The trust is measured. The escape rate isn't.

If a torque specification had escaped at 33% on any line I have ever run, the sequence would have written itself. Line stopped within the shift. Suspect stock cordoned. Containment parts on a truck to the customer, an 8D opened before the day ended, a customer auditor booking flights. Nobody would have proposed a dashboard. This week, Optro research reported that one organisation in three has acted on a wrong decision made by an AI agent. Read that with quality eyes: a 33% escape rate on a characteristic that now drafts customer communication, approves spend and pre-sorts the work humans pick up every morning. The market's answer arrived in the same news cycle – Collibra bought Trail to automate governance from policy to production, SAS shipped its AI Navigator, OneTrust rolled out an enterprise-scale overhaul. Three platforms, zero containment. Nobody read the number as a quality number. That is the actual failure.

A decision is a product characteristic

An agent's decision is not a mysterious artefact. It is a product characteristic with a spec, like any other. The inputs it is permitted to see, the authority limits it operates within, the tolerance – the conditions under which it must stop and escalate – together form that spec. It has failure modes: stale data in, confident inference from partial context, the right answer to the wrong question. PFMEA logic applies to an agent approval chain the same way it applies to a torque spec. Severity: what a wrong approval costs. Occurrence: how often the model errs in this domain. Detection: the odds the error is caught before execution. Multiply, rank, work the top of the list.

When a wrong decision is acted on, it has done what a defect does: passed every detection zone the process owned and reached its customer. The customer of an agent's decision is the human who executes it. By the book, that is an escape. Undetected, shipped, consumed.

A wrong decision someone acts on is not a governance finding. It is an escape – and escapes are counted in euros, not in slide decks.

What the platforms count – and what they don't

Policy coverage, model inventory, bias audits – these are certificate-era metrics. They tell you the paperwork exists. They are the modern equivalent of the shelfware manuals that sailed through ISO 9001:1994 audits while the line shipped scrap: certificate on the wall, scrap on the dock. Policy coverage measures documentation, the one thing documentation reliably produces. One of this week's launches was pitched explicitly at strengthening customer trust in AI.

In a plant, trust is a measured characteristic, earned at the gauges. It is not declared in a press release.

What none of the launches counts is escape accounting: detection rate on wrong decisions, cost per escaped decision in €, reaction time from wrong output to containment, repeat-escape rate. Those are the numbers I spent two decades demanding before signing off a process. They describe the product, not the promise.

I have watched this discipline work without a single governance licence. At WITTE Automotive we ran QRQC, A3 and a Q-Wall: escape rates on a board, counted daily, priced in €. Failure cost came down substantially. At SNOP, building the quality operation for a 900-strong greenfield plant, we took 70% out of defect cost and held 98% customer satisfaction the same way – the number on the wall came first, the tools followed. Nobody needed a platform to know the escape rate. It was the first thing you saw walking onto the floor.

Running QRQC on an agent

QRQC translates to agents almost unchanged. Contain, diagnose to root cause, countermeasure, verify. Concretely:

• Decision specs before go-live. For every high-consequence call the agent can make, write the spec: permitted inputs, what it may decide alone, what requires a human, the conditions that force escalation. A control plan, not a policy document.
• A designed detection zone. I designed and built MultiPS to run 63+ models in parallel with consensus synthesis because I treat a single model's output as an ungated characteristic. When the models disagree, that disagreement is the inspection signal – a go/no-go gauge for text. Split verdicts route to the gate; unanimous ones flow.
• Human gate on the vital few, not the trivial many. Pareto applies. Gate the two or three decision types whose escape costs four or five figures; let the rest run, or the ROI dies under inspection cost.
• Log every acted-on decision. Who executed it, on which inputs, at which authority level. Traceability is not surveillance; it is the control-plan record you will pray for during the first 8D.
• Weekly escape review, 8D on any repeat. A second escape of the same failure mode means the containment failed, not the model. That is a systemic finding.

Quality management, security engineering and AI engineering are the same discipline: understanding how complex systems fail and building them so they don't. The Optro number says the office side has not yet learned what the shop floor absorbed two decades ago.

The board question

The question for next quarter's board meeting is not whether we own an AI governance platform. It is this: what is our escape rate on agent decisions, and what does one escape cost us? If nobody in the room can answer, the organisation is certifying a characteristic it has never measured. Quality engineering solved this decades ago – PFMEA, control plans, QRQC, 8D – and the tooling sits on the shelf, waiting to be carried from the factory floor to the office. Acting on unverified output used to have a shorter name. We called it an escape.