After the lecture, a bank CEO approached.
Young.
Early forties.
He said:
“I was planning to do an undercover branch tour.”
The granddaughter smiled.
“Still do it.”
He looked surprised.
“After everything you just said?”
“Yes.”
“Observation is valuable.”
“Just don’t come back announcing you now understand what it means to be powerless.”
He laughed.
“Fair.”
She pointed her cane.
“And don’t fire anyone in a lobby.”
He smiled.
“I’ve seen the video.”
“Everyone has.”
The granddaughter’s last major project involved something her grandfather would never have imagined.
Synthetic customer agents.
Banks had begun deploying AI systems capable of handling increasingly complex service interactions.
Not simple chatbots.
Systems that could explain products, identify documents, schedule appointments, triage disputes, even negotiate fee waivers within policy boundaries.
Advocates said automation could reduce human bias.
Critics said it could industrialize bias.
Both were right depending on design.
The granddaughter joined an ethics review at a large cooperative bank.
The system performed well overall.
It was patient.
Never irritated by accents.
Never judged clothes.
Never got tired at five fifteen on Friday.
But analysts found a subtle pattern.
When customers used uncertain language—
“I think…”
“Maybe…”
“I’m not sure…”
—the system was more likely to route them toward generic educational materials before offering higher-level support.
Customers who spoke confidently—
“I need…”
“I qualify for…”
“I want an appointment…”
—were routed faster to specialists.
Was that bias?
Not intentionally.
The model had learned from historical service data.
Confident customers historically pushed past first-line barriers more often.
The AI interpreted confidence as readiness.
The granddaughter stared at the results.
“So now we have automated deference to assertiveness.”
The lead engineer nodded.
“Basically.”
“Who gets hurt?”
“People who are unsure.”
“First-time borrowers.”
“Some elderly customers.”
“People using a second language.”
“Anyone socialized not to sound demanding.”
The granddaughter thought of her grandfather.
A calm old man saying:
“I’d prefer a teller.”
What if he had been less confident?
What if he had accepted the ATM suggestion because he feared appearing difficult?
The system was changed.
Instead of using confidence cues as readiness signals, it asked explicit need questions.
Would you like general information, or would you like to speak with a specialist?
Choice.
Simple.
The engineer said:
“It slows the conversation.”
The granddaughter nodded.