A bankruptcy judge has delayed approval of a $10 million deal that would hand Google the internal business records of Spirit Airlines, after the carrier's flight attendants objected to the sale and asked for stronger protections covering employee data.
The agreement, reached recently, would give Google access to a wide slice of how the bankrupt airline actually runs: emails, Teams messages, software, spreadsheets and operating records. Google plans to use the material for product development and AI.
The dispute is now in front of the court. The judge put the sale on hold while the two sides sort out what happens to information about the people who worked at Spirit.
A Preview of a Fight Over Rights
The standoff is a preview of what happens when data rights questions wait until after somebody decides the information is valuable, according to Ben Lorica, who wrote about the deal in a Gradient Flow article posted on September 15, 2026.
Lorica's broader argument is that the record of how people actually do their work may be more valuable than companies have realized. Final documents can tell an AI what a company knows, he writes. Workflow history can teach it how the company operates.
That distinction matters as agents take on more tasks. Lorica writes that agents may make the process behind outputs just as important as the outputs themselves. A finished financial analysis, a resolved support ticket, shipped code or a customer recommendation shows the result. It does not show what information the employee sought, what they ignored, which tools they used, where they got stuck, what they corrected and why they decided the job was done.
"I think there are three useful layers here," Lorica writes.
The bottom layer is raw activity: messages, meetings, tickets, commits, clicks and logs. The middle layer is trajectories, which connect those fragments into sequences that run from goal to decision to action to outcome. The top layer is what can be built from trajectories, including examples, failure cases, evaluations, corrections and reusable workflows.
Not all of that material is worth keeping. "Most workplace exhaust is probably just exhaust," Lorica writes.
Where Experience Becomes a Workflow
Lorica points to a recent system that looks across the messy traces left by agents, finds recurring procedures and turns some of them into executable workflows. In one example, a task involving 34 API calls was reduced to 11 once a recurring procedure was compiled.
The savings are not automatic. Compilation involves extra work, Lorica notes, so not every workflow becomes cheaper.
The same logic applies to human work. Researchers have used software-development event logs to infer roles and generate agent specifications, according to the article. Companies are also building training environments where agents practice complete tasks and are scored on accomplishment.
Lorica compares it to how people get better at a job. They accumulate experience. They do not simply read finished work.
The raw archive is not the asset, he argues. The asset is the ability to turn operating experience into something a machine can learn from and reuse.
Stay ahead of the AI curve
The most important updates, news, and content — delivered weekly.
No spam. Unsubscribe anytime.
That is where many companies may be falling short. Lorica describes a preservation problem: decisions get separated from outcomes, exceptions disappear into chat, and fixes are applied without any recorded reason. Companies may keep plenty of activity and still lose much of the experience.
The worry for many firms is that valuable workflow data will leak. Lorica writes that companies may worry about valuable workflow data leaking when the more immediate issue is that they never captured the valuable parts.
Two Competitors, One Learning System
He imagines two competitors working with the same model and the same documents. One of them also holds hundreds or thousands of examples of how experienced employees work, connected to outcomes and evaluations. That company, he writes, builds a learning system around its own operating experience.
The choice of model still matters, but the terms have shifted. Companies increasingly have credible choices at the model layer, even if open models are not always the cheapest or best option for a particular task, according to the article. The question is not just which model works best. It is where a company is comfortable having its know-how processed and stored.
Lorica's advice for anyone starting this work is to pick the right target. Begin by identifying work that is valuable, frequent, requires expertise and produces evaluable outcomes. He offers a simple test: whether you would invest meaningful time teaching a new employee to do the task well.
He also warns against overreach. Do not start recording everyone all day. The useful material is decisions, exceptions, failed approaches, corrections and feedback.
Rights questions should be settled early, he writes. That means deciding how the company can use data for AI, what employee and customer information is mixed in, what can be transferred after an acquisition, what an outside model provider may retain, and which workflows need tighter control.
The Spirit case shows what happens when those questions are left open. Google agreed to pay $10 million for the airline's internal business data. Spirit is bankrupt. The flight attendants objected and sought additional protections for employee data. The judge delayed approval while the dispute is sorted out.
The goal, Lorica writes, is not a bigger archive but a better record of how the work succeeds. He argues that durable value is moving into what a company learns by operating its systems as access to capable models gets easier. "This is starting to move beyond theory," he writes.
Agents may need something similar, he adds.
The article also carries a direct request for reader support. "No paywall tricks, just an honest ask: consider becoming a paid supporter," it states.
Lorica's article points readers to related Gradient Flow content, including "Water, Noise,Power: The Real Costs of Data Centers" and "Navier-Stokes and AI generated work: If You Can't Explain It, Did You Really Do It?"
For Spirit's flight attendants, the immediate question is narrower than the future of AI training data. It is what happens to their emails, their messages and the records of their work now that a buyer has put $10 million on the table.