Agentic AI is moving quickly from discussion to deployment in asset-intensive industries. Unlike generative AI, which summarizes information or drafts recommendations, agentic AI can take a goal, break it into steps, pull data from multiple systems, and act on it. It can review maintenance history, check open work orders, compare crew schedules and generate a recommendation without requiring a person to manually connect every piece of information.
That capability is exactly why so many organizations are moving quickly to adopt it. According to the 2026 Gartner CIO and Technology Executive Survey, only 17% of organizations have deployed AI agents, while more than 60% expect to do so within the next two years. That’s faster than for any other emerging technology included in the survey.
In the deployments I have supported, teams have long spent too much time switching between systems, spreadsheets, inspection notes, and emails just to make a single operational decision. Agentic AI promises to reduce that friction, helping planners prepare work faster and giving technicians a clearer picture of asset history without relying purely on institutional memory.
But that promise comes with a catch that is easy to overlook in the rush to deploy: from what I have seen, agentic AI is less a modeling problem than an operational trust problem.
In asset-intensive industries such as utilities, transportation, mining, telecommunications and manufacturing, an AI agent is only as good as the data it is acting on. If that data is incomplete, outdated or inconsistent, the agent does not necessarily fail loudly. It may fail confidently, recommending work on the wrong asset, missing a dependency, underestimating risk, or misjudging what should be prioritized.
Feed an agent bad data, and it will not scale intelligence. It will scale operational risk.
The Same Old Data Problems, at a New Speed
The data quality issues agentic AI exposes are not new. In more than 17 years working on systems and enterprise asset management deployments, I have seen the same issues surface repeatedly: duplicated records, missing asset relationships, inconsistent naming conventions, outdated criticality ratings, and poor failure coding for years.
What has changed is not the nature of these problems but their potential reach.
An experienced technician might catch a bad record before acting on it, drawing on years of institutional knowledge that was never captured in the system. An AI agent does not have that instinct. It can process thousands of records far faster than a person could review them manually, applying the same flawed assumption at scale before anyone notices.
The broader enterprise implementation problem is already visible. A preliminary 2025 MIT Project NANDA report found that 95% of the organizations studied had not achieved measurable profit-and-loss impact from their generative AI initiatives. The report attributed the divide largely to learning and integration gaps, including tools that did not retain context or adapt effectively to existing workflows. In asset-intensive operations where a bad recommendation can affect physical equipment rather than just a dashboard, that gap between AI ambition and data reality is even more consequential.
Where Bad Data Becomes an Unsafe Decision
In operational settings, those consequences become physical.
Take maintenance prioritization. If an agent cannot see a recent failure, a delayed work order, or a change in an asset’s condition, it may assign the wrong priority. This can create unnecessary work in one area while allowing a more significant risk to remain unresolved elsewhere.
The stakes rise further when asset relationships themselves are incorrect. This is the scenario that raises a red flag. If a system does not accurately reflect which assets are connected, what equipment must be isolated before work begins, or what constraints apply at a particular site, its recommendations can become genuinely unsafe.
That is why organizations should think carefully before giving agentic AI full autonomy over critical decisions. The voluntary NIST AI Risk Management Framework offers a structure for incorporating trustworthiness throughout an AI system’s lifecycle rather than addressing it only after deployment. AI can gather information quickly, surface options, and summarize the evidence behind a recommendation. However, when safety, compliance, or service reliability are at stake, people need to remain involved.
At least initially, the goal should be AI that supports decisions, not AI that makes them alone.
Why Enterprise Asset Management Is the Foundation, Not the Back Office
Enterprise asset management, or EAM, is where operational reality is organized. It is where work orders, maintenance strategies, inspections, approvals, asset relationships, and equipment histories are recorded and managed.
A strong EAM foundation gives an AI agent reliable operational data to work with. A weak one is likely to be exposed almost immediately. An EY analysis of AI-enabled enterprise asset management similarly emphasizes the role of real-time data and analytics in moving organizations from reactive to proactive maintenance.
Rather than treating EAM as a back-office system, organizations preparing for agentic AI should treat it as part of the operational control layer. It shapes what the AI can access, what processes it must follow, what evidence it relies on, and where human approval is still required.
EAM also helps provide the context behind an operational decision. A maintenance recommendation should not be based only on the condition of an individual asset. It may also depend on the asset’s criticality, recent failures, connected equipment, available parts, safety procedures, crew availability, and the effect that downtime would have on the wider operation.
Without that context, even a technically sophisticated AI agent may recommend the wrong course of action.
Is Your Data Actually “AI Ready”?
Rather than launching a broad and abstract data-quality review, organizations are better served by starting with a specific use case, such as AI-assisted maintenance planning, and asking targeted questions.
These are the questions I ask first: Do we trust the asset hierarchy? Is the failure history reliable? Are job plans current? Is crew availability accurate? Are safety plans up to date?
Data does not need to be perfect. Few organizations will ever have perfect operational data. It needs to be accurate, current, and well managed enough for AI to genuinely support the people making decisions.
This approach also makes data improvement more manageable. Instead of attempting to clean every record across the organization, teams can focus on the information that directly affects the selected use case.
The risks of getting this wrong are not evenly distributed. Industries where digital decisions affect physical operations carry the greatest exposure, which is why NIST is developing a dedicated AI Risk Management Framework profile to guide critical infrastructure operators using AI-enabled capabilities. In a lower-risk process, a poor AI recommendation might cause inconvenience or additional work. In a power grid, water system, mine, transportation network, or telecommunications environment, it can affect safety, cause downtime, create compliance problems, or disrupt service.
That does not mean these industries should avoid agentic AI. It means they need to be more deliberate about data quality, governance, and human oversight before increasing its responsibilities.
A Practical Path to Agentic AI
A practical implementation path begins with a specific use case where AI can add value without introducing excessive risk.
From there, the organization should map the data the use case depends on and honestly assess whether that information can be trusted. Teams should then correct the most critical asset data, assign clear ownership, and standardize how work, failures, and operational changes are recorded.
The AI should operate within existing governance processes rather than around them. In the early stages, the agent can support planners, reliability teams, and technicians by gathering information, identifying patterns, and recommending possible actions.
This gives the organization time to compare the agent’s recommendations with real outcomes, identify data gaps, and confirm that the necessary controls are working.
Only after the results are understood should the organization consider giving the AI greater responsibility. Autonomy should increase gradually and only when the associated risks are clear, controlled, and auditable.
A successful agentic AI strategy links AI to trusted asset data and strong EAM workflows, supported by clear governance and human accountability. Deloitte’s 2026 State of AI in the Enterprise reaches the same conclusion from the top down, describing a unified, trusted data foundation as essential to scaling AI. Organizations remain responsible for the actions AI recommends or takes. That responsibility does not disappear simply because a decision was generated by an algorithm.
Agentic AI should earn its way into operations, not be handed them. Start narrow, check its recommendations against what actually happens, fix the data gaps that surface, and expand its role only when the evidence justifies it. I would rather see an agent earn a small amount of trust and keep it than be given a large amount and lose it. In asset-intensive industries, that patience is the difference between AI that earns trust and AI that quietly loses it.
