AI agents in industry: what they do, what to fence

Over the past two years, industry has learned to hold a conversation with its data: a tool summarises a report, retrieves a history, drafts a write-up. Useful, but it waits to be asked and stops at the suggestion.

The next step has a name: the agent. An agent does more than answer. It monitors, reasons and acts within a defined scope. It is the industrial topic of 2026, and it is also the one that warrants the most caution.

Copilot or agent: the real difference

Both rest on the same models; the difference is not technical, it lies in the power to act.

A copilot responds to a request. You ask a question, it proposes an answer, and nothing happens until you take back control. This is what generative AI applied to industry deployed in 2025 does: it prepares, it suggests, it leaves the decision to the human.

An agent goes one step further: it monitors a data stream continuously, reasons over what it observes, then triggers an action within a scope set in advance. This is a change of kind. Where the copilot writes a sentence, the agent creates a record, changes a plan or opens a ticket.

Slim on screen, this difference is considerable in terms of accountability: a system that acts must be bounded, traced and monitored far more strictly than a system that merely proposes.

What an agent already does on the factory floor

Three concrete cases, all within a narrow, well-defined scope:

• Maintenance. An agent monitors the vibration signatures of a rotating machine. When a signature degrades beyond a threshold, it does more than raise an alert: it creates the work order in the maintenance system, with the context of the measurement. The human receives a documented ticket, not a bare alarm.
• Quality. An agent triages the defects reported from a line: it matches a finding against past cases, proposes a category, routes it to the right channel. The triage, repetitive and time-consuming, is prepared before an inspector decides.
• Supply. An agent tracks supplier lead times. When a supplier slips, it rebalances the supply plan within the limits it has been given, or proposes a trade-off if the gap falls outside the frame.

These three cases share a common trait: narrow scope, clear input data, a modest and reversible action. That is where agentic AI delivers something tangible today, not in the fantasy of a factory that runs itself.

2026: agentic AI moves into production, carefully

2026 is the year agentic AI tips over from the lab into commercial offerings. For most sites, the dominant project is not starting from a blank page: it is evolving the copilots installed in 2025 towards agentic execution, one use case at a time.

Deloitte's 2026 report, "State of AI in the Enterprise", measures the gap: close to three companies in four plan to adopt agentic AI within two years, but only about one in five says it has a ready operating model. Appetite is running ahead of the foundations.

Gartner's finding points the same way: more than 40% of agentic AI projects could be abandoned by 2027, for three recurring reasons: unclear business value, underestimated costs and poorly anticipated complexity.

These figures do not disqualify agentic AI: they are a reminder that a project launched without a precise use case, without reliable data and without governance joins the abandonment statistic. Caution here is the condition for the project's survival.

What to automate and what to fence

The sorting principle is simple to state and demanding to apply: automate the reversible and well-bounded decisions, put a human barrier in front of the irreversible, safety-related or integrity-related ones.

A reversible decision can be undone: creating a useless work order costs a few minutes of sorting, reordering a queue of non-critical tasks puts nothing at risk. Good candidates for automation, because the error, when it occurs, stays cheap and visible.

An irreversible, safety-related or integrity-related decision does not fall into the same category. Declaring a pressure equipment item fit for service, validating a requalification, shutting down or keeping an operation running: these decisions carry a responsibility that no system bears in place of a competent person. In front of them, the agent prepares the file, but a human decides.

This boundary is the one that already structures human validation of AI outputs. Agentic AI does not move it, it makes it more critical: a system that acts without validation spreads its error further and faster.

Agent governance and security

An agent plugged into the systems that run the plant, MES, ERP or SCADA, is no longer an assistant: it is an actor in the industrial information system, to be treated as such.

The sector is taking the subject seriously. In 2026, the Open Secure AI Alliance, which brings together the likes of NVIDIA, Microsoft, Siemens, SAP, Hugging Face and CrowdStrike, is working on precisely four building blocks: the identification of agents, the management of their permissions, their isolation and the traceability of their actions.

In practice, two requirements come first. First, bounded permissions: an agent acts only on the scope explicitly entrusted to it, never beyond. Second, full traceability: every action must trace back to the data that triggered it and the rule that authorised it, failing which no audit is possible. These requirements extend those already documented by the European AI regulation for sensitive uses.

They also connect to the classic risks of AI in industry: a poorly cordoned agent opens one more path between office IT and the production network, and an over-permissive agent becomes a route for data leaks or shadow AI that no one had mapped.

There remains the most honest limit, and one often left unsaid: the real bottleneck of agentic AI is not the model, it is data quality, the data foundation on which the agent reasons. An agent that acts on bad data acts wrongly, simply faster. And the orchestration of several coordinated agents, heavily showcased in demos, remains rare in production.

An agent is only as good as what it reads, and structured inspection data is part of that. This is the niche of a platform like Integrity Loop: documentary intelligence on inspection reports, turning scattered PDFs into structured, historised and searchable data. Not a CMMS, not an ERP, not a digital twin, but one building block of the reliable data layer an agent requires before any automation.

Getting started without getting burned

• Automating an irreversible decision to impress.The demo flatters, but a safety decision handed to an agent shifts a responsibility that cannot be delegated.
• Plugging an agent into unreliable data.An agent that reads bad data acts wrongly faster. Data quality is dealt with before automation, not after.
• Granting permissions that are too broad.An agent that can act beyond its scope becomes a security risk and a leak route. Permissions are bounded to the strict minimum.
• Neglecting traceability.An action that cannot be traced back to its cause is neither auditable nor correctable. Without a log, the agent has no place in a decision chain.
• Confusing pilot with roll-out.Aiming straight for multi-agent orchestration, still rare in production, exposes you to the complexity and costs that sink close to one project in two.