Agentic AI in Manufacturing: What Must Be Ready First

Direct answer: agentic AI in manufacturing is ready for production only when a bounded workflow has a named business owner, trusted industrial context, controlled tool permissions, tested human and technical safeguards, and a baseline that shows whether the system improves the operation. If any of those conditions is missing, the next investment should repair that condition before it increases autonomy.

This is the practical difference between an impressive agent demo and an operating capability. A model may plan a maintenance response in seconds. The company still has to prove which asset it refers to, whether the sensor history is trustworthy, which work-order fields it may change, who reviews an exception, how the action can be reversed, and whether the result reduces downtime or only moves effort into supervision.

The evidence available through August 2026 points in one direction. AI use is broad, while production-scale agentic work remains narrow. The constraint is rarely access to a better model. It is the condition of the process around the model.

What counts as agentic AI in an industrial company?

An industrial AI system becomes agentic when it can pursue a defined goal across several steps, choose from approved tools, retain relevant task state, and take or prepare an action inside a workflow. That workflow might cover failure-to-restore, production exception handling, quote-to-order, schedule recovery, quality investigation, or field-service preparation.

The term should describe operating behavior, not a product badge. A chat interface that summarizes manuals is useful retrieval. A copilot that drafts a work order is useful assistance. An agent that checks the asset hierarchy, reviews condition data, prepares the work order, verifies parts availability, proposes a maintenance window, and routes an exception is performing a multi-step job. If it can write to a system of record, its identity, permission, evidence, and stop conditions become part of the product.

Many valuable industrial systems should remain at levels 1 or 2. More autonomy is worthwhile only when the next level improves the operating outcome enough to justify its added failure modes and control burden.

What the adoption data actually says

McKinsey's November 2025 survey found that 88 percent of respondents reported regular AI use in at least one function. Twenty-three percent said their organizations were scaling an agentic system somewhere, and another 39 percent were experimenting. Yet no individual business function had more than 10 percent of respondents reporting scaled agent use. Only 39 percent attributed any enterprise-level EBIT impact to AI, and most of that group attributed less than 5 percent of EBIT.

The manufacturing picture is earlier. In the Manufacturing Leadership Council's 2025 survey, 6 percent of respondents reported current use of agentic AI and 24 percent expected use within two years. The same report found that 65 percent lacked the right or useful data for AI applications, 62 percent had data that was not formatted or structured for AI use, and one in four did not measure AI value.

Official European statistics offer a broader baseline, although they measure any AI rather than agents. Eurostat reported that 17.3 percent of EU manufacturing enterprises with at least 10 employees used one or more AI technologies in 2025, compared with 20 percent across the covered business economy. Among manufacturers that had considered AI without adopting it, the most common reported obstacle was lack of expertise. Data availability or quality and incompatibility with existing equipment, software, or systems were also substantial barriers.

Those surveys use different populations and questions, so they should not be combined into a league table. They do support a careful conclusion: interest and experimentation have moved faster than process readiness and economic measurement.

An August 2026 Deloitte survey of 501 US leaders, all directly involved in organizations already piloting agentic AI, makes that gap more visible. Sixteen percent said their business processes were prepared for agentic adoption, and 5 percent said highly prepared. Even among scaled adopters, 46 percent said their processes were prepared. Deloitte also reported that 72 percent cited the lack of a unified and accessible data foundation, 70 percent the inability to trust and govern agents, and 67 percent the cost and complexity of integration.

Industrial leaders are building real capability

The evidence does not support a simple story in which industrial incumbents ignored AI. Siemens has announced a three-year EUR 1 billion investment to scale its AI offerings and is building an industrial AI operating system with partners. Holcim reports 38 large-scale AI initiatives, AI deployment in more than 100 plants, and a target of CHF 200 million in recurring EBIT benefits from AI by 2028. ABB says roughly half of its R&D employees work on digital and software development, while its Genix platform integrates operational, engineering, and IT data. Honeywell describes live, domain-trained agentic workflows inside its Forge platform.

These are company disclosures, not independent proof of enterprise-wide transformation. They show meaningful investment, operating data, and deployed products. Public material is much less able to show whether the same workflow can be reproduced across business units without a new integration program, whether agent actions improve audited P&L, or how exception handling works in daily operations. That distinction matters more than deciding which company is ahead.

Four shared layers make local results transferable

A strong use case can work without becoming infrastructure. The operating advantage begins when the company can reuse the context, permissions, process design, and learning method behind that use case. The two August 2026 research syntheses behind this guide point to four shared layers.

These layers explain why the better leadership question is not how many agents the company has. It is whether the next valid workflow becomes cheaper and safer because the previous one existed.

The seven gates of industrial agentic readiness

A workflow that fails gate 1 is an idea without an operating case. A workflow that fails gate 2 needs a data product. A workflow that fails gates 4 or 5 is not ready to act. A workflow that fails gate 6 cannot support an investment decision. A workflow that fails gate 7 may still be a good local solution, but it should not be presented as enterprise infrastructure.

Teams that have not chosen the workflow can use the industrial AI use-case decision framework first. Once the case is selected, the bounded AI workflow guide shows how to define source context, human review, and a controlled first test.

Why industrial context is harder than enterprise search

Industrial context is more than a document index. An agent may need to know that two tag names refer to the same pump after a retrofit, that the historian records in seconds while the maintenance system records by shift, that a pressure value changed unit after a sensor replacement, or that the valid procedure depends on product grade and site jurisdiction.

The hard work is often identity and time alignment: asset hierarchy, bill of materials, material and product master data, customer and order identity, event chronology, and the authority of each source. A fluent answer built on the wrong asset, revision, or unit is still wrong. This is why data lakes alone do not create agent readiness. The workflow needs a governed context layer that preserves source, meaning, ownership, and freshness.

Who should own agentic transformation?

A Chief AI Officer, head of data, or enterprise platform team can be useful when standards and investment are fragmented. That team should own shared architecture, model and vendor policy, identity patterns, evaluation methods, security requirements, and reusable tooling.

Operational accountability should stay with the line. The COO and the relevant process owner remain responsible for whether an agent changes safety, quality, uptime, service, cost, or margin. Plant engineering, OT security, IT, legal, workforce representatives, and domain experts join according to the consequence of the use case.

A title cannot substitute for those decision rights. The useful test is whether every agent-enabled workflow has one accountable owner who can explain its boundary, evidence, current performance, and stop conditions.

Use a 90-day path to a bounded decision

Days 1 to 15: establish the operating case. Select one repeated, consequential workflow. Review a bounded sample of real cases. Record cycle time, delay, rework, exception types, error cost, current owners, systems touched, and the outcome leadership wants to improve.

Days 16 to 30: build the workflow contract. Define the trigger, allowed sources, context requirements, task steps, approved tools, prohibited actions, approval points, exception route, retention, audit record, rollback, and stop condition. Separate exact rules from model interpretation.

Days 31 to 60: run in shadow mode. Let the system retrieve, classify, recommend, or prepare while people continue to make every consequential decision. Compare the agent with experienced operators on representative cases, rare events, missing data, conflicting evidence, and deliberately difficult inputs.

Days 61 to 90: allow one reversible action. If the evidence supports it, grant the smallest useful write permission. Monitor outcome, overrides, false positives, false negatives, supervision time, incidents, and user adoption. End the period with a go, revise, hold, or stop decision based on the pre-agreed thresholds.

Ninety days is enough to make a disciplined next decision. It is not a promise that an industrial workflow can be transformed in one quarter.

For the transition from evidence to delivery, use the audit-to-implementation decision path. It keeps the first sprint tied to an owner, an acceptance test, and a go or stop decision.

Measure value, control, and replication together

Model accuracy is one input. An industrial business case also needs three connected scorecards.

• Operating value: downtime avoided, yield, scrap, energy per unit, schedule adherence, quote cycle time, first-time fix, working capital, service response, or another process-specific measure.
• Control quality: error severity, overrides, exceptions, missed escalations, unauthorized attempts, recovery time, supervision effort, and cases where the evidence was incomplete.
• Replication economics: days to the next comparable deployment, percentage of reusable components, local integration hours, validation effort, retraining, and performance after transfer.

The replication scorecard is where infrastructure starts to show. If every deployment requires a new ontology, connector, permission model, evaluation set, and consulting team, the company is funding a series of projects. Shared infrastructure should make the next valid deployment faster, safer, and easier to evaluate.

Build governance into the workflow

NIST's AI Risk Management Framework provides a useful operating backbone: govern, map, measure, and manage. For an industrial agent, that becomes clear accountability, a documented task and context, pre-deployment testing, live monitoring, appeal and override, incident recovery, change control, and safe decommissioning.

European deployments also need a current legal classification. The European Commission's AI Act timeline states that the Act became broadly applicable on 2 August 2026, with exceptions. Following the July 2026 AI Omnibus, rules for Annex III high-risk uses apply from 2 December 2027, while high-risk AI embedded in regulated products under Annex I applies from 2 August 2028. An industrial use is not automatically high-risk because it runs in a factory, and classification depends on the system's intended purpose and legal context. Teams should verify the applicable duties with qualified legal and product-safety counsel.

Compliance dates should not determine the full control standard. If an agent can affect people, equipment, product conformity, the environment, customer commitments, or material financial records, its evidence and recovery mechanisms belong in the design now.

When an agent is the wrong answer

• The process has no agreed owner or stable purpose.
• A fixed rule, form, alarm rationalization, integration, or standard operating procedure solves the problem more reliably.
• The source data cannot identify the relevant asset, order, version, unit, or time period.
• Errors are difficult to detect before an irreversible consequence.
• The exception volume would turn expert staff into full-time supervisors.
• The team cannot establish a baseline or name the measure that should improve.
• The local result cannot justify the integration and control cost.

Saying no to an agent can be a sign of operational maturity. The right answer may be better master data, a clearer procedure, a deterministic control, or a narrower assistive tool.

The decision industrial leaders should make now

Do not ask whether the company should adopt agentic AI in the abstract. Ask which workflow is ready to cross one autonomy level, what evidence supports that step, who remains accountable, and what must be true before the same pattern reaches another site.

The wider pattern also appears in ShiftNode's research on where AI creates commercial value in B2B: tools become commercially useful when they improve a named decision or handoff and preserve the evidence around it.

Industrial companies already own the raw material for an advantage: operating history, engineering knowledge, installed assets, customer context, and experienced people. That material becomes valuable when it is organized into a controlled learning system. The work is specific and sometimes unglamorous. It is also the path from an AI project portfolio to an operating capability that improves with use.