Because former CEOs have a dangerous ability to turn advice into shadow authority.
“What do you think?”
“I asked first.”
“You are CEO.”
“I knew retirement would make you insufferable.”
“You promoted me to elder.”
“I did no such thing.”
Adrian laughed.
Naomi continued:
“I’m freezing automatic denial based solely on the score.”
“Manual review stays.”
“Good.”
“I’m notifying legal.”
“Good.”
“Independent analytics team.”
“Good.”
“Vendor gets notice after we preserve configuration and logs.”
“Good.”
“Board risk committee tonight.”
“Good.”
“Stop saying good.”
“Then stop doing sensible things.”
Naomi sighed.
“I hate you.”
“You’re welcome.”
Three weeks later, the cause emerged.
Not race as an input.
The system did not know race.
But one fraud feature relied heavily on address stability combined with credit-file depth.
That feature disproportionately penalized younger travelers from neighborhoods with higher rental turnover.
At several Bennett properties, those neighborhoods were disproportionately Black.
The system was not making an explicit racial classification.
It was producing a racial disparity anyway.
More importantly, the feature contributed little actual predictive value after controlling for payment verification and account history.
It had survived because nobody had tested fairness against utility carefully enough.
Naomi removed it.
The vendor resisted.
“You will increase fraud exposure.”
“By how much?” she asked.
They estimated.
The difference was tiny.
She removed it.
Then something more uncomfortable happened.
After the change, fraud did increase slightly.
Not dramatically.
But measurably.
A few executives said:
“See?”
Naomi refused the false binary.
The feature had indeed captured some risk.
It had also imposed unnecessary burdens unevenly.
The correct question was not whether it predicted anything.
Most broad proxies predict something.
The question was whether the gain justified the cost and whether a less burdensome method could achieve nearly the same protection.
They replaced it with transaction-based triggers.
Fraud returned close to baseline.
Manual review became less demographically skewed.
Nobody won a moral trophy.
The system improved.
Adrian liked that.
Naomi later asked him to attend the board debrief.
He refused.
“Why?”
“You handled it.”
“That’s why you should come.”
“No.”
“You need to stop turning me into institutional grandfather.”
“You are literally old enough.”
“Naomi.”
“Fine.”
The debrief proceeded without him.
That mattered.
An institution is not reformed if the reform depends on the original reformer remaining alive.
A year later, Naomi called again.
This time laughing.
“We have a problem.”
“What?”
“People are gaming the culture audits.”
Adrian smiled.
“Of course.”
“You sound happy.”
“I’m relieved.”
“Why?”
“Because now you’re in the normal phase.”
The Crown Review had become important enough that managers wanted good scores.
Predictably, some learned to optimize appearances.
Complaint resolution times improved suspiciously before audits.
Training completion reached one hundred percent while employee interviews suggested half the staff clicked through modules without attention.
One regional director had instructed managers to avoid coding guest disputes under bias-related categories unless the guest explicitly used the word discrimination.
Exactly the old problem wearing a new dashboard.
Naomi discovered it because an internal auditor interviewed employees instead of trusting the metrics.
The board was furious.
Adrian was less surprised.
“Any metric tied to prestige eventually becomes target.”
Naomi said:
“So what?”
“Make gaming harder.”
“How?”
“Random qualitative review.”
“Anonymous employee sampling.”
“Compare free-text complaints with categories.”
“Rotate metrics.”