Why We're Not Trying to Replace Your Level III
Every conversation about AI and weld inspection in Canada eventually gets to the same question, usually asked carefully: is this trying to replace the Level III?
No. And it's worth explaining why that's not a hedge — it's how the system has to work.
Under CSA Z662, a CGSB Level III has to review and approve the NDT procedures behind every inspection. That's not a formality a vendor can automate around, and nobody serious is trying to. The certification requirement exists because a signature attaches accountability to a specific qualified person, and that accountability is the entire point of the regulatory structure. An AI model doesn't hold a CGSB certification. It shouldn't be asked to, and it can't stand in for one.
What actually eats a Level III's day isn't the judgment calls — it's the volume of images that don't need one. A busy pipeline project pouring welds across a long right-of-way generates radiographs continuously, and a large share of them are clean. Someone certified still has to look at each one before it's accepted, because CSA Z662 doesn't grade on "probably fine." That grind — steady, high-volume, low-variance — is exactly the kind of work a second-reader model is good at doing first.
HotPass AI reads every image before your team does and sorts them: here's what's almost certainly clean, here's what needs your eyes first. The Level II or III still makes every call. What changes is which images they see first, and how much of their day goes to scrolling through frames that were never going to be the interesting ones.
This is also why "augment, don't replace" isn't a marketing line here — it's a compliance requirement. The sign-off stays exactly where CSA Z662 puts it. What HotPass AI removes is the time between the image landing and a qualified person actually getting to look at the ones that matter.
Where this goes next: if the real constraint on your shop is throughput — getting more certified eyes on the images that need them, faster — the AI read is one half of that. The other half is usually a workflow problem: how images get routed, queued, and packaged for review in the first place. Enrisk builds that half too, and it's often the bigger time-saver of the two.