A metals-sector client of Deloitte cut its mean time to repair by 45% through a maintenance routine driven by AI agents, says Tim Gaus, principal at Deloitte Consulting. Other deployments have produced no meaningful change, he cautions. That contrast puts the industrial AI opportunity in practical terms: value depends on connecting intelligence to work that people can act on.
Repair Performance Depends on What Happens After Prediction
Deloitte’s AI Institute report, The AI Dossier, examines applications across industrial assets, infrastructure and worker safety. Gaus describes a broad technology range, encompassing machine learning and deep learning alongside large language models, agentic systems and physical AI. For plant leaders, those categories matter less than the operational task each can support.
Maintenance offers a clear test. Predicting an equipment failure is useful only if the warning reaches the right worker and helps determine the next action. An agent can extend that workflow by monitoring sensor inputs, identifying anomalies, supporting diagnosis or generating and scheduling work orders. Simulation and reinforcement learning provide further potential applications.
Gaus also identifies a less visible opportunity: helping teams connect machine information with maintenance histories when those records are difficult to access and compare. That addresses an information bottleneck rather than a mechanical problem. For a reliability team, easier comparison could make existing evidence more useful without requiring a fully autonomous maintenance system.
A distribution-sector client illustrates how the pieces can fit together. Gaus says the company combined downtime prediction with agents that deliver findings to employees across a large asset portfolio. Users can adjust the system’s output, reducing irrelevant notifications. That feedback mechanism makes adoption part of the technical design.
Alert volume is a deployment issue, not just an interface detail. A system that repeatedly interrupts technicians with weak recommendations risks becoming another source of operational friction. Gaus’s example points toward a more useful model: predictions delivered through a workflow that employees can refine.
Before selecting a maintenance pilot, Gaus recommends checking:
- The operational problem. Start with a specific task or failure pattern, then assess whether AI can help.
- Data readiness. Use information the team trusts, without making perfect data a prerequisite for experimentation.
- Worker involvement. Design around the people receiving recommendations and carrying out repairs.
- Expansion potential. Consider whether the approach can be repeated across other assets or locations.
- IT and OT collaboration. Bring technology specialists together with those who understand plant operations.
These choices explain why a promising demonstration may fail to become a useful operating capability. Better predictions alone don’t resolve unclear responsibilities, inaccessible records or excessive alerts. The investment needs a path from model output to maintenance action.
Adoption Is Expanding, but Deployment Remains Bounded
A Manufacturing Leadership Council survey from early 2025 found that 22% of manufacturers expected to deploy generative and agentic physical AI within two years, compared with 9% already using it. Those figures describe intentions as well as current adoption; they do not establish that the expected deployments will happen.
Gaus says Deloitte’s research shows adoption growing, although much of the existing activity remains confined to narrowly defined settings. That is consistent with a practical deployment direction: solve a bounded problem, learn how employees use the result, then decide whether expansion is justified.
For executives, the distinction between experimentation and repeatable operation is central. A pilot can demonstrate technical capability while leaving maintenance burden or workforce requirements unresolved. The next investment decision should therefore examine whether the application can remain useful beyond its initial setting.
Knowledge transfer provides a particularly concrete example. Gaus describes a company facing retirements among experienced capital engineers involved in program implementation. Their expertise existed partly in personal judgment and partly in accumulated emails and files. Losing access to either would leave less experienced engineers with a thinner base for decisions.
The company used interviews to capture tacit expertise and combined that material with existing documents. An AI agent then made the collected knowledge available through conversation and automated document review. Junior engineers could consult the system rather than search separately through scattered records.
This approach changes how institutional knowledge is accessed. It can provide some of the review support associated with a senior engineer, as Gaus describes it, but that should not be treated as proof of equivalent judgment. A sensible deployment would make clear where the agent can assist and where an experienced person must still assess the result.
Mobile Inspection Extends AI Into the Facility
Physical inspection follows a different technical path. Gaus notes that industrial vision sensors have existed for 25 years; the shift is in their affordability and ease of deployment. He offers an illustrative price contrast of perhaps $20 against $40,000, rather than a universal cost benchmark.
Mobility adds another layer. Robots on wheels or legs can carry sensors through a facility, collecting images, measuring temperatures and reading gauges. That gives inspection systems access to places that are awkward for employees to reach or easy to overlook during routine work.
For maintenance teams, the practical value is more consistent observation. Gaus describes the familiar situation in which a technician arrives after a breakdown and discovers evidence, such as leaking oil, that might have been visible earlier. Automated patrols could help surface those signs before the opportunity to intervene has passed.
The business consequence differs from that of a knowledge agent. One makes existing expertise easier to consult; the other increases the opportunity to collect physical evidence. Both still need a route into human decisions.
| Application | How it enters the workflow | Deployment consideration |
|---|---|---|
| Agent-supported maintenance | Connects monitoring and diagnosis with work-order activity | Employees need actionable findings and manageable alerts |
| Engineering knowledge agent | Supports questions and document review using captured expertise | Its assistance needs a clearly defined scope |
| Robotic inspection | Collects readings and images while moving through a facility | Safe operation and ongoing maintenance must be planned |
Humanoid robots attract attention, but Gaus says their technical complexity makes them unsuitable for broad deployment at present. His expectation is that collaboration between people and robots will change more substantially than robot shape. More independent machine decisions may become part of shared workplaces, with corresponding demands on safety controls.
The Investment Case Must Include Upkeep and Workforce Skills
Gaus cautions that some applications remain too immature for the job. Devices carrying AI also introduce maintenance requirements that employees may not previously have faced. Technical feasibility therefore leaves several questions open: can the system be sustained, can the organization maintain it, and does the return justify the commitment?
Robotics purchasing models add a financial consideration. Gaus notes that many advanced robotics providers offer robotics-as-a-service, often treated as operating expenditure rather than capital expenditure. That changes how organizations assess returns and should be reflected in the investment case alongside operational benefits.
Workforce planning belongs in the same discussion. Robotic inspection may reduce difficult physical observation tasks, while knowledge agents may help newer engineers access accumulated expertise. Neither removes the need to decide who supervises outputs, maintains the technology or handles exceptions.
Industrial AI matters now because adoption expectations are rising while practical examples are already exposing what successful deployment requires. The strongest investment case connects a specific operational problem to a usable workflow, then accounts for safety, skills and upkeep. For plant leaders, that means judging the technology by its contribution to sustained maintenance performance, rather than by the sophistication of the demonstration.
