The Tower of Invisible Decisions

The Tower of Invisible Decisions

The rain had been falling since before dawn, the kind of cold, deliberate October rain that turns city sidewalks into mirrors and makes umbrellas feel like acts of optimism. By the time Mira Sandoval turned off Lake Shore Drive onto the service road behind Northwind Bank's glass headquarters, her shoes were soaked, her phone had vibrated seventeen times with calendar reminders she did not need, and the leather folder on the passenger seat — the one containing her signed offer letter, her ID badge photo that she refused to look at, and a single printed page of the bank's "AI Governance Charter, v0.3" — had slipped twice onto the floor.

She was not, by nature, a person who arrived early to things. But the email from the chief risk officer had been unusually firm: "Please be in the lobby at 8:30. We will walk you in personally. Do not use your badge until then." The phrasing had felt almost ceremonial, like being met at a temple gate.

The Northwind tower rose out of the rain like a slab of polished obsidian. Forty-two stories, LEED Platinum, designed by a firm that had won awards for "honest materials." Mira had read about the building twice. The first time, four months ago, when she had been asked to interview for a job she did not entirely understand. The second time, two weeks ago, after she had accepted and begun Googling phrases like "model risk management" and "what does an AI governor do" at two in the morning.

She still did not have a clean answer to the second question. The job title on her offer letter was *Head of AI Governance, Northwind Bank, Senior Vice President.* The title on the org chart she had been sent was *Head of AI Governance, Northwind Bank, Senior Vice President.* The title in the LinkedIn message from the recruiter, a woman named Priya who typed in complete sentences with actual punctuation, had been *Head of AI Governance, Northwind Bank, Senior Vice President.* And yet none of these three identical strings of words had prepared her for what the chief risk officer, a tired-eyed man named Henrik Voss, said to her in the back seat of a town car four minutes later, after a security guard had silently waved them through a side entrance.

"Mira," he said, not looking at her, "I need you to understand something before we go upstairs. The trading floor you passed on the way in? That is not the bank. That is the museum. The bank is on the seventh floor. It has no windows. It does not exist on any diagram we give to regulators. And as of yesterday afternoon, it was making roughly twenty-three million decisions a day without a single human reading the logs."

He paused. The car was idling under a porte-cochère that smelled faintly of wet concrete and money.

"That number," he added, "is going to scare you. Good. You should be scared. The previous person who sat in your chair was scared too. He quit in March."

"Why?" Mira asked.

"Because he figured out what he was actually sitting on top of," Henrik said. "And no one would let him do anything about it."

---

The Scale and Invisibility of AI Decisions Inside Large Enterprises

The seventh floor of Northwind Bank did not look like a bank. It looked like a mid-sized hospital data center crossed with a graduate-school library that had been allowed to age badly. The carpet was industrial gray, the ceiling was a grid of acoustic tiles interrupted by bundles of cable trays that had been added so many times over the years that they resembled mechanical vines. There were no trading screens. There were no customer service representatives. There were no private bankers in bespoke suits. There were, instead, several hundred people working at long shared benches under flat panel lights, each one staring at code, dashboards, or spreadsheets that Mira could not yet interpret.

A woman named Adaeze Okonkwo met them at the elevator. She was the head of "model operations," a title Mira would later realize meant "the person who actually keeps the lights on." Adaeze had been at Northwind for eleven years. She was the kind of engineer who wore the same black Northwind hoodie three days in a row and could tell you, from memory, the difference between a logistic regression and a gradient boosted tree while walking and drinking cold coffee at the same time.

"Mira, welcome," Adaeze said, shaking her hand with a warmth that felt almost defiant in such a sterile space. "I'm going to show you something. Please don't say anything until we're done."

She led Mira down a corridor, past a row of small offices with the doors closed and the lights off, and into a windowless room that contained exactly one thing: a wall.

The wall was a wall in the way the Great Wall is a wall. It stretched almost fifteen feet across and ten feet tall, and it was covered, edge to edge, in printed pages — flowcharts, dependency diagrams, red-team memos, quarterly attestations, deployment tickets, vendor invoices, screenshots of dashboards, screenshots of errors, screenshots of screenshots of errors. At the top, in large block letters that had been printed and taped rather than mounted, was a phrase: Northwind Production Model Inventory, Q3.

Adaeze handed Mira a pair of reading glasses she had not asked for.

"Count them," Adaeze said.

Mira counted.

It took her eleven minutes. When she was done, she had written a number on the back of her visitor badge. 3,012.

"Three thousand and twelve production AI and machine learning models," Adaeze said, as if she had said this sentence many times before and still did not entirely believe it. "As of this morning. Eighteen hundred of them make decisions about customers directly. The rest make decisions about employees, vendors, the building, the cloud bill, the ATMs, the marketing spend, and each other."

"How many decisions a day?" Mira asked.

"Yesterday: twenty-three point one million. The day before: twenty-one point seven. The day before that: twenty-two point four. We have a dashboard." Adaeze nodded toward a small monitor mounted near the door. The dashboard showed a number ticking up in real time. 18,442,103. Underneath, in smaller type, were six other counters for fraud, credit, marketing, HR, treasury, and "other." The "other" counter was the largest by an order of magnitude, and Adaeze did not explain why.

Mira stared at the wall. She was a person who had spent her career in regulation. She had written compliance memos. She had reviewed audit reports. She had argued with bank examiners in conference rooms with very good coffee. She had never, in her entire professional life, been inside a place where the actual mechanisms of decision-making were visible at the same time as the actual decisions. That, she would realize over the coming weeks, was the point of this room. Most banks did not have such a room. Most banks did not even know such a room was necessary.

"Who built this wall?" she asked.

"I did," Adaeze said. "Last month. Took me two weeks. Took three of my engineers two weeks to pull the inventory. The number we got back from IT was two thousand four hundred. The number we got back from the business was nine hundred. The real number is three thousand and twelve. That gap, between what IT thinks exists and what actually exists, is what your job is going to be about."

Henrik had been silent the entire time. Now he spoke.

"In 1995," he said, "Northwind Bank made perhaps thirty thousand consequential decisions a day. Loans, trades, wire transfers, account openings. Each one was made by a person whose name appeared on a piece of paper. We knew who decided what, why, and on the basis of which rule. The audit trail was the paper. The bank was, in a real sense, legible to itself."

He gestured at the wall.

"In 2025, this bank makes more than twenty million consequential decisions a day. Most of them are made by code. The code is written by people whose names appear in a Git commit history. The decisions are reviewed by other