Same Radiograph, Two Calls
Nobody in the NDT industry pretends that two certified readers always reach the same conclusion on the same film. Human-factors research into radiographic and ultrasonic interpretation has documented reader-to-reader variability for decades — it's a known, studied property of the method, not a secret and not an accusation against any individual inspector. Fatigue, time pressure, viewing conditions, and plain experience differences all move the needle on a borderline call.
Weld RT in Canada is built for exactly the kind of volume where that variability shows up. CSA Z662 requires a minimum 15% of every production weld poured each day to be inspected, on top of 100% coverage on the highest-risk welds — road, rail, and water crossings, and anything that won't be pressure-tested. On a large right-of-way project, that's a lot of film moving through a lot of reads, often on rotating crews, often at the end of a long shift. That's the environment where inter-reader variability is most likely to bite — not because anyone's cutting corners, but because that's when human judgment naturally drifts.
This is the part of the workflow an AI second read is actually good at fixing, and it's a narrower, more honest claim than "AI is more accurate than a person." The value isn't that a model out-judges a CGSB Level II or III on a genuinely ambiguous indication — it's that it applies the exact same criteria, at 2am on hour ten, that it applied on the first film of the shift. Consistency, not superiority. Every film gets a second, unfatigued opinion before a human signs off, and that second opinion is logged and attached to the record.
That logged second read matters as much for what happens after the inspection as during it. If a weld's read is ever questioned — during a CWB audit, an owner's oversight review, or a CER investigation — having a documented, consistent AI read alongside the certified human read is a materially stronger record than the human call standing alone.
Where this goes next: a consistent second read is only as useful as the record it produces. If your shop's current audit trail is a folder of scanned reports and an email thread, the workflow automation side of this — structured, queryable records tied to every film — is usually worth building at the same time as the AI read itself.