The Corporate Connectome
Training AI on chaotic chat logs instead of pristine wikis teaches agents the unwritten human workarounds needed to improvise when formal systems fail.
By Theo Marsh
Sparked by Google has acquired the data of failed US airline Spirit · discussion
The Register reported yesterday that Google scooped up bankrupt Spirit Airlines' internal corporate data at auction. Over on Hacker News, the immediate consensus was to ask what happens when you train a neural network specifically to lose luggage.
I get it. The privacy implications of buying a defunct company's Slack archives are deeply uncomfortable. And the jokes practically write themselves.
But those jokes rest on a deeply flawed mental model of what a modern enterprise model actually needs to learn. We assume you teach an AI to work by feeding it pristine documentation. You scrape the corporate wiki, embed the employee handbook, and ingest every procedural guideline.
Except standard operating procedures are a hallucination.
They represent how a bureaucracy wishes it operated on a sunny day with infinite resources. The formal wiki assumes every API returns a 200 OK and every employee is perfectly aligned. If you genuinely want to build an autonomous agentic workforce, those happy-path wikis offer almost no signal. You're training on the ideal, not the reality. A bankrupt airline's chaotic, desperate, stressed internal chat history provides arguably the most valuable AI training set on earth.
To understand why, you have to look at how physicists study the fundamental building blocks of the universe. You don't just stare at a stable atom. A stable atom is boring. It hides its secrets. Instead, you hurl particles at each other at near-light speed inside a massive magnetic ring and measure the shrapnel. You look at the decay trails. As Ted Benson pointed out years ago while discussing data-driven computer science, you have to smash a complex system open to see its true components.
Bureaucracies follow the exact same physics.
Imagine a side-by-side contrast. On the left, you have a clean, hierarchical Org Chart. It looks perfectly symmetrical, utterly rigid, and completely divorced from reality. On the right, you have the actual communication flow -- a chaotic web of direct messages bypassing the hierarchy entirely to fix a broken process.
The formal chart claims the Shift Manager talks to the Regional Director when a plane is delayed. But the collision data proves something else entirely. It proves the Shift Manager actually ignores the Director and immediately direct-messages a specific Level 2 database engineer named Dave in Slack.
Because Dave is the only person who knows how to manually reset the crew-scheduling lock.
And Dave doesn't use the official ticketing portal. He uses a hacked-together Python script on his local machine.
I call this invisible graph The Corporate Connectome.
This connectome is the latent web of relationships, favors, and unapproved workarounds defining how a company functions when the primary systems fail. Crucially, it only becomes visible during a crisis.
A canceled flight during a blizzard, a missing flight crew in Newark, a broken third-party payment gateway. These are the high-energy collisions of the enterprise world. When those events happen, the polite Org Chart instantly dissolves, replaced by a flurry of raw problem-solving. Far from being useless noise, those leaked chat logs serve as perfectly labeled, human-generated traces of how to resolve catastrophic error states.
Right now, the missing layer of enterprise AI is recovery intuition. We build agents equipped to execute a rigid task, provided the environment behaves perfectly. We give them a neat little JSON manifest of tools to query a flights database or issue a refund via Stripe. Yet the second an API returns an unexpected 503 Service Unavailable, or a database lock prevents a row update, the agent halts. It throws an exception. It gives up and kicks back a polite "I'm sorry, I couldn't complete that request."
Agents fail because we train them on the job description. We teach them the rules while completely ignoring the unwritten exceptions to the rules.
Consider what a human support agent actually does when a passenger is stranded at 2:00 AM because a partner airline canceled a connecting flight.
The wiki says to rebook them on the next available partner flight. It's a simple if/then flowchart. But the chat logs reveal the messy reality. The next partner flight isn't for 14 hours. The passenger is increasingly frantic because they have a medical emergency. To make matters worse, the inventory system is throwing a silent error because it incorrectly thinks the original flight is already boarded.
So the human agent improvises. They bend a policy to issue a direct competitor's ticket. They simultaneously DM a supervisor on Slack to retroactively approve the cost overage. They manually override the inventory lock by dropping into an ancient AS/400 terminal emulator that technically shouldn't even be accessible.
The Spirit Airlines archive provides millions of these exact timestamped examples. It captures humans encountering unexpected failure modes, diagnosing the root cause across different software silos, and coordinating a messy, imperfect fix. It is a massive, multi-modal dataset of error correction in the wild.
Mining a collapsed bureaucracy provides Google with something far more potent than a discount travel bot. They are extracting the exact sequence of corporate actions required to navigate extreme operational complexity. They are mapping the connectome so they can teach software how to panic, improvise, and recover exactly like an exhausted operations manager.
My guess? In five years, Google's real product launch will just be a foundation model that can perfectly simulate the back-office of a 5,000-person company.. bugs, workarounds, and Dave included.