When the Asset Starts Taking Care of Itself: The Rise of Agentic AI in Maintenance

Predictive maintenance has taught us to detect what is going wrong and estimate when failure may occur. But what happens when an asset can also decide what should happen next? In his closing keynote at Asset Performance 2026, Professor Diego Galar will explore the move from predictive analytics to agentic AI, and introduce a concept that could fundamentally change maintenance and asset management: the self-preserving asset.

At first, self-preserving assets sound like something from science fiction. Machines that understand their own health, anticipate threats and take action to protect themselves seem a considerable step beyond the dashboards, condition monitoring systems and predictive models used in industry today. Yet Diego Galar, Professor at Luleå University of Technology and R&D VP at SISTEPLANT, argues that the building blocks are already emerging.

“A self-preserving asset can understand its own condition, anticipate threats, evaluate the consequences and initiate the most appropriate response within clearly defined boundaries,” he explains.

The important words are not only understand and anticipate. They are initiate a response. For Galar, that distinction marks an important shift in the evolution of maintenance technology. For years, industry has become increasingly capable of detecting anomalies, diagnosing faults and predicting failure. But knowing that something will happen is not the same as deciding what to do about it.

Prediction was never the final destination

Predictive maintenance has created enormous progress in the ability to understand asset condition. Sensors, condition monitoring, advanced analytics and machine-learning models can increasingly tell organisations what is happening inside their equipment and, in certain applications, how much useful operating time may remain.

But this still leaves a gap. “Predictive maintenance is focused on accurate diagnosis and prognosis,” says Galar. “It tells us what is going on and when a failure is expected to happen. But it is disconnected from the action.”

That final step is often still left to people. An engineer interprets the prediction, checks the operating context, considers production requirements and evaluates the consequences of different interventions. Should the machine be stopped immediately? Can it continue until the next planned shutdown? Should its operating conditions be changed? Could production be rescheduled? Is another maintenance action possible?

The prediction is therefore only one input into a much broader decision.

Prescriptive approaches started to address this by adding scenarios: What happens if we choose option A rather than option B? Self-preserving assets go another step further. According to Galar, the question becomes: what action should now be taken to preserve both the health of the asset and the mission it is expected to perform?

It is this connection between information, decisions and ultimately action that makes the concept fundamentally different.

Agentic AI closes the loop

The rapid development of AI is making that connection increasingly realistic.

The first wave of industrial AI was largely analytical. Machine-learning algorithms helped identify patterns in large quantities of operational data. More recently, large language models have created an intuitive interface between people and complex information. Now agentic AI adds another capability: AI systems that can work towards a defined objective and determine the steps needed to reach it.

“The clear difference with agents is that they connect the information to the action,” Galar says. Within a carefully defined framework, an AI agent can gather information from different sources, interpret the situation, evaluate alternatives and initiate an appropriate next step. Applied to maintenance, that creates a very different proposition from an algorithm that simply raises an alarm.

A future asset could combine knowledge about its current condition with its operating requirements, maintenance history, reliability models and information about the wider production context. Rather than merely reporting that a component is deteriorating, the system could help determine how the organisation should respond.

That does not necessarily mean immediately sending a robot to replace a bearing. The appropriate response might be much less dramatic: adjusting operating parameters, changing the production schedule, requesting an inspection or preparing a maintenance intervention.

The important shift is that the system begins reasoning towards an operational objective rather than stopping at the diagnosis.

And that is where Galar believes AI becomes particularly interesting for asset management.

Autonomy without governance is not autonomy

Allowing AI to participate in operational decisions immediately raises another question: how much freedom should it have?

Maintenance decisions have consequences far beyond equipment health. An intervention can influence safety, product quality, production output and cost. An AI system optimising one variable without understanding the wider constraints could therefore make a technically rational decision with unacceptable business consequences.

For Galar, this is precisely why the discussion about autonomous maintenance cannot begin with AI alone.

One of the misconceptions surrounding industrial AI, he argues, is the belief that a sufficiently advanced technology can compensate for weak maintenance fundamentals.

It cannot. “If you have a really bad maintenance policy, artificial intelligence will not suddenly make maintenance good,” he says.

The intelligence of an agent depends on the environment in which it is allowed to operate. Companies need solid maintenance policies, reliable sources of data and clear rules defining acceptable risk. Those principles become the boundaries within which an agent can act.

Risk management therefore moves from being a parallel management discipline to becoming part of the architecture of autonomous decision-making. The organisation must translate its policies into constraints: what can the system decide independently? When should it request approval? Which risks are acceptable? Which decisions must remain outside its authority?

Without that framework, more autonomy does not necessarily produce better decisions. It can simply produce faster uncontrolled ones.

For maintenance organisations attracted by the possibilities of agentic AI, that is an important reality check: the road towards autonomy starts with governance, not with an algorithm.

The human does not disappear — but the job changes

The obvious response is to assume that a person must therefore remain at the end of every decision loop. Galar does not necessarily agree.

Some decisions may continue to require human judgement because they involve uncertainty, broader consequences or contextual knowledge that cannot easily be encoded. Other decisions, however, could eventually be taken autonomously.

“The question is not whether we always put a human at the end of the loop,” he says. “It is where human knowledge, the human view of the context and understanding of the consequences are most valuable.”

That distinction could have major implications for maintenance organisations.

Highly qualified people still spend considerable time moving between fragmented information: checking work-order histories, comparing measurements, searching documentation, gathering data and reconstructing context before they can actually make a decision. AI agents could take over part of that preparatory and coordinating work.

Human expertise would then move towards the situations in which it adds the greatest value.

And as autonomous technologies develop further, some actions may no longer stop at a recommendation. Agents could potentially interact with automated systems — and eventually autonomous robots — allowing selected decisions to move from detection through reasoning to execution without requiring human intervention at every intermediate step.

Where exactly those boundaries should lie is not a purely technological question. It is one of the issues maintenance and asset-management organisations themselves will have to define.

The companies that start with the basics may move fastest

For organisations wondering how close all of this really is, Galar’s advice is surprisingly traditional.

Do not start by looking for a self-preserving asset.

Start by strengthening the foundations that such an asset would need.

Maintenance policies must be clear. Data about asset health must be available and trustworthy. The system also needs information about the asset’s mission and operational context. Structured and unstructured information must become accessible in a form that AI can use. Above all, the organisation needs a clear approach to risk.

These are hardly new maintenance principles. What changes is their importance when software is no longer merely presenting information but potentially acting upon it.

“The maintenance policy needs to be provided by humans,” Galar stresses. “AI is not going to replace that foundation.”

That may ultimately separate useful industrial AI from technology demonstrations. The organisations best positioned for agentic maintenance may not be those experimenting with the largest number of AI tools, but those capable of combining new technology with decades of reliability and maintenance knowledge.

A new generation will redefine maintenance

Looking ten years ahead is risky in a field moving as quickly as artificial intelligence. Yet Galar sees another development that may prove just as significant as the technology itself.

Within the next decade, an increasing proportion of maintenance professionals entering operational roles will have been born entirely in the digital era. Unlike previous generations who moved from analogue to digital technologies during their careers, these engineers and technicians will have grown up interacting naturally with digital systems and AI.

Galar describes himself as a “digital immigrant”. The coming generation will be different.

“They are not simply going to be replaced by AI or by robots,” he says. “They are going to have a relationship with AI that allows them to achieve their goals in a much more professional way.”

But there is an important warning attached to that prediction. Being fluent in AI does not remove the need to understand maintenance. Quite the opposite.

As AI becomes increasingly capable of transforming information into action, domain knowledge becomes crucial for determining whether that action makes sense. Reliability engineering, maintenance strategies, asset behaviour and risk management provide the knowledge and constraints within which intelligent agents can operate.

The maintenance professional of the future may therefore combine two worlds: a natural ability to work alongside intelligent systems and a deep understanding of the engineering principles built up over decades. That combination, rather than AI alone, may ultimately determine how far autonomous maintenance can go.

From intelligent predictions to intelligent assets

The concept of the self-preserving asset raises uncomfortable as well as exciting questions. How much autonomy should an asset receive? How can organisational risk policies be translated into boundaries an AI agent can understand? Which maintenance decisions still need human judgement and which ones are we keeping manual simply because they have always been manual?

There are no universal answers.

And Galar does not intend to provide all of them in advance.

During his closing keynote at Asset Performance 2026, he will take the concept beyond the theoretical framework and show a concrete example of an AI agent performing self-preserving actions. Exactly what that agent does — and how far the chain from information to decision and action can already be taken — is something he is saving for Antwerp.

What is already clear is that the discussion about AI in maintenance is entering a new phase.

The question is no longer only whether we can predict what an asset will do next.

It is becoming whether the asset itself can decide what should happen next.