AI in Inspection ERP — What It Actually Does, and What It Cannot

Start by separating the AI that helps from the AI that sells

Every business-software vendor now claims AI. For an inspection company the claim is worth testing against a single question: does it remove work a human was doing badly, or does it produce output a human must now check? The first saves money. The second moves the work and adds risk.

Below is the honest split, based on what the technology can actually do inside an inspection company's operations — certification records, calibration, scheduling, field capture and reporting.

Where AI genuinely earns its place

Report drafting from structured findings. When examination data is already captured as structured records — locations, readings, indications, acceptance criteria — generating the narrative sections of a report is exactly the kind of pattern work language models do well. The technician still reviews and signs; the blank page disappears. This is the single largest time saving available in most inspection businesses, because report turnaround is the bottleneck between finished work and an invoice.

Certification and calibration risk. A system that knows every technician's methods, levels and expiry dates, and every instrument's calibration due date, can forecast which jobs in the next quarter are at risk of having nobody qualified or nothing calibrated to run them. That is arithmetic with a good interface rather than intelligence, but it is the arithmetic nobody does until an audit forces it.

Scheduling under constraints. Matching crews to jobs across qualification, availability, travel and client-approval status is a genuine optimisation problem, and software beats a whiteboard once a company runs more than a couple of crews. Treat the output as a strong proposal a planner accepts or overrides, never as an instruction.

Anomaly review as a second pair of eyes. On high-volume encoded data — corrosion mapping, tube inspection — algorithms flagging regions that deviate from their neighbours help an analyst prioritise where to look first. As triage this is valuable. As a verdict it is not.

Where it is theatre

Dashboards that "use AI" to restate what a filter already showed. Chat interfaces bolted onto reports that answer questions the report already answers, more slowly. Predictive claims about equipment the system holds no condition history for — a model cannot forecast corrosion from an invoice ledger, however confidently it phrases the output.

The one place it must not decide

Acceptance. Whether an indication is rejectable under the governing code is a determination a qualified person makes and signs, and every code in this industry is written around that fact. A system may surface, sort, measure and draft — but the moment a machine's classification is recorded as the acceptance decision, the evidence chain that makes an inspection defensible is broken. Ask any vendor claiming automated acceptance who signs, and what happens in litigation when the answer is "the model".

Related reading: machine learning for defect detection · human-in-the-loop workflows.

What to ask a vendor

Three questions separate substance from marketing quickly. What data does the feature learn from — mine, or a pooled dataset? (If pooled, ask what leaves your tenancy.) What happens when it is wrong, and who sees that it was wrong? Can I turn it off and keep the underlying function? A feature that cannot be disabled without losing the workflow is not a feature, it is a dependency.

See how the operational layer is built at Atlantis: NDT inspection software · the ERP for inspection companies · the evaluation checklist to run vendors against.