Why would Google pay millions of dollars for data belonging to an airline that no longer operates? The answer may have less to do with aviation than with the broader race to develop AI systems capable of performing many kinds of office work.
Google won a bankruptcy auction this month by offering $10 million for the corporate data of Spirit Airlines, which ceased operations in May. Its bid surpassed a $7.5 million offer from Mercor, an AI data company. The collection reportedly contains approximately 100 million emails, 500 million Microsoft Teams messages, and about 30 million lines of computer code.
At first glance, the purchase may seem like an unusual extension of the current AI competition. Over the past year, AI companies have focused heavily on building coding agents, and those systems have benefited from the vast amount of publicly available source code used for training. However, creating agents that can handle other forms of professional work may require a different kind of resource: large quantities of private, realistic records showing how employees communicate, make decisions, coordinate projects, and use business software.
If companies can successfully turn that information into training environments, the consequences could extend well beyond the technology industry and affect the labor market. “Buying Spirit Airlines—that’s not centrally a coding-agent purchase,” says Nick Heiner, head of reinforcement-learning environments at Surge, a data company that supplies leading AI laboratories. “That’s a purchase if you believe that your agent is going to generalize into the rest of the economy.”
Reinforcement learning
One of the most important developments in AI over roughly the past 18 months has been the growth of reinforcement learning from verifiable rewards. This approach trains models through repeated attempts: an AI agent performs a sequence of actions, receives a reward when it reaches a desired result, and becomes more likely to repeat strategies that work.
The environments used for this training are often simulations of familiar software systems. An agent might be asked to modify a codebase, update a database, complete a workflow, or manage information across several applications. Anthropic, for example, has discussed spending more than $1 billion annually on reinforcement-learning environments.
Coding has produced some of the clearest gains from this method because software provides an unusually precise feedback mechanism. A program either runs correctly or it does not, allowing the system to receive a relatively immediate and objective reward. Coding models also had a major advantage from the beginning: enormous amounts of source code were already available online, giving them extensive training material.
The longer-term ambition of AI companies, however, is much broader than programming. They want agents that can perform substantial portions of administrative, analytical, and other white-collar jobs. Spirit’s records could be valuable because they may show what those jobs look like in practice rather than in a carefully designed simulation.
Why the data could be useful
According to Heiner, reinforcement-learning environments are only as effective as the information used to populate them. Companies such as Surge and Mercor commonly employ people to build realistic scenarios, sometimes from scratch and sometimes with assistance from AI tools. Those workers may create examples of customer-service interactions, scheduling tasks, document workflows, or internal decision-making.
Yet artificially constructed examples remain somewhat removed from the complexity of actual business operations. “Even that is a little bit removed from literally having actual data that was used in the real world,” Heiner says. “That’s where deals like Spirit come in.”
Although Spirit ultimately failed as a business, its records may still preserve valuable patterns about how people work. The data could reveal how employees communicate across departments, divide responsibilities, resolve problems, pursue shared objectives, and follow the unwritten conventions of the airline industry. It might also show how workers interact with the software systems that support daily operations.
Even mistakes can provide useful training material. An AI developer could identify an inefficient decision or unsuccessful workflow, turn it into a task, and then create a better outcome for the agent to pursue. As Heiner explains, a poor example can be transformed into a lesson by showing the system how the process might have been handled more effectively.
The challenge is that office work is not as easy to evaluate as code. A software program can often be tested automatically, while a business email, staffing decision, or project plan may be reasonable in several different ways. Determining whether an action was truly effective may require a human expert to interpret context and exercise judgment.
Companies such as Mercor and Surge provide that human feedback to AI developers, but the process is expensive, slow, and inherently subjective. Consequently, it remains uncertain whether using real corporate records will allow AI companies to automate white-collar work as quickly as they have advanced in software development. Real-world data may make systems more capable, but it may also expose the difficulty of judging the quality of complex professional decisions.
Privacy concerns
Google has not explained precisely how it intends to use Spirit’s information. A company spokesperson said only that the data could help improve Google’s products and AI models. The spokesperson also said that an independent third party would rigorously remove personal information before the dataset was delivered and that customer data would not be included.
The proposed safeguards have not eliminated concerns. After Google placed its bid, a flight-attendants’ union objected to the transaction. The union acknowledged that the deal would remove personal information but argued that the dataset would retain “referential integrity”—the connections linking records across different types of data.
Those connections are important because they preserve the context that makes a dataset useful for training AI. For example, an email, a chat message, and a task record may remain associated with the same employee, department, or event even after names and other obvious identifiers are deleted. The union contends that preserving those relationships could make it possible to reconstruct information about individual members, despite the data being described as anonymized. It is asking for stronger privacy protections, and a judge is expected to rule on the issue on September 9.
The auction may not be over. Another AI data company, Micro1, submitted a $12.5 million offer after the auction had already closed, potentially complicating the outcome.
Regardless of which company ultimately obtains the dataset, the transaction signals that the AI industry is increasingly looking beyond publicly available code and toward the records of ordinary organizations. Google’s willingness to spend $10 million on the data of a failed budget airline suggests that the company may see Spirit not as an aviation investment, but as a detailed record of how a real business operated.
“If you believe this is really generalizable, you see a path to all these different industries—not just the ones that Spirit was involved in—being something you can replace pretty quickly,” Heiner says. In other words, the value of the data may lie not in teaching an AI system how to run an airline specifically, but in helping it learn the broader patterns of communication, coordination, and decision-making that appear across many workplaces.
Discover more from Headline Hub
Subscribe to get the latest posts sent to your email.
