Manufacturing AI: connect the information, respect the controls
Useful industrial AI begins with a concrete information problem and a disciplined boundary between advisory work and machine control.
Start with a decision people already make
Manufacturing AI can support work around production without taking over machine control. Consider document retrieval, maintenance preparation, quality-record organization, or analysis of approved historical data. Choose a specific decision and identify the information a person currently assembles to make it. This reveals where interpretation or preparation may help and where established controls must continue to determine behavior.
An illustrative maintenance-planning workflow might combine approved equipment documentation, historical work orders, and a current issue report into a review packet. The packet identifies relevant references and questions for a qualified technician. It does not diagnose a machine with certainty, issue a control command, or replace a lockout procedure. State the intended use directly so a helpful summary is not mistaken for an approved operating instruction.
Map business systems and operational technology
Business information may live in an ERP, a maintenance system, a quality platform, documents, and spreadsheets. Operational technology has different constraints around availability, timing, equipment, and the physical process. NIST SP 800-82 Rev. 3 treats OT security in the context of those operational requirements. Use that distinction when mapping the workflow and deciding where data can be accessed safely.
Record each source, its owner, the permitted interface, and the direction of data movement. Include the units, timestamps, identifiers, and freshness required to interpret the data. An apparently simple “machine status” field can be ambiguous if its sampling time or operating context is missing. Domain expertise is necessary to decide what the data means and which combinations support a valid conclusion.
Choose an integration boundary deliberately
OPC UA provides an industrial interoperability framework with information modeling and security capabilities. Its presence does not mean any application should receive unrestricted access to equipment. Work with the responsible controls and IT teams to identify an approved interface and a constrained data path. The specific topology, authentication, and permissions depend on the site's architecture and requirements.
For an initial analysis workflow, consider a reviewed, read-only data path into an appropriate environment. Decide how data is filtered, validated, and made available without introducing an uncontrolled route into operational systems. Any proposed write path needs a separate engineering assessment and change process. An agent's ability to invoke a tool is not evidence that invoking it is appropriate for a production asset.
Preserve deterministic control and safety functions
Do not position an LLM as a replacement for deterministic machine-safety controls. Safety-related functions, interlocks, and established protective procedures belong within the site's engineered control and safety systems. An AI-generated recommendation requires evaluation in its intended operational context. Keep the authority to change machine behavior with the people and systems responsible for that function.
This distinction should appear in the user interface and operating procedure. Label generated analysis as advisory where that is its approved role. Show its sources, uncertainty, and data age. If the data is stale or a required signal is absent, the output should make the limitation visible. A confident narrative can be particularly misleading when the underlying system has changed since the available record was produced.
Evaluate with operational specialists
Develop acceptance cases with the people who understand the process: operations, maintenance, quality, controls engineering, and IT as appropriate. Include mismatched units, incomplete records, changed equipment identifiers, and misleading correlations. Check whether the workflow preserves the distinction between a measured observation and an inferred explanation. A summary that loses that distinction can send a team toward an unsupported action.
Measure usefulness in the actual decision process. Does the packet retrieve the correct approved reference? Does it make missing information easier to identify? How much review is required before use? Avoid inventing uptime, scrap, productivity, or savings improvements. Those claims require a defined baseline, controlled measurement, and supporting records. A pilot should first prove the bounded information task it was designed to perform.
Plan ownership and change control
Assign owners for the data interface, interpretation rules, model-assisted output, and operational approval. Document how equipment changes, schema changes, and new procedures affect the workflow. Maintain a way to disable the AI-assisted component without weakening the established operating process. This makes the integration more understandable to the next engineer and reduces reliance on the person who built the demonstration.
An implementation partner should be able to explain the proposed boundary and identify where specialist review is required. Begin with feasibility and source access, then a controlled pilot, then measured acceptance. The result should be a useful connection between business information and industrial expertise, with clear evidence about what the system has actually been validated to do.
The source record
Sources & editorial notes
Primary documentation checked September 29, 2026. Our implementation recommendations are editorial analysis. Illustrative workflows are proposed examples, not completed client case studies or performance claims.
- Guide to Operational Technology Security, SP 800-82 Rev. 3 NIST
- OPC Unified Architecture OPC Foundation
- Control your dot OpenAI
- Connect computers and apps to your dot OpenAI
Found something that needs updating? Contact the editorial team with the passage and a supporting source.