Agentic AI for predictive maintenance

---

Unplanned downtime now costs the world’s 500 largest companies roughly $1.4 trillion a year (Siemens, 2024). Predictive maintenance can cut machine downtime by 30 to 50% (McKinsey), yet most efforts stall in pilot because a prediction is only an alert. Someone still has to act on it. Agentic AI closes that loop: it contextualises the anomaly, checks parts, schedules a technician, and raises the work order, with a person supervising. The hard part is integration with the CMMS, ERP, and MES. We at Vstorm build that integration as production-grade, observable systems through our TriStorm methodology.

---

Moving from scheduled downtime to condition-based intervention

Unplanned downtime is still a board-level cost. At the start of 2026, stopping a production line costs more than it ever has. Siemens’ The True Cost of Downtime 2024 report puts the annual loss for the world’s 500 largest companies at roughly $1.4 trillion, equal to 11% of total revenues and up from 8% in 2019 and 2020, a 62% increase in five years (Siemens, The True Cost of Downtime 2024). For an automotive plant, an idle line can cost up to $2.3 million per hour (Siemens, 2024).

For a mid-market manufacturer the per-hour figure is smaller, but the exposure is not. Margins are thinner, schedules carry less slack, and a handful of unplanned hours can wipe out a week of production targets. Downtime is no longer a maintenance line item. It is a board-level number that shapes capital decisions and customer commitments.

How maintenance is handled today

Most plants run on a mix of two strategies. Reactive maintenance fixes a machine after it breaks. Preventive, or scheduled, maintenance replaces parts at fixed intervals whether or not they are worn. Both work, up to a point, and both waste money. Reactive maintenance trades a low upfront cost for unplanned stoppages. Scheduled maintenance trades predictability for parts replaced too early and lines shut down that did not need to stop.

Condition-based maintenance (CBM) changes the trigger. Instead of a calendar, the trigger is the equipment’s actual state, so a machine comes offline only when its readings warrant it. This depends on continuously monitoring equipment performance: IoT sensors, together with techniques such as vibration and oil analysis, feed performance data that shows how each asset is behaving. McKinsey notes this increases the time between repairs compared with fixed-interval preventive maintenance (McKinsey, Establishing the right analytics-based maintenance strategy). Done well, predictive maintenance cuts machine downtime by 30 to 50% and extends machine life by 20 to 40% (McKinsey, Manufacturing: Analytics unleashes productivity and profitability), and over the long term it reduces maintenance costs. The prize is real. The difficulty is in reaching it.

Why predictive maintenance stalls before it scales

The uncomfortable truth is that most predictive-maintenance efforts never leave the pilot stage. McKinsey identifies three recurring barriers: data that is insufficient, inaccessible, or of low quality; technology that is inadequate, with too few sensors or weak IT infrastructure; and the difficulty of prioritising which assets to cover (McKinsey, Prediction at scale).

There is a fourth barrier that receives less attention. A prediction, on its own, is only an alert. Someone still has to read it, check the maintenance history, confirm the parts are in stock, find a qualified technician, and raise the work order. The model does the easy part. The coordination does the rest. There is also a quieter risk: a model that over-predicts can erase its own savings, as McKinsey found in one case where a 10% false-positive rate cancelled out the gains (McKinsey, analytics-based maintenance strategy).

This pattern is not unique to maintenance. MIT’s NANDA research found that despite $30 to $40 billion invested, 95% of organisations see no measurable return from generative AI, with poor integration into existing workflows named as a central cause (MIT NANDA, The GenAI Divide: State of AI in Business 2025). The failure is rarely the model. It is everything around it.

From prediction to intervention: what the agentic layer adds

The difference between predictive analytics and an agentic system is the difference between insight and execution. A predictive model forecasts equipment failures: it tells you a bearing is likely to fail. An agent does something about it. This is the core of agentic AI for predictive maintenance.

In practice, an AI-powered agentic layer takes the anomaly and closes the loop. It contextualises the signal against the asset’s maintenance history, checks spare part availability and cost, identifies a technician with the right skills and availability, schedules the maintenance activities into the next planned maintenance window, and generates a fully populated work order for the maintenance tasks that follow. A human stays in the loop where judgment matters, approving the action or handling the exception. This is supervised autonomy, not unsupervised automation. The agent runs the routine coordination; the team keeps control of the decisions that carry risk.

“The prediction model is the easy 10% of the problem. The value lives in the 90% that follows: reading the plant’s own systems, acting inside them, and learning from what happened. That is an engineering problem, not a data science one.”

— Antoni Kozelski, CEO and Founder, Vstorm

The integration problem is the real problem

The hard part of condition-based intervention is not the prediction. It is AI agent integration with existing systems: the platforms that already run the plant. A monitoring system can monitor equipment and flag a fault, but on its own it stops at the alert. Maintenance history sits in the CMMS, parts and cost in the ERP, and the production schedule in the MES. An agent that cannot read and write across all three can detect a fault but cannot act on it.

Off-the-shelf platforms assume the data is clean and consistent: that a machine in the MES maps to the same asset in the CMMS and the same cost centre in the ERP. In most plants it does not. Bridging that gap is where projects either deliver or stall, and it is the work that templates do not do. The table below sets out what changes when the loop is closed.

How we build condition-based intervention at Vstorm

We build agents that act inside existing workflows, not dashboards that wait for someone to act. We saw this directly in our text-to-workflow platform for an engineering client, where agents convert plain instructions into validated, executed workflows rather than recommendations a person has to retype (Vstorm case study: text-to-workflow agentic AI platform). The same principle applies to maintenance: the value is in the agent completing the task, not flagging it.

Production-grade systems also need to be observable. Every agent decision should be traceable and auditable, which is why we instrument our systems from the first build rather than adding monitoring later. Our journey from a single agent to a hybrid agent-graph architecture with Pydantic AI shows how we keep complex agent behaviour transparent in production (Vstorm case study: hybrid agent-graph architecture).

This sits inside our TriStorm methodology. Transformation Consulting identifies which assets justify intervention and builds the ROI case; Agentic AI Engineering builds the closed loop and integrates it with the plant’s systems. One team carries the work from roadmap to deployed system, so the strategy and the build do not drift apart.

Where to start: sequence by asset criticality

The mistake is trying to instrument everything at once. The better path is to start where downtime costs the most, prove the closed loop on a bounded set of critical assets, then scale once the value is demonstrated. This mirrors how the most successful predictive-maintenance programmes are built, and it keeps the first investment small and the first result measurable.

The end state is a shift in what the maintenance team does: less time spent coordinating reactive work, more time spent supervising a system that handles the operational layer, delivers operational efficiencies, and escalates only what needs a person. That is the move from scheduled downtime to condition-based intervention. Not a better alert, but a faster and more reliable response.