As AI moves from analyzing industrial data to making decisions and potentially taking action, businesses face a critical question: how can they trust it? Digital twins are emerging as a safe virtual testbed, allowing AI outcomes to be validated before they affect machines, production lines and factories.
From Simulation to Validation
Digital twins are already well established across industry. By combining models of physical assets and processes with operational data, they allow manufacturers to simulate behaviour, explore changes and optimise performance without disrupting what happens in the real world.
But the rise of industrial AI — particularly AI agents capable of reasoning, making decisions and carrying out multi-step tasks — is now expanding their role. For Jan Van Den Bossche, Regional Vice President, Software & Control EMEA at Rockwell Automation,
“In many ways, the digital twin is becoming the proving ground for industrial AI. It gives manufacturers a practical pathway to move from experimentation to trusted deployment.”
Rather than allowing an AI system to make changes to a real production line, manufacturers can let it propose an action, simulate the consequences and assess the result first. Roy Clarke, Solutions Consultant at PTC, sees this less as a completely new application than an evolution of what twins already do:
“Digital twins provide an environment to test out improvements virtually. What is evolving is that we are now faced with testing out something that can reason and make its own decisions.”
Putting AI Through its Paces
Testing an AI agent may involve placing it within a virtual representation of a machine, process or factory — and then exposing it to different operating conditions. Manufacturers can then see how it reacts to bottlenecks, equipment failures, resource constraints or abnormal process conditions. They can also deliberately introduce difficult scenarios that would be expensive, disruptive or unsafe to reproduce physically.
Jesus Hernandez, Senior Vice President EMEA at AVEVA, says this allows organizations to examine whether an AI system
“stays within operating limits, responds appropriately to abnormal conditions and balances competing objectives”.
At AVEVA customer ISU Chemical, for example, process simulation combined with hybrid AI modelling achieved 99.7% reactor-yield prediction accuracy, demonstrating how AI and process models can work together to generate reliable industrial insights.
The attraction here is the ability to explore far more possibilities without putting production at risk. Siemens, for example, is using its digital twin technology to create environments in which manufacturers can simulate factories, virtually commission systems, train autonomous robots and evaluate AI-driven scenarios, before applying changes in the physical world.
But no twin can simulate reality in its entirety. Roy Clarke is particularly cautious about assuming that a virtual factory can reproduce all the “noise” of a physical one.
“We can get a twin to throw in deliberately messy data to test out specific scenarios, but can we second-guess and test for everything that might happen in reality? Probably not.”
Jesus Hernandez makes a similar point. No virtual environment can reproduce every unexpected event, he argues, meaning simulation still needs to be combined with real-world monitoring, controlled deployment and human review. Its purpose is
“to reduce uncertainty, rather than eliminate it”.
Start Where the Consequences are Manageable
For manufacturers considering the technology, one useful principle emerges: begin with decisions where mistakes can be detected and corrected without creating unacceptable risk.
Production scheduling, maintenance planning, energy optimisation and resource allocation are obvious candidates. Their consequences can be modelled and — unlike split-second safety decisions — an imperfect result can often be adjusted. For Roy Clarke,
“The question to ask here is what happens if an AI-generated decision turns out to be wrong? Scheduling and energy optimization seem like good places to start.”
Robot optimization and automation logic move further up the risk scale. Digital twins can already validate robot paths, machine interactions and control logic before deployment. Rockwell Automation, for example, says manufacturers can test AI-generated control logic, machine configurations and automation workflows virtually. But allowing AI to rewrite and deploy PLC logic autonomously is another matter says Clarke:
“The sensible approach is for the AI to propose the change, the twin to test it, and a qualified engineer or operator to approve it.”
Siemens draws a similar distinction between using AI to support decisions and allowing it to execute them. At the planning level, AI can analyze production scheduling, resource allocation and maintenance strategies. Digital twins can also test robot behavior and automation logic before physical deployment. But as decisions move closer to physical execution, validation and human judgement become increasingly important. For manufacturers, then, autonomy should not be treated as a single switch. Different decisions justify different levels of AI authority.
Can You Trust the Twin?
Simulation may help solve one trust problem, but it creates another: if a digital twin is being used to validate an AI, how do you validate the twin? For Jan Van Den Bossche,
“A model is only as useful as the quality of the underlying engineering data, process understanding and operational assumptions used to build it.”
This puts data foundations at the center of the equation. Twins must remain aligned with factories as equipment, processes and product mixes change. AI systems also need reliable contextual data if their recommendations are to be meaningful.
Jesus Hernandez therefore sees simulation as only one element of industrial AI assurance. For him,
“Industrial AI still requires validation, traceability, appropriate guardrails and human oversight.”
This means “tested enough” will mean different things for different applications.
Siemens is developing metrics and measurement systems to address precisely this issue. Ronny Hendrych, Head of Agentic AI for Operations Software at Siemens Digital Industries Software, says the company is aiming for
“a quality level between four and five sigma before agents are allowed to act autonomously in operations, without a human in the loop”.
In other words, a very high level of demonstrated reliability. But the acceptable threshold depends on the consequences of failure, says Hendrych:
“Lower-criticality actions can tolerate a higher number of mistakes and still be acceptable. For high-criticality actions, where something going wrong could cost 100,000 euros or put human life in danger, we may never let an agent run autonomously; there would always be a human in the loop making the final decision.”
Towards the Self-optimizing Factory
The longer-term vision is a continuous loop: operational data updates the digital twin; AI identifies an opportunity or problem; an agent proposes a response; simulation tests it; and an approved action is sent back to the physical factory. The results then become new data, allowing both twin and AI to improve.
Elements of this model already exist. AVEVA points to semi-autonomous optimization in operations with mature data and clearly defined control boundaries, particularly around anomaly detection, maintenance, energy and production decisions.
Siemens also sees the building blocks coming together. Its latest Digital Twin Composer combines engineering and simulation information with operational data and AI, while work at the Siemens Electronics Factory in Erlangen is exploring more adaptive and AI-powered production. Yet none of the four interviewees expects humans to simply disappear from industrial operations. According to Jan Van Den Bossche,
“The future is less about fully autonomous factories and more about trusted autonomy where AI can increasingly propose, test and optimize actions while people retain oversight of decisions that have significant operational consequences.”
Roy Clarke expects the propose-test-validate-deploy cycle to spread steadily, beginning with planning and moving closer to the shop floor as confidence grows:
“A fully ‘lights out’ factory with no human intervention? I’m not sure it’s the right destination. But I think we are moving rapidly towards AI doing the heavy lifting, with humans making the final call.”
A Practical Route to Greater Autonomy
For manufacturers across industrial sectors, the message is therefore not to wait for the arrival of some fully autonomous factory. The more useful approach is incremental: strengthen the industrial data foundation, build and maintain digital twins that accurately represent important processes, and start applying AI where the consequences of failure are manageable. From there, autonomy can expand as evidence and confidence accumulate.
Digital twins cannot prove that an AI system will never make a mistake, any more than they can reproduce every event a real factory might encounter. What they can provide is something increasingly valuable as AI moves closer to physical operations: a controlled space in which machines can make mistakes without those mistakes immediately becoming real.
The factory of the future may therefore be less about removing people than changing their role — with AI increasingly doing the proposing and testing, digital twins providing the proving ground, and humans deciding when the technology has earned the right to add value through its actions.
