AI drew your CAD model. Your APQP still assumes a human did.

The first time I watched a 3D scan become editable CAD in under four minutes, my thought wasn't about the geometry. It was about the revision history. Backflip AI's CAD Copilot does what the vendor demos promise. Point a scanner at a physical part, feed the mesh to the algorithm, receive parametric CAD with a full feature tree – geometry that opens in CATIA or SolidWorks and behaves as though a drafter spent three days building it. As Head of Manufacturing Engineering Technical Authority for Airbus North America, I own the technical integrity of every manufacturing engineering decision in our scope. So when a tool produces a master model from a point cloud, I don't first ask whether the surfaces are accurate to ±0.05 mm. I ask who made the engineering decisions embedded in that geometry, and whether our change control process even knows a decision was made.

The scan captures what is. The CAD decides what should be.

These are different engineering activities, and the AI conflates them. A structured-light scan of a stamped bracket captures the part as manufactured – springback included, die wear included, every deviation frozen in the point cloud. That scan is a measurement. When the algorithm "cleans" it into nominal geometry, it makes a choice: which deviations are intentional design features, which are manufacturing error. That distinction used to require a PFMEA, a tolerance stack-up, a dimensional layout report, and a sign-off from a design engineer who understood the functional intent behind each surface.

I have worked under IATF 16949 and AS9100 for two decades. Both frameworks assume the person creating or revising a model has a reason – a design intent they can articulate in an 8D or defend in a design review. The algorithm has no intent. It has a loss function and a prior trained on a dataset of parts it has never seen your specific requirements for. When it fillets an edge your drawing specifies as sharp, or smooths a blend that exists to manage stress concentration, it is not making an error in its own terms. It is optimising for geometric plausibility. Your stress team does not review plausibility. They review intent.

Probability is not a substitute for purpose, and a clean surface is not a clean audit.

Your ECN system was designed for humans

Engineering change notice systems – whether you run them in Teamcenter, Windchill, or a SharePoint list with good intentions – are built around a simple assumption: someone, deliberately, decided to change something. That person documented the reason. The reason was reviewed. The review was approved. The approval propagated downstream.

AI-generated geometry breaks this chain because the changes are stochastic and invisible. The model updates between scan sessions because the algorithm refined its reconstruction. Nobody files a change, because nobody knows a decision was made. I built the Routing Verification KPI framework at Airbus that cut internal lead time by 97%, and the reason it worked is that every step in the digital thread is explicitly verified against the step before it. The moment you insert an AI-generated model as the master without an equivalent verification gate, you have a gap in the thread that your KPIs cannot see. The routing looks clean. The data underneath is a statistical best guess dressed up as a controlled drawing.

I design and deploy autonomous multi-agent AI systems – currently running 63-plus models in parallel with consensus synthesis. Build enough of these and you learn that the most dangerous output is not the obviously wrong one. It is the one plausible enough to pass a visual check and wrong enough to fail a tolerance stack-up. Scan-to-CAD tools produce exactly that category of output at industrial scale.

Every downstream artifact inherits the gap

CNC programs reference the master model. CMM inspection plans derive coordinates from it. Workholding design, tooling paths, operator work instructions – all trace back to the geometry file. When I ran quality at SNOP across a 900-person greenfield plant, we built the control plan on the principle that every inspection characteristic maps to a design intent. If the master model was generated by an algorithm and quietly revised without an engineering review, your CMM is inspecting against geometry that was never validated against what the part is supposed to do.

This is not theoretical. In aerospace, a surface profile tolerance of 0.25 mm against an aerodynamic fairing exists for a reason that lives in a wind tunnel report, not in the point cloud. If the AI smooths that surface because its training data suggests aerodynamic parts should look clean, you get a model that machines beautifully, passes inspection against itself, and fails its functional requirement. QRQC catches it on the shop floor eventually. By then you have machined 400 parts. The ECN system shows no discrepancy, because it was never told a change occurred. Your audit trail is clean. Your parts are wrong. I have seen those two conditions coexist often enough to know which one costs more.

Scan-to-CAD belongs in the toolkit. I use AI systems daily, I build them, and I believe they belong in manufacturing engineering. The question is not whether the technology works. It does. The question is whether your engineering change process can tell the difference between a decision a person made and a decision a model made for them. APQP was written for engineers with intent. AI delivers geometry with probability. If your quality system cannot distinguish the two, the algorithm is not accelerating your process – it is quietly becoming your source of truth, one invisible revision at a time.