The Future of AI in Manufacturing: From Assistant to Autonomous Process Participant
An expert forecast of AI evolution in manufacturing across three stages — assistant, copilot, and autonomous agent — with a sober assessment of what is realistic within five years.
Serhiy Andrejev
Lead Machine Learning Developer
Manufacturing stands at an inflection point in its relationship with artificial intelligence. After years of pilot projects and incremental deployments, the industry is beginning to see the outlines of a more fundamental transformation — one in which AI evolves from a passive tool that answers questions into an active participant that monitors, recommends, and ultimately makes decisions within defined operational boundaries. This article offers a practitioner's forecast of that evolution, grounded in current data, real barriers, and a deliberate separation of what is probable from what is aspirational.
Introduction: Where We Are and Where We Are Heading
The manufacturing sector's engagement with AI has accelerated sharply. According to McKinsey's 2025 survey of over 100 manufacturing COOs, only 7% plan to keep digital and AI spending below 1% of cost of goods sold over the next five years, compared to one-third who spent that little in the preceding period. Nearly one-third of respondents intend to spend at least 5% of COGS on digital and AI initiatives [1]. Among the Global Lighthouse Network — the World Economic Forum's benchmark of advanced manufacturing facilities — 90% of technology use cases now incorporate AI [1].
Yet investment intensity has not translated into maturity. Approximately two-thirds of manufacturing COOs indicate their AI implementation is still at the exploration or targeted-implementation stage, with only 2% reporting AI fully embedded across all operations [1]. More broadly, McKinsey's State of AI report finds that nearly two-thirds of organizations have not yet begun scaling AI across the enterprise, and only about 5.5% see real financial returns from their AI investments [2].
This gap between investment and impact is not a failure of technology. It is a predictable consequence of deploying AI as a point solution — answering isolated questions, generating individual reports — rather than embedding it as a systemic participant in operational workflows. The trajectory of AI in manufacturing is best understood not as a linear improvement in model capability, but as a series of qualitative shifts in the role AI plays within the production environment.
In our practice of integrating AI into manufacturing and process industries across Ukraine and the EU, we observe three distinct stages in this evolution: assistant, copilot, and autonomous participant. Each stage demands not only different technology but different organizational structures, governance frameworks, and levels of trust. The remainder of this article examines each stage, the barriers between them, and what is realistically achievable within the next five years.
Stage 1: Assistant (2024–2025) — AI That Answers
The assistant stage represents the current state of AI deployment across the majority of manufacturing enterprises. At this level, AI functions as a responsive tool: it answers questions, retrieves information, summarizes documents, and generates reports on demand. The human operator initiates every interaction and retains complete decision authority.
Typical applications at this stage include:
- Documentation search. Natural language queries against technical manuals, standard operating procedures, and regulatory documentation. An engineer asks, "What is the torque specification for valve assembly V-4200?" and receives a direct answer sourced from the relevant document.
- Report generation. Automated drafting of shift summaries, production reports, and quality overviews from structured data sources. The system compiles; the human reviews and approves.
- Translation and communication. Multilingual document processing for enterprises operating across markets with different regulatory languages.
- Data exploration. Non-technical personnel querying ERP or MES data in natural language rather than through specialized interfaces.
The assistant stage delivers value primarily through time savings on information retrieval and document handling. In our observation, well-implemented assistant deployments can reduce the time engineers spend searching for information by 40–60%, freeing capacity for higher-judgment work.
However, the assistant stage has clear structural limitations. The AI has no persistent context between sessions. It does not monitor processes proactively. It cannot detect anomalies unless asked to look for them. It does not coordinate across systems or recommend actions based on multi-source analysis. Most critically, it takes no actions — it produces text and waits for the human to decide what to do with it.
The assistant stage is where the vast majority of manufacturing enterprises operate today. It is valuable, but it is not transformative. The transformation begins when AI stops waiting to be asked.
Stage 2: Copilot (2025–2027) — AI That Recommends and Monitors
The copilot stage represents a qualitative shift from reactive information retrieval to proactive operational partnership. At this level, AI continuously monitors data streams, identifies patterns and anomalies, generates recommendations, and automates routine decisions — while the human retains authority over consequential actions.
This stage is characterized by several key capabilities that the assistant lacks:
Continuous monitoring. Rather than responding to queries, the copilot continuously ingests data from SCADA, MES, ERP, and IoT systems. It detects deviations from specification before they become defects, identifies maintenance needs before they become failures, and flags supply chain disruptions before they halt production.
Contextual recommendations. When the copilot detects an anomaly, it does not merely alert. It correlates the deviation with equipment history, maintenance logs, environmental conditions, and similar prior incidents, then proposes specific corrective actions ranked by likelihood of success and operational impact. The operator receives not a data point but a decision package.
Routine automation. For decisions that fall within well-defined parameters — reorder triggers, shift scheduling adjustments, standard quality disposition — the copilot can execute autonomously, subject to predefined rules and post-hoc review. This is not artificial general intelligence; it is intelligent automation guided by organizational policy.
Cross-system coordination. The copilot integrates information across operational silos. A quality deviation on one production line triggers assessment of affected downstream processes, inventory implications, and potential customer impact — a correlation that typically requires multiple phone calls and spreadsheet analyses when performed manually.
The World Economic Forum's 2025 report on Physical AI emphasizes that this transition is accelerating: breakthroughs in intelligent systems that combine perception, reasoning, and action are enabling machines to perceive, learn, and respond to complex environments with far greater flexibility than traditional rule-based approaches [3]. Early industrial adopters are already reporting improved efficiency and faster response times.
Gartner predicts that by 2028, at least 15% of day-to-day work decisions will be made autonomously through agentic AI — up from effectively 0% in 2024 — and 33% of enterprise software applications will include agentic AI capabilities, up from less than 1% in 2024 [4]. This timeline aligns with our assessment: the copilot stage is not a distant vision but an imminent reality for enterprises with adequate data infrastructure.
Based on our implementation experience, we estimate that 20–30% of manufacturing enterprises with mature digital infrastructure will operate at the copilot level by the end of 2027. The limiting factor is not AI capability but organizational readiness — data integration, process standardization, and governance framework maturity.
Stage 3: Autonomous Participant (2027+) — AI That Decides and Acts
The autonomous participant stage represents the most ambitious vision for AI in manufacturing: a system that plans, decides, and executes within defined operational boundaries, with human oversight that is periodic rather than per-action.
At this level, AI agents orchestrate complex workflows across multiple production systems. They autonomously adjust production parameters in response to real-time quality data, reallocate resources based on demand signals and equipment availability, coordinate maintenance windows with production schedules, and manage inventory replenishment across supply chain networks — all within organizational policies and safety constraints.
Gartner's manufacturing predictions quantify this trajectory: by 2030, semi-autonomous AI agents will orchestrate 10% of key production operations, quality, and maintenance use cases, up from approximately 2% today, with humans retaining final approval authority [5]. Additionally, 15% of process manufacturing plants are expected to deploy closed-loop digital twin systems for real-time optimization, achieving projected 20% reductions in downtime and emissions [5].
However, the autonomous participant stage demands capabilities that significantly exceed the current state of the art in enterprise deployment:
Reliable long-horizon execution. Current AI agents reliably complete tasks equivalent to 60–90 minutes of human work, a task horizon that has approximately doubled every seven months since 2023 [6]. While impressive in software engineering contexts, industrial applications require sustained autonomous operation over shifts and days, not minutes.
Guaranteed safety boundaries. In safety-critical manufacturing environments — chemical processing, food production, pharmaceutical manufacturing — autonomous action requires not just high reliability but provable compliance with safety constraints. The margin for error in adjusting a reactor temperature is fundamentally different from the margin for error in editing a document.
Multi-system coordination under uncertainty. Autonomous operation across production, quality, maintenance, and supply chain systems requires handling cascading effects, conflicting objectives, and incomplete information — capabilities that remain at the research frontier rather than production readiness.
This stage is where the discourse most frequently outpaces reality. The technology trajectory is encouraging, but the gap between "technically possible in controlled conditions" and "reliably deployable in a 24/7 production environment" remains substantial. We expect meaningful autonomous operation to emerge first in non-safety-critical domains — logistics optimization, procurement automation, predictive maintenance scheduling — before extending to core production processes.
Technical Barriers to Evolution
The progression from assistant to autonomous participant is constrained by several technical challenges that merit explicit discussion.
Reliability at industrial scale. Large language models, the foundation of current AI agents, are probabilistic systems. They produce different outputs for the same input and can generate plausible but incorrect results — a phenomenon known as hallucination. In a manufacturing context, a 1% error rate on production decisions translates to thousands of incorrect actions per year. Industrial deployment demands deterministic reliability within defined operating parameters, which current probabilistic architectures cannot guarantee.
Real-time performance. Manufacturing processes operate on timescales from milliseconds (machine control) to minutes (process adjustments) to hours (production scheduling). Current AI agent architectures, which rely on large language model inference, introduce latency measured in seconds — acceptable for advisory functions but inadequate for real-time process control. The copilot and autonomous participant stages require hybrid architectures that combine ML-based reasoning with deterministic control systems.
Explainability. When an AI agent recommends adjusting a production parameter, the operator and the regulatory framework require understanding of why. Current large language models offer limited introspection into their reasoning processes. For safety-critical manufacturing decisions, "the model suggested it" is not an acceptable justification. Explainability is not merely a technical feature — it is a prerequisite for regulatory compliance and operational trust.
Data quality and integration. AI systems are only as reliable as the data they consume. Manufacturing environments are characterized by heterogeneous data sources — legacy SCADA systems, modern IoT sensors, manual quality records, ERP transactions — with varying levels of completeness, accuracy, and timeliness. The copilot and autonomous participant stages require data integration infrastructure that most manufacturing enterprises have not yet built.
Organizational Barriers to Evolution
Technical capability is necessary but insufficient. In our experience, organizational barriers are more frequently the binding constraint on AI evolution in manufacturing.
Trust. Anthropic's empirical research on agent autonomy, analyzing millions of human-AI interactions, identified a phenomenon they term the "deployment overhang": a significant gap between what AI can do and how much autonomy humans are willing to grant [7]. Their data shows that independent estimates suggest AI can handle tasks requiring approximately five hours of human effort, yet users grant an average of only 42 minutes of autonomous operation — a roughly 7x gap [7]. In manufacturing, where errors can affect product safety and human welfare, this trust gap is even wider.
Interestingly, experienced users exhibit a counterintuitive pattern: they auto-approve routine actions more frequently (over 40% of the time versus 20% for new users) but also interrupt more frequently on complex tasks (9% versus 5%) [7]. This reflects a maturation from micromanagement to strategic oversight — exactly the supervision model that industrial AI deployment requires.
Responsibility and liability. When an AI agent makes a production decision that results in a defective batch, who bears responsibility? The operator who approved the deployment? The engineering team that defined the authority boundaries? The vendor that supplied the model? Current regulatory frameworks do not provide clear answers, and this ambiguity inhibits autonomous deployment in regulated industries.
Regulation. Manufacturing operates under extensive regulatory oversight — HACCP in food production, GMP in pharmaceuticals, ISO standards across industries. These frameworks were designed around human decision-making and human accountability. Integrating autonomous AI agents requires not just compliance with existing regulations but evolution of the regulatory frameworks themselves — a process that inherently lags technological capability.
Workforce qualification. The World Economic Forum projects that by 2030, nearly 40% of core skills in advanced manufacturing will change [8]. Deploying AI copilots and autonomous agents requires personnel who can define operational boundaries, interpret AI recommendations, override autonomous decisions when appropriate, and audit agent behavior. This skill profile does not yet exist at scale in the manufacturing workforce.
Organizational structure. AI agents that operate across functional silos — coordinating production, quality, maintenance, and supply chain — challenge traditional organizational hierarchies where each domain has its own authority structure. The copilot and autonomous participant stages require new organizational models for cross-functional AI governance.
Evolution Roadmap: Three Stages of AI in Manufacturing
The following table summarizes the capabilities, barriers, and expected adoption levels at each stage of AI evolution in manufacturing.
| Dimension | Stage 1: Assistant (2024–2025) | Stage 2: Copilot (2025–2027) | Stage 3: Autonomous (2027+) |
|---|---|---|---|
| Core function | Answers questions, retrieves information | Monitors, recommends, automates routine | Decides and acts within boundaries |
| Human role | Initiates all interactions and decisions | Reviews recommendations, approves actions | Sets boundaries, provides periodic oversight |
| Data integration | Single-source queries | Multi-source correlation | Cross-system orchestration |
| Memory | Session-scoped | Persistent context | Organizational knowledge base |
| Key barrier | Data accessibility | System integration, trust | Reliability, regulation, accountability |
| Adoption estimate | 60–70% of large manufacturers | 20–30% by end of 2027 | <10% by 2030 (non-critical processes) |
Table 1. AI evolution stages in manufacturing: capabilities, barriers, and adoption trajectory. Sources: McKinsey [1][2], Gartner [4][5], author's assessment.
Figure 1. Estimated adoption of AI by evolution stage among large manufacturers. Midpoint values: 65% for Stage 1, 25% for Stage 2 (by end of 2027), 8% for Stage 3 (by 2030, non-critical processes only). Sources: McKinsey [1][2], Gartner [4][5], author's assessment.
Readiness Requirements by Stage
The transition between stages requires progress across multiple organizational dimensions simultaneously. The following assessment maps readiness requirements across five critical axes for each evolution stage.
| Axis | Assistant | Copilot | Autonomous |
|---|---|---|---|
| Technology | LLM access, basic API integration | Real-time data pipelines, tool integration, agent frameworks | Hybrid AI-control architectures, edge computing, digital twins |
| Data | Digitized documentation, structured databases | Integrated OT/IT data, real-time streams, quality baselines | Comprehensive data mesh, closed-loop feedback systems |
| Trust | Minimal — advisory only | Moderate — validated recommendations inform decisions | High — verified through track record, audit trails, and fail-safes |
| Regulation | Standard IT governance | AI-specific risk assessment, audit requirements | Regulatory frameworks for autonomous industrial AI |
| Organization | IT-led deployment | Cross-functional AI governance, updated SOPs | New roles (AI operations), redefined authority structures |
Table 2. Readiness requirements across five dimensions for each AI evolution stage. Source: author's assessment based on implementation experience and Gartner AI maturity framework [4][5].
Figure 2. Required readiness levels across five organizational dimensions for each AI evolution stage (1 = minimal, 5 = comprehensive). Source: author's assessment based on Table 2.
Critically, these dimensions are interdependent. An enterprise cannot leap to autonomous operation by advancing technology alone while trust, regulation, and organizational structure remain at the assistant level. The most common failure pattern we observe is precisely this: investing heavily in AI technology while neglecting the organizational and governance foundations it requires.
What Is Realistic: An Expert Assessment
The temptation in any forward-looking analysis is to project technological capability curves and assume organizational adoption will follow. Our experience suggests a more measured outlook.
What is realistic by 2028:
The copilot stage will become the standard expectation for large and mid-size manufacturers with mature digital infrastructure. AI will continuously monitor production parameters, generate contextual recommendations, and automate routine decisions in quality, maintenance scheduling, and inventory management. The majority of value will come not from spectacular autonomous feats but from systematic elimination of information latency — getting the right data to the right person with the right context, faster than any human-driven process can achieve.
Gartner's prediction that over 40% of agentic AI projects will be cancelled by end of 2027 due to escalating costs, unclear business value, or inadequate risk controls [4] is sobering but instructive. The projects most likely to survive are those with clearly bounded scope, measurable ROI, and organizational sponsorship beyond the IT department.
What is realistic by 2030:
Autonomous operation will be demonstrated in specific, well-bounded industrial use cases: predictive maintenance execution where the AI schedules and dispatches maintenance based on equipment condition data; logistics optimization where agents manage warehouse operations and delivery routing; quality disposition for standard non-conformance categories where the decision criteria are fully specified. Gartner's estimate that semi-autonomous agents will orchestrate 10% of key production, quality, and maintenance use cases by 2030 [5] appears reasonable — perhaps even optimistic given current organizational readiness levels.
What remains aspirational beyond 2030:
Fully autonomous production management — where AI agents coordinate across planning, execution, quality, and supply chain without per-decision human approval — remains aspirational for safety-critical manufacturing. This is not primarily a technology constraint. It is a trust, liability, and regulatory constraint. The technology to execute may arrive within five years. The organizational and regulatory frameworks to permit it will take longer.
In our assessment, the most productive investment for manufacturing enterprises today is not in pursuing autonomous AI but in building the foundation that autonomous AI will eventually require: integrated data infrastructure, standardized processes, robust governance frameworks, and workforce capabilities. The enterprises that build this foundation during the assistant-to-copilot transition will be the ones positioned to adopt autonomous operation when it becomes viable.
Author's Assessment and Conclusions
The evolution of AI in manufacturing from assistant to autonomous participant is not a matter of if but of when and how. Based on our experience integrating AI into industrial environments, we offer the following conclusions:
First, the three-stage framework — assistant, copilot, autonomous participant — provides a useful mental model, but the transitions are not clean boundaries. Enterprises will operate at different stages simultaneously across different functions. Quality monitoring may reach the copilot level while procurement remains at the assistant stage. The organizational challenge is managing this heterogeneity without creating governance gaps.
Second, the binding constraint is organizational, not technological. AI models will continue to improve. Data infrastructure will continue to mature. But trust, regulation, workforce capability, and accountability frameworks evolve on institutional timescales, not technology timescales. The enterprises that treat AI evolution as an organizational transformation — not merely a technology deployment — will capture disproportionate value.
Third, the "deployment overhang" identified by Anthropic's research [7] is perhaps the most important concept for manufacturing leaders to understand. The gap between what AI can do and what organizations permit it to do is currently approximately 7x — and in manufacturing, where the consequences of errors are more severe than in software engineering, this gap will persist longer. Closing it requires not just better AI but better governance, better audit mechanisms, and a structured approach to graduated autonomy.
Fourth, the immediate priority for most manufacturing enterprises is not to pursue autonomy but to master the copilot stage. The copilot delivers the majority of realizable value — reduced information latency, improved decision quality, automated routine work — while building the organizational muscle and trust that autonomous operation will eventually demand.
The future of AI in manufacturing is not a chatbot that becomes smarter. It is a gradual, governed expansion of the boundary between human decision-making and machine decision-making — an expansion that must be earned through demonstrated reliability, transparent reasoning, and organizational readiness, not assumed through technological optimism.
References
- McKinsey & Company. "From Pilots to Performance: How COOs Can Scale AI in Manufacturing." 2025. https://www.mckinsey.com/capabilities/operations/our-insights/from-pilots-to-performance-how-coos-can-scale-ai-in-manufacturing
- McKinsey & Company. "The State of AI: How Organizations Are Rewiring to Capture Value." 2025. https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai
- World Economic Forum & Boston Consulting Group. "Physical AI: Powering the New Age of Industrial Operations." September 2025. https://www.weforum.org/publications/physical-ai-powering-the-new-age-of-industrial-operations/
- Gartner. "Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027." June 2025. https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027
- Gartner. "Manufacturing Predicts 2026: AI and Digital Twins." 2026. https://www.gartner.com/en/industries/manufacturing-predicts
- IEEE. "Advancing Manufacturing Through Artificial Intelligence: Current Landscape, Perspectives, Best Practices, Challenges, and Future Direction." IEEE Access, 2024. https://ieeexplore.ieee.org/document/10677409/
- Anthropic. "Measuring AI Agent Autonomy in Practice." February 2026. https://www.anthropic.com/research/measuring-agent-autonomy/
- World Economic Forum & McKinsey. "Empowering Frontlines: Bridging the AI Skills Gap in Manufacturing." September 2025. https://www.weforum.org/publications/empowering-frontlines/