Shell and C3 AI are extending predictive maintenance beyond the alert. Their next generation of agents is intended to investigate equipment anomalies, identify likely causes and prepare the operational steps needed to respond across a vast industrial estate.
Shell already monitors around 30,000 assets with predictive systems. That scale creates a second problem: detecting an unusual signal is only the beginning. Engineers still have to combine sensor history, maintenance records, operating conditions, spare-parts data and expert knowledge before deciding what to do.
Agentic AI promises to connect those fragments. In a refinery or energy facility, however, a plausible explanation is not enough. Every recommendation must respect physical safety, process constraints and established authority.
Predictive maintenance needs a complete chain of action
Traditional systems can warn that a pump, compressor or valve behaves differently from its baseline. The alert may arrive early, but it can still leave teams with hours of diagnosis and coordination.
C3 AI’s proposed agents can explore root causes, search previous incidents and draw on both operational technology and enterprise systems. The workflow may then draft a work order, check inventory and begin procurement steps for a replacement component.
This is where the economic value can grow. Avoiding an unplanned shutdown matters, but so does ensuring the right technician, procedure and part are available during the maintenance window. An insight that never reaches execution has limited operational value.
Connecting industrial data also exposes inconsistencies. Sensor identifiers, asset hierarchies and SAP records may describe the same equipment differently. Before autonomy, organisations need a reliable mapping between the machine in the field and its digital record.
Human approval defines the first stage of autonomy
Shell’s approach keeps people in control while the technology proves itself. Agents can assemble evidence and propose actions, but engineers approve decisions with safety, production or financial consequences.
That boundary is sensible because an apparently routine work order can interact with a wider process. Taking one asset offline may alter pressure, capacity or risk elsewhere. Local optimisation must not override plant-level judgement.
Teams need to see why an agent suspects a cause, which records it used and how confident it is. They should be able to distinguish measured evidence from an inference and record why they accepted or rejected the recommendation.
Over time, low-risk actions may become automatic. An agent could gather documents or reserve an available part without further permission, while equipment isolation and procurement commitments retain explicit approval. Autonomy should expand by action type, not through a single switch.
Auditability supports both safety and learning. A complete trail can show whether a failure came from sensor quality, an incorrect asset relationship or weak reasoning. The resulting corrections improve the whole maintenance system rather than one alert.
The last mile determines industrial AI value
Industrial AI has often been strongest at prediction and weakest at organisational follow-through. Agents are attractive because they can bridge that last mile between an analytical model and the enterprise processes that schedule work.
Success should be measured through avoided downtime, fewer emergency orders, planning accuracy and the quality of completed maintenance. Counting agent recommendations would reward activity without proving that equipment became more reliable.
Cybersecurity and separation of duties are essential. An agent that reads operational data does not automatically need permission to alter controls. Procurement authority should be limited, and every integration should have a defined failure mode and rapid revocation path.
The Shell project shows why industrial agents will develop differently from consumer assistants. They operate more slowly, face stricter evidence requirements and depend on specialists who understand the physical system.
If the deployment succeeds, its importance will come from disciplined coordination rather than dramatic machine independence. The agent notices, investigates and prepares; the organisation retains responsibility for acting. That partnership can turn predictive maintenance from a stream of warnings into a repeatable operating capability.
