The governance is budgeted. The gauge study isn't.

Every plant I have ever set foot in had one instrument that settled every argument. Early in my career it was the CMM in a stamping plant's metrology lab. Two shift leaders disagreed about a bore diameter, they walked the part upstairs, and whatever the machine printed became the truth. The machine had cost €400,000. Nobody had ever run a gauge study on it. We measured everything with it and never once measured it. This week the trade press has discovered that the hidden cost of AI is governance: IT Pro says so in as many words, Planview has hired an executive advisor to sell the fix, MIT Sloan is pushing urgency on ethical adoption. Three headlines, one pattern. The industry has correctly diagnosed a measurement problem and answered it with a purchase order. The trusted, unstudied gauge never went away. It has been reborn as the corporate copilot.

A model is a gauge nobody ever studied

Two decades in automotive and aerospace quality drilled one habit into me: qualify the instrument before you trust a single reading. Under IATF 16949, any gauge touching a critical characteristic needed a measurement-system analysis before first use. Repeatability, the same operator running the same part ten times over. Reproducibility, three different people on the same gauge. Stability, tracked for months against a reference; bias, compared to a known master. If the gauge's own variation ate 30% of the tolerance, the auditor stopped you using it, and rightly so. You were no longer inspecting parts. You were inspecting noise.

Now map that discipline onto the instrument your company actually runs on. Ask the copilot one question five times and watch the answer migrate. That is repeatability failing in real time. Two engineers phrase the same request differently and pull materially different conclusions – on the shop floor we would call that a reproducibility failure; in the boardroom it gets called a prompt problem. Stability: the model is re-versioned under you mid-shift, no calibration event, no log entry. Bias: nobody maintains a reference standard, so nobody can say. A gauge R&R on a bore gauge cost me a day of a technician's time and a few hundred euros. I have yet to meet the company that has spent the equivalent on the model drafting its supplier evaluations and its 8D narratives.

An instrument nobody has studied is not a gauge – it is a rumour with a calibration sticker.

Consensus is not independence

I built MultiPS to run 63-plus models in parallel on a single question and synthesise their answers, because I do not trust single readings. I want the spread, not just the average. Even in private beta, the lesson holds: agreement is data, not verdict. Most large models are built on substantially the same corpus – the same crawl, the same forums, the same answer threads – so their errors correlate. Averaging correlated instruments does not cancel error. It launders it. When 58 of 63 models agree, the instinct is to call that a verdict. Resist it. Fifty-eight instruments raised on the same corpus have agreed; the number describes the corpus, not the answer.

The dissenting five are where the money is. At SNOP we took 70% out of defect cost, and little of that came from adding inspection. It came from treating small anomalies as signals while containment was still cheap. The same reflex applies. When the minority dissent clusters – five models objecting for overlapping reasons – that usually marks the blind spot the majority shares. My ethical-hacking side reads it the same way from the other direction: the vulnerability worth hunting is never in the room everyone watches, it is confident agreement in a channel everyone trusts. Three appraisers using the same deformed fixture will show you beautiful reproducibility and reject every good part. Correlated error dressed as consensus has fooled better auditors than any AI council.

What a governance budget should actually buy

I have nothing against governance spending. I object to buying the wrong thing with it. A control plan for AI outputs costs little and changes everything: for every output that touches a decision, write down what it may decide, how it is verified, how often, and who reacts when it drifts. Those four lines have carried automotive control plans since the nineties. Drift containment is QRQC muscle memory. When a model starts producing degraded readings, contain within 24–48 hours, quarantine every output since the last verified check, and run an A3 on the cause – version change, prompt drift, poisoned context – before a bad reading ships inside a customer deliverable.

Price one unqualified reading. A supplier scorecard built on a drifted model. A PFMEA severity graded by an instrument with no reference standard. An audit response that quietly misstates a clause. Each surfaces months later as an escalation, by which point containment is a negotiation. Against that, the line items nobody budgets are almost embarrassing in their cheapness: a maintained evaluation set with known-good answers, regression suites that re-run critical prompts on every model version, drift monitors, and one named owner for the reaction plan. Call it €5,000 and a quarter of an engineer's year. One escalation traced to an unqualified reading costs multiples of that, plus a slice of credibility no budget line covers.

Purchasable governance is document control – it tells you who approved the output. Practised governance is measurement – it tells you whether the output was any good. The organisations that come through AI adoption intact will be the ones that already knew how to qualify an instrument, because they will look at the corporate copilot and see it plainly: the newest gauge on the floor, and the only one nobody has ever studied.