Key findings

• Keep generative AI outside safety instrumented and direct machine-control loops.

• Preserve authoritative OT data and segment production networks from model services.

• Validate by product, machine, material, shift, and operating condition.

• Measure false negatives, scrap, downtime, and recovery, not only model accuracy.

• Do not automate physical action when a failure can harm people, product, equipment, or the environment.

Manufacturing use cases with practical value

An SME can use AI to retrieve approved work instructions, classify maintenance notes, summarize shift logs, extract supplier certificates, prepare nonconformance records, group defect descriptions, assist visual inspection, and compare scheduling scenarios. It can help planners understand constraints and help technicians find relevant procedures. These are support functions around production evidence, not replacements for machine safeguards, engineering approval, or operator skill.

Begin read-only. A cited answer from current work instructions is safer than generating a new procedure. A second-pass vision check is safer than automatically accepting product. A maintenance queue is safer than commanding equipment. A scheduling scenario is safer than changing the production plan. Choose a bounded line or product family with sufficient examples, stable measurement, and an owner from operations, quality, maintenance, or planning.

• Good pilots: instruction retrieval, supplier document extraction, shift summaries, maintenance triage, and quality-review assistance.

• Higher control: defect detection, predictive maintenance, production scheduling, and parameter recommendations.

• Exclude: safety interlocks, direct PLC writes, autonomous equipment restart, final disposition, and unreviewed process changes.

An OT-safe reference architecture

Collect approved data from the ERP, MES, QMS, CMMS, historian, or an OT data broker into a segregated analytics zone. Where risk warrants, use one-way transfer or tightly controlled gateways. An orchestrator retrieves only the machine, batch, product, and time window needed. The model returns a summary, classification, anomaly, or recommendation with source tags and uncertainty. Deterministic rules enforce operating envelopes and route results to an operator, quality engineer, planner, or maintenance lead.

The PLC, distributed control, safety instrumented, MES, QMS, CMMS, historian, and ERP systems remain authoritative. Generative services should not receive unrestricted OT credentials or write directly to control equipment. Record sensor and document sources, model and version, operating context, recommendation, human decision, and outcome. Design local fallback, queueing, and safe degradation for network loss. Test changes in a lab, digital twin, or isolated environment before any production connection.

Production data, trade secrets, and OT security

Manufacturing data can expose recipes, tolerances, yields, machine behavior, customer specifications, supplier performance, export-controlled material, and facility layout. Classify data and contracts before external processing. Minimize fields, encrypt transfer and storage, restrict service accounts, set retention, and prevent vendor training on confidential inputs. Keep product and customer isolation where required. Images and logs can reveal more than their file names suggest.

NIST SP 800-82 Rev. 3 addresses OT security while recognizing performance, reliability, and safety requirements. NIST's Manufacturing Profile identifies objectives including human safety, environmental safety, product quality, production goals, and trade secrets. Use those objectives to prioritize controls. Network segmentation, asset inventory, controlled remote access, backups, incident response, and coordination between IT and OT are prerequisites for connecting an AI service. A pilot should not create an unmanaged bridge into the plant.

Human-machine oversight

Assign a process owner and technical authority for each workflow. Operators approve operational action, quality owns inspection and disposition, maintenance owns work and restart, engineering owns process changes, EHS owns safety and environmental controls, and security owns OT connectivity. Reviewers need the relevant trend, image, work instruction, machine state, and uncertainty. They must be able to reject the recommendation without bypassing production discipline.

NIST research on human-machine teaming for manufacturing digital twins focuses on how people can work with AI for complex tasks. That teaming principle is essential: automation should improve situational awareness and consistency while preserving expertise. Apply management of change to model, threshold, camera, sensor, and data-pipeline updates. Revalidate after product, supplier, tooling, lighting, firmware, or process changes. Sample accepted results and investigate near misses, not just recorded defects.

KPIs for production and model performance

Measure the operational baseline by line and product. Relevant metrics include first-pass yield, scrap, rework, defect escape, unplanned downtime, mean time to repair, schedule adherence, changeover time, work-instruction search time, and supplier-document cycle time. For vision or anomaly systems, track precision, recall, false negatives, false positives, and detection latency by defect class. A high aggregate accuracy can hide a dangerous missed-defect category.

Add review acceptance, operator override, maintenance false alarms, time to safe fallback, stale-sensor events, model drift, unauthorized connection attempts, and recovery test results. Compare changes using controlled trials where feasible and document other process changes. Do not attribute yield or downtime improvements solely to AI when maintenance, materials, staffing, or demand changed. Expansion requires an operational benefit with no unacceptable deterioration in safety, quality, cybersecurity, or operator workload.

A 90-day manufacturing rollout

Days 1-30: choose one line and read-only workflow, involve operations, quality, maintenance, EHS, IT, and OT security, map assets and data, baseline performance, classify sensitive information, and build an evaluation set covering products, shifts, machines, defects, normal variation, sensor loss, bad images, and malicious files. Define physical-action exclusions. Prototype outside production control.

Days 31-60: run in shadow mode, compare recommendations with operator and quality decisions, test by operating condition, inspect network paths, and rehearse loss of connectivity and rollback. Days 61-90: deploy a limited advisory workflow with named review and no direct control write. Train each shift, monitor daily at first, and use management of change for updates. Expand only after stable performance across conditions, a passed OT security review, and signed safety, quality, and operations approval.

• Day 30: asset and data map, control boundary, owners, baseline, and condition-rich evaluation set complete.

• Day 60: shadow mode passes defect, drift, network, fallback, and operator-workload gates.

• Day 90: advisory production use has no direct control authority and has signed operational approval.

When not to automate

Do not automate when sensors are unreliable, labels are inconsistent, process conditions are changing, network segmentation is weak, or safe fallback is unavailable. Avoid cloud processing of trade secrets or controlled data when contractual and security requirements are unmet. Do not deploy a model trained on one machine or product as if it were validated for every line.

Keep safety interlocks, emergency response, equipment restart, process setpoints, final quality disposition, environmental controls, and changes to validated procedures under authorized human and engineered control. Do not let a prediction extend maintenance beyond safe limits or overrule an operator who sees conditions absent from the data. Deterministic control and statistical process control remain better for many stable, well-understood tasks. If failure can immediately harm people or equipment, advisory AI must remain outside the control loop.

Sources and methodology

This article synthesizes the primary sources below as of the publication date. Forecasts and recommendations are directional scenarios, not guarantees; they should be tested against your workflow, data, risk tolerance, and current vendor documentation.

National Institute of Standards and Technology: Guide to Operational Technology Security (accessed 2026-07-30)

National Institute of Standards and Technology: Cybersecurity Framework Manufacturing Profile (accessed 2026-07-30)

National Institute of Standards and Technology: Human/Machine Teaming for Manufacturing Digital Twins (accessed 2026-07-30)

Cybersecurity and Infrastructure Security Agency: Cross-Sector Cybersecurity Performance Goals (accessed 2026-07-30)

National Institute of Standards and Technology: AI Risk Management Framework (accessed 2026-07-30)