Key takeaways
- Factory knowledge management is not a document-upload project. It is the controlled conversion of experience, records, decisions and operating context into source-linked knowledge objects that people can review and reuse.
- The retirement risk is measurable: McKinsey reports that the share of U.S. manufacturing employees over age 55 rose from about 10% in 1995 to about 25% in 2025, while North Carolina’s manufacturing share aged 55 and over reached roughly 30% in 2024. These figures are evidence of exposure, not a prediction for every plant.
- A useful industrial AI assistant must return the correct asset, revision, state, timestamp, source and uncertainty, then route consequential actions to an authorized human. Natural language fluency is not an acceptance criterion by itself.
- The best first use cases are bounded and evidence-rich: troubleshooting, shift handover, work-instruction retrieval, change-impact review, onboarding and maintenance-history search. Start read-only before adding workflow actions.
- ISA-95 boundaries, OPC UA information models, NIST AI RMF practices and OT security controls provide a stronger foundation than a generic vector database with broad access to plant systems.
The knowledge-loss problem is operational, not sentimental
A senior process engineer retires. A maintenance specialist moves to another site. A controls expert leaves after three decades. The organization still has the drawings, standard operating procedures, work orders and quality records, but the plant becomes slower at answering the questions that were never fully written down: Which symptom usually appears before this failure? Which parameter should be checked first? Which workaround is safe only on this asset revision? Which supplier document explains the exception in the field?
That is the factory knowledge problem. It is not solved by placing every file in a shared drive or by asking a general-purpose chatbot to summarize a document folder. Knowledge management becomes valuable when experience is captured with context, connected to authoritative records, kept current through change control and returned in a form that a qualified person can verify.
The retirement pressure is not hypothetical. McKinsey reports that the share of U.S. manufacturing employees over age 55 increased from about 10% in 1995 to about 25% in 2025, while the total U.S. manufacturing workforce declined from 20.5 million to 15.0 million over the same period. McKinsey describes the resulting “brain drain” as knowledge that is not passed to new joiners. [1] A separate North Carolina Department of Commerce analysis found that workers aged 55 and over represented roughly 30% of manufacturing employment in 2024, compared with 24% across all industries in the state. [2]
These figures are industry and regional indicators of exposure, not a universal forecast for every plant. They do, however, make the business question concrete: if a critical decision is known by one person, where is the evidence that the next person can find, understand and safely apply it?
A factory does not lose knowledge only when a person leaves. It loses knowledge whenever a decision cannot be reconstructed from the records, context and reasoning that produced it.
What counts as factory knowledge?
Factory knowledge is broader than controlled documentation. It includes formal artifacts, operational records and the reasoning that connects them. A useful knowledge program makes those categories explicit instead of treating every text fragment as equivalent.
| Knowledge class | Examples | What must travel with it | Typical AI-assisted use |
|---|---|---|---|
| Authoritative instruction | SOPs, work instructions, recipes, inspection plans, safety procedures | Owner, approval state, effective date, revision, applicability and superseded version | Find the correct instruction for an asset, product or operation |
| Engineering intent | Design rationale, calculations, constraints, deviation decisions, change reviews | Requirement, object identifier, revision, assumptions, approver and linked evidence | Explain why a design or parameter exists and assess change impact |
| Operational experience | Troubleshooting notes, shift handovers, operator observations, maintenance narratives | Asset, timestamp, symptom, conditions, action, outcome and confidence | Retrieve similar failures and prepare an evidence-backed investigation |
| Contextual data | Historian tags, alarms, events, quality results, CMMS history, MES records | Tag or object mapping, time window, units, state, data quality and source system | Put a document or recommendation in the right operating context |
| Learning material | Demonstration videos, mentoring notes, onboarding questions, lessons learned | Skill, role, prerequisites, capture date, reviewer and validity boundary | Shorten time to proficiency without pretending training replaces qualification |
The distinction matters because a retired engineer’s explanation is not automatically an approved work instruction. A shift note is not automatically a root cause. A historian trend is not automatically evidence of causation. An AI system should preserve those differences and show them to the user.
Why a document repository is not a knowledge system
A repository answers, “Where is the file?” A knowledge system must answer, “Which fact or decision applies to this situation, why should I trust it, and what should I do next?” The difference is the information model around the content.
A practical factory knowledge object should have at least six parts: a subject such as an asset, product, process step or failure mode; a source such as a document, work order, tag, event or expert recording; a state such as draft, released, obsolete or superseded; a time context such as effective date or observation window; an applicability boundary such as site, line, model, revision or operating mode; and an accountability trail showing who captured, reviewed, approved or corrected it.
Retrieval-augmented generation can help assemble an answer from those objects, but retrieval does not create authority. If the index contains two conflicting instructions and no revision status, the model can produce a fluent compromise that is operationally wrong. If the same pump has different identifiers in the CMMS, historian and engineering repository, the assistant may join the wrong evidence. If a workaround was valid only during a temporary production campaign, it must not appear as a timeless rule.
This is why the highest-return work is often knowledge preparation rather than model selection. Define identifiers. Resolve aliases. classify document states. Capture effective dates. Preserve units and time zones. Record source links. Mark uncertainty. Establish owners for correction. The model should be the last layer in a chain that makes evidence reviewable.
A four-layer architecture for industrial knowledge continuity
A production-grade design separates the knowledge lifecycle into four layers. The separation makes it easier to test failure modes and to assign responsibility.
1. Capture: turn experience into reusable evidence
Capture should happen while the expert is still available and while the work is visible. Interviews alone are not enough because people often omit what feels obvious to them. A better session starts with a real event, task or decision. Ask the expert to explain the symptom, the first discriminating check, the alternatives rejected, the safety boundary, the expected result and the evidence that would change the decision.
For a troubleshooting session, record the asset identity, operating mode, alarm or symptom, time window, observed conditions, diagnostic sequence, intervention, result and unresolved uncertainty. For an engineering lesson, record the design object, revision, requirement, constraint, tradeoff, approval and downstream effect. Audio or video can be useful, but the resulting transcript must be reviewed, linked to the source recording and broken into bounded, searchable knowledge objects.
McKinsey describes one industrial manufacturer that created a legacy-skills video library in which soon-to-retire employees demonstrated difficult skills step by step, including the reasons behind each action, common pitfalls and techniques learned through experience. The videos were segmented and accompanied by documentation in an LMS. [1] The lesson is not “record everything.” It is “capture the hard-to-learn decisions in the context in which a new worker will need them.”
2. Structure: create the factory’s vocabulary and lineage
Structure connects human language to industrial objects. “Line 4 filler,” “F-04,” “asset 001847” and a vendor serial number may all refer to one asset, or they may refer to different revisions over time. The knowledge layer needs a governed alias map and a clear rule for resolving conflicts.
ISA-95, also known as ANSI/ISA-95 or IEC 62264, provides a useful boundary model for enterprise-control integration. ISA’s public overview describes Level 3 as manufacturing operations management, including MES and SCADA, and Level 4 as business planning and logistics, including ERP. It also describes information exchange and transactions across those interfaces. [5] The model does not prescribe a single AI product, but it helps a team decide where a knowledge object originates, who owns it and what interface should expose it.
OPC UA offers a complementary technical foundation for industrial information. The OPC Foundation describes information models that represent structure, behavior and semantics, alongside communication, conformance and security capabilities. OPC UA can expose current and historical data, alarms and events, and supports information discovery through typed address spaces and relationships. [6] That is materially different from treating plant data as a bag of unlabelled text.
In April 2026, the OPC Foundation announced an initiative to make more than 430 OPC UA Companion Specifications easier to use for retrieval-augmented generation, semantic search and AI-assisted engineering workflows. The announcement describes a prototype with Markdown, image descriptions, token-optimized chunks, embeddings and MCP/REST interfaces. [4] This is an important direction for machine-readable industrial semantics, but it is an initiative and prototype announcement, not proof that every specification is already a production-ready knowledge base.
3. Retrieve: return evidence, not just an answer
A factory assistant should retrieve by industrial context before it retrieves by wording. A request such as “How do I clear the high-temperature trip?” should resolve the asset, line, product, operating state, alarm code and applicable revision before ranking passages. The answer should show the evidence used and identify what it does not know.
A useful response contract includes the following fields:
| Response field | Acceptance question |
|---|---|
| Object | Did the answer identify the correct asset, product, document or work order? |
| Applicability | Does the evidence apply to this site, line, model, revision and operating state? |
| Source | Can the user open the originating record, image, video, trend or event? |
| Freshness | Is the evidence current, and does the system show its effective or observation date? |
| Reasoning boundary | Does the answer distinguish reported fact, inference, recommendation and unknown? |
| Next step | Is the proposed action safe, authorized and routed through the existing workflow? |
This contract changes the evaluation target. The question is not whether the assistant sounds like a senior engineer. The question is whether a qualified engineer can inspect the result quickly enough to accept, correct or reject it.
4. Apply: connect knowledge to work without bypassing authority
The first applications should reduce search and preparation time while keeping decisions in existing operating processes. A knowledge assistant can prepare a troubleshooting brief, assemble the relevant work instructions, identify similar maintenance events, draft a shift handover or list the engineering records affected by a proposed change. It should not silently change a control parameter, release a procedure, close a quality deviation or create a work order without an explicit authorization and audit trail.
NIST’s AI Risk Management Framework is intended to help organizations incorporate trustworthiness into the design, development, use and evaluation of AI systems. Its core functions, Govern, Map, Measure and Manage, provide a practical structure for this layer. [7] For a factory knowledge assistant, that means defining ownership, mapping intended use and failure modes, measuring retrieval and answer quality, and managing incidents, drift and corrections over time.
The best first use cases are bounded by a decision
A knowledge project should start with a costly information delay, not with a promise to “chat with all plant data.” The use case must have a named user, a finite evidence set and a measurable current-state baseline.
| Use case | Why it is a strong first candidate | Safe first output | Measure before expanding |
|---|---|---|---|
| Troubleshooting assistant | Repeated faults generate searchable history and clear expert review points. | Ranked evidence, diagnostic checklist and unresolved questions. | Time to assemble an investigation, evidence coverage, correction rate. |
| Shift-handover copilot | The work is time-bound and already depends on structured context. | Draft handover with open issues, asset state, actions and owners. | Handover preparation time, omitted critical items, follow-up questions. |
| Work-instruction retrieval | Operators often need the right revision under pressure. | Source-linked instruction with applicability and revision status. | Search time, wrong-revision events, supervisor acceptance. |
| Maintenance-history search | CMMS narratives contain prior decisions that are difficult to reuse. | Similar events, actions, outcomes and evidence links. | Time to find precedent, duplicate investigations, false matches. |
| Engineering change impact review | Changes cross drawings, specifications, procedures and asset records. | Candidate impact set with source and confidence for engineer review. | Review cycle time, missed links found in review, rework. |
| Onboarding and skill transfer | Experts can demonstrate tacit techniques before they leave. | Reviewed learning path with video segments, prerequisites and checkpoints. | Time to proficiency, mentor load, qualification outcomes. |
The priority is not necessarily the use case with the most AI novelty. It is the one where the organization can assemble a trusted evaluation set. If the plant cannot agree on the correct answer or the authoritative source, adding a larger model will not fix the governance problem.
A 90-day pilot that earns trust
A practical pilot can be organized around four stages. The durations below are planning ranges, not a delivery promise. Actual effort depends on the number of systems, data quality, security review, subject-matter availability and the required validation evidence.
| Stage | Planning activity | Exit evidence |
|---|---|---|
| Define | Select one site, asset family, role and decision class. Baseline search time, escalation rate, rework and current answer quality. | Signed use-case boundary, risk assessment and representative test set. |
| Prepare | Inventory documents, work orders, tags, events, recordings and owners. Resolve identifiers, revisions, permissions and retention rules. | Data contract, source register, alias map and reviewed gold-standard questions. |
| Assist | Build read-only retrieval and answer generation with citations, filters, feedback and an abstention path. | Human evaluation showing source correctness, applicability, citation coverage and safe refusal. |
| Operate | Put the assistant into one existing workflow with logging, correction ownership and weekly review. | Measured time saved, correction patterns, incidents, drift indicators and scale decision. |
The evaluation set should contain normal questions, ambiguous questions, obsolete documents, conflicting instructions, missing evidence, wrong-asset traps and deliberately unsafe requests. A system that only succeeds on clean demonstrations is not ready for a factory.
The pilot should also test the human operating model. Who owns a correction? Who approves a new knowledge object? What happens when a work instruction changes? How does a supervisor see that an answer was generated from an older revision? What is the escalation path when the assistant abstains? The answers belong in the operating procedure, not only in the application backlog.
What to measure: time, trust and continuity
Factory knowledge management needs more than a chatbot usage count. Measure the delay that the program is intended to reduce and the quality risk it is intended to control.
| Measurement area | Example metric | Interpretation |
|---|---|---|
| Retrieval efficiency | Median time to locate the correct source and revision. | Shows whether the system reduces searching without hiding evidence. |
| Evidence quality | Percentage of responses with a correct, applicable source. | Separates grounded assistance from plausible language. |
| Human correction | Correction rate by use case, source type and failure mode. | Reveals where data or retrieval needs improvement. |
| Operational continuity | Time for a qualified replacement to complete a known task under supervision. | Tests whether captured knowledge helps transfer capability. |
| Workflow impact | Investigation preparation time, handover preparation time or change-review cycle time. | Connects knowledge work to a plant decision rather than vanity usage. |
| Safety and quality | Escaped wrong-revision instructions, unauthorized actions, quality deviations or near misses. | Defines non-negotiable guardrails and stop conditions. |
| Knowledge health | Stale-object rate, owner coverage, unresolved conflicts and review age. | Measures whether the knowledge base remains maintainable. |
Deloitte’s 2025 survey of 600 executives from large manufacturing companies found that 92% believed smart manufacturing would be the main competitiveness driver over the following three years. Respondents reported average net impacts of 10% to 20% in production output, 7% to 20% in employee productivity and 10% to 15% in unlocked capacity after smart-manufacturing implementation. [3] Those are survey-reported averages from a defined sample, not a promise for a knowledge-management pilot. They are useful context for why manufacturers are investing, but the pilot still needs its own baseline and measurement method.
Security and governance: keep the assistant outside the control path
Factory knowledge often contains sensitive process parameters, product designs, supplier information, quality evidence and operational details. The assistant therefore needs least-privilege access, source-level permissions, audit logging, retention controls and a defined boundary between retrieval and execution.
NIST SP 800-82 Rev. 3 emphasizes that OT systems have distinct performance, reliability and safety requirements and provides guidance for securing systems that monitor or directly affect the physical environment. [8] For a knowledge assistant, the practical implication is straightforward: start with read-only access through approved interfaces, keep model and retrieval services separated from direct control paths, and require an authorized workflow for every consequential action.
Deloitte’s survey shows why this cannot be treated as an afterthought. Sixty-five percent of respondents ranked operational risk among their top two smart-manufacturing concerns; 55% identified unauthorized access in OT as a high concern, 47% cited intellectual-property theft and 46% cited operational disruption. [3] The statistics describe survey sentiment, but they align with the architectural reality that more connected data and more automation create more paths to govern.
A sound control set distinguishes at least four identities: the reader who can search, the recommender who can prepare a result, the approver who can authorize a change and the executor who can apply it. An AI system may assist the first two roles. It should not inherit the authority of the latter two simply because it can call an API.
Common failure modes and how to avoid them
“We will upload everything”
A large corpus can increase uncertainty if it includes obsolete procedures, duplicate exports and informal notes with no ownership. Start with a source register and a document-state policy. Exclude or label material that cannot be tied to an owner, applicability boundary or review status.
“The model will know which revision is right”
Revision selection is a data and workflow responsibility. Retrieval can rank candidates, but the system should receive explicit metadata for released, draft, superseded and obsolete states. When the metadata is missing or contradictory, the correct behavior is to abstain and request review.
“We will capture tribal knowledge through interviews”
Interviews can produce polished summaries that omit the actual decision boundary. Capture real work, demonstrations, failure cases and rejected alternatives. Ask a second expert to review the result and test it against a representative scenario.
“A high answer score means production readiness”
A knowledge assistant can score well on a clean question set and still fail on wrong-asset, stale-revision or ambiguous-state cases. Evaluate negative cases, source applicability, safe refusal, latency, permissions and the human correction process.
“AI will replace the expert”
The defensible goal is continuity and leverage, not the removal of process authority. An expert should spend less time repeating the same evidence hunt and more time handling exceptions, validating changes and improving the knowledge system.
A decision checklist for manufacturing leaders
Before approving a wider rollout, ask five questions. First, can the team name the decision, user and site boundary that the assistant serves? Second, can every answer expose the source, revision, applicability and timestamp that support it? Third, can the system abstain when evidence is missing, conflicting or outside scope? Fourth, can a qualified owner correct a knowledge object and see where the correction propagates? Fifth, is every action that can affect production, quality, safety or compliance routed through an authorized and auditable workflow?
If the answer to any of these questions is no, the next investment should be in the data contract, operating process or evaluation set, not in a larger model.
Bottom line: preserve judgment, accelerate access
Manufacturers do not need another generic content repository. They need a way to preserve the decisions, context and practical judgment that make engineering and operations work repeatable. AI can help capture demonstrations, structure records, search across systems, compare prior events and prepare a clear evidence package. It cannot make an uncontrolled source authoritative, and it should not erase the boundaries between recommendation, approval and execution.
The strongest starting point is a bounded, read-only workflow where the plant already knows what good evidence looks like. Identify the experts and decisions most exposed to retirement. Capture the real work before it disappears. Map the assets, revisions and interfaces. Evaluate answers against difficult cases. Then connect the assistant to the existing operating process with explicit authority and feedback.
The goal is not to make the factory dependent on a chatbot. The goal is to make hard-won knowledge findable, reviewable and transferable, so that the next engineer can begin with the organization’s best available evidence instead of starting over.
References
[1] [McKinsey & Company: Investing in the manufacturing workforce to accelerate productivity](https://www.mckinsey.com/industries/aerospace-and-defense/our-insights/investing-in-the-manufacturing-workforce-to-accelerate-productivity)
[2] [North Carolina Department of Commerce: North Carolina’s Manufacturing Workforce Faces Growing Retirement Risk](https://www.commerce.nc.gov/news/the-lead-feed/north-carolinas-manufacturing-workforce-faces-growing-retirement-risk)
[3] [Deloitte Insights: 2025 Smart Manufacturing and Operations Survey](https://www.deloitte.com/us/en/insights/industry/manufacturing-industrial-products/2025-smart-manufacturing-survey.html)
[4] [OPC Foundation: OPC UA for the AI Era with Companion Specifications Optimized for Agentic AI](https://opcfoundation.org/news/press-releases/opc-foundation-advances-opc-ua-for-the-ai-era-with-companion-specifications-optimized-for-agentic-ai/)
[5] [International Society of Automation: ISA-95 Series of Standards](https://www.isa.org/standards-and-publications/isa-standards/isa-95-standard)
[6] [OPC Foundation: OPC Unified Architecture - Part 1: Overview and Concepts](https://reference.opcfoundation.org/specs/OPC-10000-1/4)
[7] [National Institute of Standards and Technology: AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework)
[8] [National Institute of Standards and Technology: Guide to Operational Technology Security, SP 800-82 Rev. 3](https://csrc.nist.gov/pubs/sp/800/82/r3/final)
