In 1993 I was sent to run big-ticket delivery operations for Sears Canada. 35% of customers were telling us their delivery wasn’t perfect. Most of the problem traced back to one issue: the merchandise simply wasn’t there on the morning of the promised delivery.
Previous management’s solution? Hire a night shift with clipboards to walk the staging area, count the defects, and write them down.
That drove me crazy. So I went for a walk in a million-square-foot warehouse I didn’t manage.
What I found was obvious in hindsight: trains full of merchandise parked against the wall after the day shift had gone home. The root cause was sitting there untouched until after the delivery trucks left.
We moved the trains. A year later delivery costs were down $2 million annually and customer satisfaction hit 98% perfect — not because the inspectors got better, but because the process improved.
The lesson has a name. Philip Crosby taught it decades ago: Quality is free, but only if you build it in. Everything else is the price of nonconformance — prevention (which pays) versus appraisal (which just counts).
That same night shift is back in 2026 — and it’s enormous.
We now have output guardrails, hallucination checkers, citation validators, human review farms, and red team evaluations. All downstream inspection on processes that still produce defects. Consultancy reports with fabricated citations have already made the headlines.
Prompt engineers are the new salespeople — hand-carrying workarounds because the core process fails. Bloated retrieval systems dump dozens of documents into context windows because precision is poor. And sycophancy? That’s not a bug — it’s the reward signal working as designed.
The big difference today? The price of poor quality arrives itemized on your API bill. Every retry, every re-run, every “great question.” Almost nobody reads their usage logs or invoices as quality documents.
I recently walked my own 22-million-record usage logs and found changes I never authorized — including a default switch that affected model routing and cost. The trains are against the wall again, in buildings I don’t control.
So what does a quality person do when they can’t move the trains?
I write about it. I build instruments (TokenScope) so the cost of nonconformance becomes visible. And most importantly, I build prevention.
That’s Michelangelo — patent-pending governance that sits upstream of execution. In a pharmaceutical patent intelligence pipeline, it took fabrications from 74 to zero. Not by adding more clipboards, but by changing the process so the inspection has nothing to find.
This is the finish line for AI technology.
Some percentage of your AI output is generating angry calls and liability somewhere.
Who walks your building?