The agent is live. The yield isn't.

Two numbers did the rounds this week: 79% of enterprises report AI agents running in production, and 15% can tie any of that investment to the P&L. I have seen this film. I have walked plant floors where SPC charts hung beside the line, corners curling, and not one supervisor on shift could tell me what the chart had said last Tuesday. The auditors admired them. Deployment without measurement is not adoption; it is decoration – and a 64-point gap between what companies run and what they can price is not an IT-maturity problem. It is the oldest problem in quality: the gap between plants that pinned charts to the wall and plants that read them.

An agent is a process, so it gets a spec

Strip the vendor language away and an AI agent is a process. It takes inputs, transforms them, emits an output, consumes cost and can fail. I have run PFMEA workshops over stamping, welding and assembly; the failure logic does not change because the transforming machine is a language model. The failure modes change. Severity, occurrence and detection do not care what the machine is made of.

Before any agent ships, it needs an output specification: what a conforming output looks like, what tolerances apply, what is rejectable, and who signs conformity. If nobody in the building can write that spec, you have not deployed a process. You have deployed a demo.

The past fortnight of announcements says the market is solving a different problem. Trusted runtimes for securing agents in production. A second generation of autonomous enterprise agents. A governed rollout at an insurer, a pre-seed round promising agents on factory floors. All guardrails and horsepower. I hold a Certified Ethical Hacker credential among thirty-odd security certifications, so I take runtime security seriously. But securing the agent and rating its output are separate disciplines, and a perfectly sandboxed agent producing wrong answers at industrial speed is a contained defect generator.

First-pass yield works on machine work too

First-pass yield is the bluntest metric on any floor: of a hundred units started, how many pass every specification first time, no rework, no sorting. It transfers to agent output unaltered. Take a thousand outputs – quotes drafted, orders routed, supplier emails answered. Count how many left the agent fit for use, how many needed human correction before release, how many escaped unseen. That is your FPY, and it is the first adoption number a board should ask for.

Escapes deserve line discipline. QRQC taught me the fifteen-minute rule: from detection to containment you stop, isolate, get to the point of failure, find the cause under real conditions, verify the countermeasure there – not in a meeting room. An agent escape gets the same treatment. Which prompt, which context window, which model revision produced the defect; systematic or sporadic. Then decide what containment stops it recurring tomorrow morning. The method was built for presses and weld cells. It does not care what failed, only that something did.

When I designed MultiPS – the platform that runs sixty-three language models in parallel and synthesises consensus – the non-negotiable was never model count or interface polish. It was traceable, rateable outputs. Every answer carries its lineage, and disagreements between models are logged, not averaged away. Consensus is a detection control, and a detection control earns its keep only when what it catches is recorded and priced.

An output you cannot rate is a cost you cannot see – and a cost you cannot see compounds daily.

Put the agents on the same ledger as the scrap

At SNOP I built the cost-of-poor-quality ledger for a greenfield plant of more than nine hundred people, from a blank page. Nobody hands you the chart of accounts; you build the categories yourself – scrap, rework minutes, sorting, expedited freight, customer claims – and you make finance co-sign each one so the numbers survive contact with a controller. That ledger is what turned a 70% reduction in defect costs into a fact rather than a claim – visible because the baseline had been priced first.

The same categories price agent output. Cost per escaped error. Rework minutes per thousand outputs. The cost of the human review layer that exists only because the yield is unknown. An escaped email quoting a tolerance that does not exist is an 8D and a claim in a plant; in most offices it is not even logged. Price the categories, trend them weekly, and the arithmetic turns unsentimental: an agent that saves four analyst-hours and generates six hours of downstream correction is a net liability presenting as a productivity win. None of this touches the agent estate today; no clause in IATF 16949 obliges anyone to look. Absence of a clause has never meant absence of a defect.

This week's automotive news shows both ends of the discipline. BYD spent it publicly apologising for a wheel-rim error while Chinese regulators tighten the screws on vehicle quality; Sona Comstar's driveline division collected a Deming Prize for total quality management. One escape, named and costed, against seventy years of TQM machinery doing its job. The tools exist. They are simply not pointed at the agents.

The sixty-four-point gap will not be closed by better agents. Plants did not improve because someone pinned a chart to a wall; they improved when a named person became accountable for reading it and acting on it. Agents will get there eventually – the hard way, through an escaped defect with a price on it, or the structured way, through a spec, a yield and a ledger. When an agent's first-pass yield sits on the same page as the press's, reviewed in the same meeting, the 15% stops being a survey finding and becomes a floor. We have run this project before. The substrate was metal. The discipline is identical.