Can AI Replace the Inspector? The Line Between Preparing and Deciding

The question comes up at every demonstration of an artificial intelligence tool in maintenance, and it is asked with suspicion far more often than with curiosity: is this machine going to decide in my place, and who carries the responsibility when it gets it wrong. It is a good question. It deserves better than a sales answer.

The answer fits in one sentence that the rest of this article merely expands: an AI prepares, a human decides, and that line is not a temporary precaution while we wait for the technology to catch up. It is the very structure of a job where a single misreading can leave in service a piece of equipment that should not be.

What AI genuinely does well

We have to start with the concrete, because suspicion often stems from vague promises. On inspection and maintenance data, AI excels at four precise tasks, all of them sitting upstream of the decision.

It reads unstructured documents. An inspection report in PDF format, a scanned contractor write-up, a technical datasheet: it pulls out the equipment tag, the recorded measurements, the findings and the due dates, and files them in a usable form. This is the re-keying work that nobody wants to do and everybody puts off, with scattered records as the consequence. The topic is covered in detail in extracting data from a PDF inspection report.

It retrieves from a large history. Every asset that has shown the same damage mechanism, every job tied to a common cause, every condition monitoring location whose wall thickness has dropped faster than expected. On a fleet of several hundred assets, that is the difference between a half-day dig and an immediate answer.

It connects scattered pieces of information. The same anomaly reported in three different reports by three different contractors, under three different wordings, that a human would not have linked for want of having read them on the same day.

It formats. A structured write-up from notes, a turnaround summary drawn from dozens of jobs, a note for management built from raw data. The form is prepared, the content still needs checking.

None of these four tasks decides anything. All of them free up time on work that never needed human judgement: document handling and searching.

What it must never decide on its own

The line reads as the mirror image of the tasks above. Wherever a conclusion bears on safety, compliance or operation, the decision stays human.

An AI can prepare a ranking of assets by risk. It does not decide which one gets inspected first: that choice commits a maintenance policy and a responsibility. It can calculate a degradation rate and a due date. It does not pronounce the fitness-for-service of a piece of pressure equipment, which falls to a competent person and a regulatory framework. It can flag that a wall thickness is approaching the retirement limit. It does not decide whether the asset stays in operation, nor the date of a shutdown.

That line is not a cautious choice we would later drop. It follows from the nature of the consequence. In an office role, a mistake produces a document to redo. In inspection, a misreading can lead to running a degraded asset, with consequences for safety, the environment and contamination. That asymmetry does not disappear as the model improves. A better model gets it wrong less often, but every error stays just as serious, and nothing in a model carries legal responsibility.

Why human verification is not a formality

A tool that processes three hundred reports will get some of them wrong. That is not a flaw in a particular piece of software, it is a property of any system that interprets imperfect documents: a skewed scan, a handwritten annotation, a misaligned table, an ambiguous unit.

The skill to build, then, is neither to trust blindly nor to redo everything by hand. It is to know where to look first. The outliers, the wall thicknesses that rise, the inconsistent dates, the near-identical equipment tags that betray a mix-up. This is exactly the reasoning an inspector already applies to a measurement campaign, carried over to a document process. That skill is developed in the skills of the maintenance manager.

This verification has a cost, and it has to be said, because the savings calculations often forget it. AI does not bring processing time down to zero: it shifts it from tedious re-keying towards targeted verification, shorter and more skilled. That is a good trade, but it is a trade, not an elimination.

The same line, three industries

The rule is the same everywhere, but what it protects changes in nature by sector, and a concrete example beats a principle.

In petrochemicals, AI extracts the wall thicknesses of hundreds of points on the columns, exchangers and kilometres of piping in a unit. It calculates the rates, it flags the points that drop away. What it does not do: decide to extend a campaign until the next turnaround, a trade-off that weighs a production loss of several million against a risk of loss of containment. That trade-off falls to the inspection manager.

In pharmaceutical, it can connect the qualification records of a sterile vessel and an exchanger, and flag a slow drift. It does not pronounce the requalification, which commits batch compliance and patient safety. The criticality of a defect on equipment in contact with an injectable is judged, it is not calculated.

In food and beverage, it can read all the maintenance write-ups for a spray-drying tower and retrieve past jobs on its cladding. It does not replace the physical inspection that looks for a micro-crack towards the insulation, because the risk is to health and invisible, and no document holds it until someone has gone to look.

Three sectors, three consequences, a single line: the machine prepares the case, the human carries the decision.

What the regulatory framework says

On this point, the texts are clearer than people think, even though they were not written for AI.

The pharmaceutical quality risk management framework holds that any assessment ties back ultimately to the protection of the patient, and that the decision commits an identified human responsibility. The framework for manufacturing sterile medicines requires that the criticality of a defect be determined in light of its impact on the patient and the route of administration: that is a judgement, not something an algorithm can compute. In pressure equipment, periodic requalification falls to authorised bodies and competent persons, terms that designate accountable humans, not systems.

None of these frameworks forbids using a tool to prepare a decision. All of them assume that a person owns it. A piece of software that produced a regulatory conclusion without human validation would not merely be risky: it would be unusable, because no one could rely on it in front of an auditor.

So, does AI replace the inspector?

No, and the way the question is framed lights up the answer. What disappears is the time spent searching, re-keying and formatting. What remains, and gains in value, is the interpretation, the judgement on a doubtful case, the decision that commits.

The job shifts rather than shrinks. An inspector whose tool prepares the ground spends less time assembling a file and more time deciding on the cases that deserve it. That is a step up in skill, not a disappearance. The real threat, for a professional, is not the tool: it is not knowing how to use it while others learn.

There remains one category of site where caution must be at its highest: those where the consequence of a defect is to health and invisible, such as cross-contamination through a loss of equipment integrity. There, no automation replaces human control of integrity, and the subject goes beyond report reading to touch on physical inspection. This case is covered in tracking the condition of industrial equipment.

• Handing a decision to the tool because it is right most of the time.How often it is right changes nothing about the seriousness of the remaining error, nor the fact that no one owns it in law.
• Accepting an output you cannot verify.That is signing a document you have not read. The verification skill is the real prerequisite for using it.
• Believing verification will disappear as the model improves.A better model gets it wrong less often, not less seriously. The line follows from the consequence, not from performance.
• Underestimating verification time in the savings calculation.AI shifts the time, it does not remove it. An honest calculation counts the verification.
• Presenting a tool as decision-making to sell it.A piece of software that claims to decide on its own in regulatory inspection is unsellable to anyone who knows the framework: no one could rely on it.