Inspection doesn't scale. AI output does.

Tricentis published survey data this week showing banks are releasing code into production without testing it. Not occasionally. Not in edge cases. As standard practice. AI-assisted development has accelerated code generation past the point where traditional QA cycles can keep up, so organisations are skipping them. If you've never stood on a production floor at two in the morning watching a press stamp parts faster than your inspectors can measure them, this probably sounds like a uniquely reckless software story. It isn't. It's the universal failure mode of any system where generation velocity exceeds verification velocity. I've watched it play out across automotive and aerospace for two decades.

The gap you can't staff away

Generation is automated. Verification isn't. That asymmetry is structural, and adding people to the verification side doesn't fix it — it delays the breaking point while burning money linearly against exponential output. When I built the greenfield QA/QC department at SNOP for 900+ employees, production ramped faster than any manual inspection regime could absorb. The conventional playbook said throw more inspectors at the line. For the first few weeks, that's what it demanded. But every inspector I added was fixed-cost overhead against a production rate that kept climbing. The math collapses. You hit a ceiling where inspection becomes the bottleneck, and someone — usually someone with a spreadsheet — decides the bottleneck is worse than the risk. That's what the banks are doing. They've decided QA is the bottleneck. The answer was never more inspectors. At SNOP we cut defect costs by 70% and hit 98% customer satisfaction not by scaling inspection but by eliminating the need for it. In-process controls. Mistake-proofing. Routing Verification KPIs that caught deviations at the source rather than thirty minutes later at end-of-line. The 97% internal lead-time reduction I delivered at Airbus came from the same principle: quality embedded in the process flow, not bolted on as a downstream gate. Every inspection step introduces its own time, cost, and failure modes.

Prevention scales. Detection doesn't.

The QRQC discipline I implemented at WITTE Automotive taught me something that applies to every generation pipeline, AI included. Rapid response built into the process rhythm beats post-hoc detection every single time. QRQC — Quick Response Quality Control — works because it compresses the loop between deviation and countermeasure to minutes, not days. It doesn't try to inspect quality into the product. It makes quality the default state of the process. This is the shift anyone deploying AI-generated output at scale needs to make. You cannot test your way to safe AI output any more than you can inspect your way to a zero-defect aircraft component. Verification capacity will always lag generation capacity, and the gap widens with every model upgrade, every parameter increase, every automated pipeline added to the stack.

You don't solve a velocity asymmetry by running faster on the slow side. You redesign the system so the slow side is no longer the one that matters.

What built-in verification looks like for AI output

PFMEA logic — Process Failure Mode and Effects Analysis — exists for exactly this situation. Map failure modes before they occur, rank them by severity and likelihood, engineer controls that make the failure impossible or immediately detectable at the point of generation. The same discipline applies to AI output pipelines, and it means architectural work, not prompt engineering or fine-tuning hopes. Start with constraint design. Constrain the generation space so the model can't produce certain classes of output — banned content, out-of-spec parameters, non-compliant language. This is poka-yoke applied to language models: design out the defect rather than catching it downstream. The model never gets the chance to fail in that particular way. I built MultiPS to handle the harder problem. Running 63+ models in parallel and synthesising their outputs surfaces divergence automatically. When models disagree, you've found the risk. When they converge, you have higher confidence than any single-model output can deliver. Verification embedded in the generation step itself, not appended after the fact. It costs compute, not headcount — and compute scales. Then there's the process layer. Actual architectural controls that enforce routing, logging, escalation, and human-in-the-loop breakpoints where the risk profile demands it. Same logic as a manufacturing process control plan: define the critical parameters, monitor continuously, have a predetermined response when they drift. The principle transfers from AS9100 environments to AI pipelines because the physics of failure are identical. Complexity increases, failure modes multiply, and the only control strategy that scales is one designed into the generation process itself.

The economics haven't changed

The defect cost curve is the same one it's always been. A defect caught in design costs one unit. Caught in process, ten. Caught by the customer, a hundred. AI didn't change that curve. It compressed the timeline between generation and field failure to something approaching zero. When a model generates output and that output ships directly to a customer without a human in the loop, design, process, and field collapse into a single instant. That's what makes the banking situation dangerous. Banks shipping untested code aren't taking a calculated risk. They've lost the ability to calculate the risk because the feedback loop that would inform the calculation has been removed entirely. The Tricentis data isn't a warning about banking. It's the canary in the mine for every organisation automating output without automating its quality architecture. The ones who recognise this as a process engineering problem — not a testing problem, not a tooling problem — will build systems that scale safely. The ones who don't will keep shipping untested output until the field failures arrive, and by then the time they thought they'd saved will look very expensive indeed.