Executive Summary

Most discussions about Artificial Intelligence in manufacturing start with technology: large language models, computer vision, predictive analytics, or digital twins. The usual sequence is familiar — first choose a technology, then look for problems it might solve.

This analysis takes the opposite approach.

It starts with real production problems observed across Russian and international manufacturing enterprises. Only after identifying the most persistent and costly problems does it examine where digital tools and AI can actually help.

The central finding is straightforward:

The majority of recurring operational problems in manufacturing are not caused by the absence of modern AI models. They are caused by the absence of an effective system for capturing, storing, searching, and reusing production knowledge — what can be called the digital memory of the enterprise.

The working hypothesis is therefore:

The primary constraint on effective AI adoption in industry today is not the maturity of the models themselves, but the lack of high-quality, structured, and accumulated production knowledge. AI can significantly improve operational efficiency only when it has something real to work with.

Recommended sequence:

  • Build a unified digital knowledge base (the enterprise’s digital memory).
  • Establish reliable data accumulation and documentation practices.
  • Deploy intelligent services on top of that foundation (semantic search, recurring failure analysis, recommendations, automated documentation, prediction).

This post is a preliminary analysis based on open sources. The hypothesis requires field verification through interviews on actual production sites.

A whole document is available by link

Why a Roadmap for AI Cannot Start with Technology

A serious roadmap for AI in manufacturing should not begin with model selection, platforms, or infrastructure.

It should begin with a more fundamental question:

Which production problems actually need to be solved?

Skipping this step produces collections of fashionable digital initiatives that have little measurable impact on productivity.

The principle used here is simple:

First — real production problems. Then — analysis of their causes. Then — evaluation of existing digital tools. Only then — the role of Artificial Intelligence.

This is the approach used by leading industrial research centres and is the only reliable way to avoid technology for technology’s sake.

Methodology

Sources and priorities

The analysis prioritised materials closest to primary sources:

Priority Source
A Interviews with engineers, maintenance technicians, shift supervisors, process engineers, production managers
A Government productivity studies
B Academic papers based on shop-floor interviews
B Industry association materials
C Company annual reports and conference materials
D Consulting reports
E Press releases

The review covered Russian enterprises as well as material from Germany, the United States, Japan, China, Fraunhofer studies, ethnographic research of production processes, and maintenance organisation literature.

The goal was not statistical representativeness, but identification of patterns that appear independently of country, industry, and level of digitalisation.

Use of AI tools

Modern AI tools were used for search, aggregation, structuring, and preliminary analysis of open sources. This expanded the volume of material examined and accelerated pattern detection.

At the same time, the limitations are clear: possible incomplete sampling, interpretation errors, data quality issues, and publication bias. All conclusions remain preliminary working hypotheses that require verification through direct interviews on industrial sites.

Research limitations

A full AI roadmap cannot be built from internet sources alone. Field research is required.

A minimally representative sample would include at least ten enterprises from one industry, with interviews across the full range of roles: machine operators, maintenance engineers, repair teams, shift supervisors, process engineers, quality specialists, IT staff, production managers, and plant leadership.

Only such a sample can reliably identify systemic problems, quantify economic impact, and determine which digitalisation directions deliver the highest return.

Map of Recurring Production Problems

Despite large differences in countries, automation levels, and IT landscapes, the most frequently reported problems show a striking degree of consistency. Most of them have little to do with the absence of AI.

Workforce-related

  • Loss of production knowledge when experienced employees leave or retire
  • Shortage of qualified specialists
  • Long onboarding time for new engineers
  • Dependence on a small number of key experts

Information-related

  • Time spent searching for information before equipment repair
  • Missing or incomplete repair history
  • Absence of a unified knowledge base
  • Knowledge scattered across paper journals, Excel files, email, shared folders, and individual memory
  • Difficulty finding similar past cases
  • Outdated or incomplete technical documentation

Organisational

  • Information transfer between shifts
  • Heavy manual documentation burden
  • Loss of information in handovers
  • Duplication of records

Operational

  • Long fault diagnosis times
  • Repetition of previously known errors
  • Insufficient time for preventive maintenance
  • Extended equipment downtime

Causal model

Experienced specialists leave
            ↓
Loss of production knowledge
            ↓
Absence of a unified digital memory
            ↓
Long information search  ·  Repeated errors  ·  Slow onboarding
            ↓
Lost time and money

Virtually all of the most common problems are downstream effects of one root cause: the lack of an effective system for accumulating and using production knowledge.

Why This Matters Now

Knowledge management problems have existed for decades. Several converging factors make them critical today.

Aging engineering workforce

A significant share of the most experienced specialists is approaching retirement. Simultaneously, enterprises face a shortage of younger engineers. When 20–25 % of the workforce over 55 leaves, undocumented knowledge leaves with them.

Increasing equipment complexity

Modern machines require deeper and more specialised knowledge for operation and maintenance.

Accumulation of digital data

Enterprises already hold large volumes of electronic documentation, repair logs, photos, and other records. Most of this material remains fragmented and underused.

Maturity of language models

For the first time, modern language models can work effectively with large volumes of internal enterprise documentation in natural language — provided that knowledge has first been captured and structured.

Proposed Direction: Digital Memory of the Enterprise

The working hypothesis follows directly from the problem map:

The main constraint on AI adoption in industry is not the insufficient maturity of AI models, but the shortage of high-quality, structured, and accumulated production knowledge.

If confirmed by field research, the first stage of digital transformation should be the creation of a unified digital memory of the enterprise.

Such a system should bring together:

  • technical documentation
  • repair history
  • photos and video
  • diagnostic results
  • work instructions
  • engineer recommendations
  • shift logs
  • accumulated practical operating experience

Once this foundation exists, AI becomes a natural next step rather than an isolated project. Typical uses include:

  • intelligent (semantic) search
  • analysis of recurring failures
  • detection of operating patterns
  • recommendations for engineers
  • support for new employees (onboarding)
  • automated documentation
  • analysis of accumulated statistics
  • prediction of potential equipment failures

In short, AI starts working on the enterprise’s own experience instead of generic internet knowledge.

Recommended sequence

Stage 1 — Unified production knowledge base
Object-oriented digital card for every equipment unit. Integration with existing ERP, MES, CMMS, and file stores (a layer on top of current systems, not a replacement).

Stage 2 — Reliable data accumulation and documentation culture
Simplified input (mobile, voice, photo), incentives for staff, data quality control, versioning.

Stage 3 — Intelligent services
Semantic search with source citations, recurring failure analysis, recommendations, automated documentation, prediction.

Organisational conditions for success

Technology alone is insufficient. Critical conditions include:

  • dedicated time and incentives for documentation (otherwise data entry remains formal)
  • interfaces designed for shop-floor reality (mobile devices, gloves, noise)
  • data quality control
  • integration with the existing system landscape rather than creation of another silo
  • ability to operate in a closed contour (on-prem / private cloud), compliance with critical information infrastructure requirements, and use of domestic language models or the ability to fine-tune them

Practical Example: Manufacturing Knowledge Management System (MKMS)

One possible implementation of the approach described above is the Manufacturing Knowledge Management System (MKMS).

MKMS is positioned as the first infrastructural layer of digital transformation, not as a complete AI solution. Its primary purpose is to form the digital memory of the enterprise by giving every equipment unit its own digital card that consolidates documentation, repair history, media, and engineer comments.

This creates the foundation on which intelligent search, analysis, and decision-support services can later be built.

What digital memory does not solve

Creating a digital memory of the enterprise does not eliminate:

  • physical wear of equipment
  • lack of investment
  • design errors
  • capacity shortages
  • the need for genuine engineering expertise

It does, however, substantially reduce the time and knowledge losses that occur at almost every stage of the equipment life cycle, and it creates the necessary conditions for effective subsequent use of AI.

Next Step: Field Verification

The logical next stage is verification of the working hypothesis on a representative sample of industrial enterprises (for example, in Moscow) through structured interviews conducted directly on production sites.

Objectives:

  • confirm or adjust the problem map and priorities
  • assess current maturity of production knowledge management
  • collect quantitative data (search time, share of documented repairs, onboarding cost)
  • formulate a grounded AI roadmap based on real needs rather than technology expectations

A minimum sample of 8–12 enterprises from one or two priority industries, with interviews covering the full range of roles listed earlier, would provide a solid empirical base.

Conclusion

Discussion of Artificial Intelligence in manufacturing should not begin with models, platforms, or computing infrastructure.

The primary task is to create the conditions under which enterprise knowledge becomes accessible, accumulated, structured, and suitable for analysis.

Production knowledge should be treated as a strategic asset of the enterprise — alongside equipment, technology, people, and intellectual property.

Artificial Intelligence does not create this asset. It only makes it possible to use the knowledge the enterprise has already accumulated (or managed to preserve) far more effectively.

Building the digital memory of the enterprise is therefore a necessary preparatory stage for any serious application of AI in industry.

The next step is field verification of this hypothesis. Only after such verification will it be possible to construct a justified AI roadmap grounded in the real needs of modern manufacturing rather than in technological expectations.

Related Publications

Case study: manufactory knowledge management system (KMS)
AI applied to the Government: Transforming Public Administration into a Self-Learning System
Artificial Intelligence in Education and Management: Strategic Framework, Risks, and Practical Implications