Generative AI can answer questions, analyse information and recommend what to do next. Agentic AI goes a significant step further: carrying out actions across industrial systems and workflows, raising both the potential value and the stakes when AI gets things wrong.
For manufacturers still getting to grips with generative AI, another evolution is already gathering momentum. AI is shifting from a technology that primarily responds to human prompts towards one capable of pursuing goals, coordinating workflows and taking actions across connected systems.
Agentic AI describes systems that can interpret a goal, gather information, reason through a problem and use software tools to act. Instead of simply warning a maintenance engineer that a machine may fail, for example, an agent might investigate the likely cause, create a work order, schedule a repair around production priorities and notify the relevant personnel.
For David Villalon, co-founder and CEO of Maisa, a Spanish AI startup that provides “Digital Workers” designed to automate complex, multi-system processes for industry,
“The simplest distinction is that a copilot helps and suggests, but a person still has to execute the work. An agent, on the other hand, carries out the action itself and hands back a finished result.”
This changes the stakes. A misleading chatbot response can simply be rejected by its user. An incorrect action executed across a supply chain, maintenance system or production environment may have real operational consequences.
Vishwanathan Ramakrishnan, Global Vice President & Head, Digital & AI Software Engineering at ABB, describes agentic AI as
“the next step in moving from insight to action”.
The objective, he says, is not simply to give AI more autonomy, but to connect intelligence with action in a way that delivers measurable operational value.
Where Agents Can Add Value
The most convincing near-term opportunities for agentic AI are not within autonomous factories, but clearly defined processes where people currently spend significant time moving between systems, interpreting information and coordinating routine responses.
Maintenance is an obvious candidate. Predictive AI can already detect anomalies and warn that equipment may fail; an agent could connect that insight to the subsequent investigation, planning and maintenance workflow. Other promising areas include quality management, production scheduling, engineering, energy and resource management, and supply chains. Global consulting firm McKinsey similarly identifies applications across advanced industries ranging from engineering and procurement to manufacturing and maintenance.
One real-world example comes from a Maisa customer in Spain. A 20-person manufacturing team was struggling to handle daily alerts about delayed shipments and component shortages. One of Maisa’s AI Digital Workers, connected to the plant’s operational systems and trained on its processes, took over much of the review and handling of these alerts. The Spanish startup says resolution time fell 85% and errors by 95%, while daily throughput increased sixfold.
Jan Van Den Bossche, Regional Vice President, Software & Control EMEA at Rockwell Automation, expects increasing value to come from agents working across traditional organisational boundaries:
“Manufacturers are increasingly looking for ways to connect engineering, operations, maintenance and supply chain processes into a more continuous and responsive workflow.”
How Much Autonomy Is Too Much?
Allowing AI to recommend an action is one thing. Allowing it to execute that action is another. All three experts see industrial autonomy developing progressively, with the freedom given to an agent determined by risk and potential consequences. Generating reports, retrieving information, coordinating workflows or scheduling tasks within predefined limits may require little supervision. Safety-critical decisions, major process changes or actions with regulatory implications are different.
David Villalon proposes a useful test: consider how reversible a decision is and how far its consequences travel:
“If a wrong move can be caught quickly and undone without stopping a production line or breaching a contract, the agentic system can work autonomously. If it can’t, a human must stay in the loop.”
Jan Van Den Bossche calls the likely model “human-supervized autonomy“: agents handling increasing amounts of routine decision-making and orchestration, while people retain responsibility for governance, strategic decisions and exceptions.
When the Agent Gets It Wrong
That oversight matters because AI agents will make mistakes. For industry, the challenge is making those errors visible, contained and traceable, Jan Van Den Bossche says:
“Industrial AI cannot be evaluated using the same standards as consumer AI. Manufacturing requires reliability, traceability, and accountability.”
Agents therefore need access to trusted information and industrial context, clearly defined permissions and boundaries, audit trails and mechanisms for escalation. Digital twins and simulation could provide another safeguard, allowing manufacturers to test AI behavior virtually before changes reach physical assets.
David Villalon argues that reliability should rest on three guarantees: reproducibility, explainability and auditability. A convincing explanation generated after an event is not enough; manufacturers need to be able to establish what an agent actually did and why.
Crucially, agentic AI need not replace the deterministic control systems already responsible for running industrial processes. ABB envisages AI providing advanced analytics and decision support, and increasingly supporting defined actions, while real-time and safety-critical process control remains separate. For Vishwanathan Ramakrishnan:
“Ultimately, the objective should not be autonomy for its own sake. It should be about applying the right level of autonomy where it improves an operational outcome.”
The Integration Problem
Before agents can orchestrate an entire factory, they need to understand it. Industrial information is typically scattered across ERP, MES and SCADA platforms, maintenance systems, engineering tools, spreadsheets, sensors and equipment from multiple vendors. For Jan Van Den Bossche:
“Most manufacturers don’t suffer from a lack of data. They suffer from disconnected data.”
Vishwanathan Ramakrishnan argues that operational technology, information technology and engineering technology increasingly need to come together, giving AI the context behind individual data points. Yet this need not mean ripping out established automation. Industrial architectures can evolve progressively, connecting new AI capabilities while protecting existing assets and operational continuity.
Villalon offers another perspective: people already navigate fragmented systems every day. Agents could potentially do likewise (reading documents, calling APIs, accessing applications and cross-checking information). The bigger requirements, he argues, are clear access governance, process ownership and baseline measurement.
The Factory Five Years From Now
As agentic AI enables greater autonomy, none of the three experts expects people to disappear from factories and wider industrial settings. Instead, agents are likely to assume more of the information gathering, routine analysis, coordination and execution that currently consumes the time of engineers and operators. People can then concentrate more heavily on complex problem-solving, judgement, strategy and exceptions.
For Vishwanathan Ramakrishnan,
“As AI takes on more routine analysis and execution, operators and engineers can focus more of their time on what people do best: applying experience, exercising judgement, solving complex problems and making decisions where context matters.”
The result may be factories containing multiple specialized agents working across maintenance, production, quality, engineering and supply chains, sometimes acting independently, sometimes asking people for approval. This is also broadly consistent with the direction identified by global consulting firm Bain, which sees enterprise agentic AI progressing from individual task agents towards supervized cross-system workflows and, eventually, collaborative multi-agent systems.
But accountability provides an important limit. As Villalon points out, regulators, insurers and customers ultimately hold people and companies responsible, not software. So the industrial future may be neither human-run nor self-managing. Much as physical AI is giving machines greater ability to respond to the world around them, agentic AI will give industrial software greater ability to act on what it knows.
For manufacturers, the question is no longer simply what AI can tell them. Increasingly, it is what they are prepared to let AI do.
