harmoni Is Using Tablets and AI to Transform Traditional U.S. Factories — TechOrange reports on Harmoni’s approach to modernizing existing factories and its shop-floor applications.
