Why Is Google Bidding $10 Million for Spirit Airlines' Emails?
Google's $10M bid for Spirit Airlines' old emails and Teams data shows how AI labs are buying corporate archives to train agents on knowledge work.
What is Google actually bidding on?
In August, Google placed a $10 million winning bid for Spirit Airlines’ internal records during the bankrupt carrier’s asset auction, beating a $7.5 million offer from a company called Merkor. The haul isn’t planes or gates. It’s an inventory reported to include around 100 million emails and roughly half a billion Teams messages, plus code and business documents. A sale hearing is scheduled for October 14, so the deal isn’t final, but the bid itself tells you something new about what AI labs think is valuable: not Spirit’s business, but the digital residue of everyone who worked there.
Why would a company buy failed Spirit Airlines emails instead of generating cleaner data?
Because labs are chasing a specific, unproven theory: that if you feed an AI system enough raw records of how real knowledge work actually happened, Slack and Teams threads, email chains, meeting follow-ups, documents in various draft stages, the system can learn to do that work itself. Spirit’s archive is attractive precisely because it’s messy and real. It’s years of actual humans doing actual jobs, including the parts that never show up in a tidy case study: chasing down a missing approval, correcting a supplier’s invoice, scheduling a follow-up meeting because the first one didn’t resolve anything.
Built like a system. Not vibe-coded.
Remy manages the project — every layer architected, not stitched together at the last second.
This matters because most of the AI progress people talk about so far comes from verifiable domains. Code either runs or it doesn’t. A chess move is legal or illegal. Knowledge work doesn’t offer that kind of clean feedback loop. Nobody can point to the exact sentence in a report that made it good, or the specific email that made a negotiation succeed. Training an agent to do that kind of work requires some proxy for success, and corporate archives are being treated as that proxy.
How are these deals structured, and who profits?
The emerging pattern has three parties. The distressed company (or its bankruptcy estate) sells the data and gets paid. A broker arranges or facilitates the sale and takes a fee. The buyer, in this case Google, hopes to turn the archive into agentic capability it can eventually resell as a product. The employees whose daily work generated every one of those emails and messages get nothing and generally have no say.
A newer player, a data broker called GP.AI, is already formalizing this pipeline for smaller companies, offering to package business records (messages, code, documents) for sale at prices reportedly ranging from $20,000 to $5 million, then reselling bundles to AI labs. The pattern across these deals is consistent: a third party packages the record of your work and monetizes it, usually without your direct input.
Merkor, the losing bidder on Spirit, is worth watching too. It also announced plans in July to acquire Deep Tune, a company that builds “training environments,” meaning the simulated software settings, assignments, and success checks used to train agents. That pairing matters: Merkor isn’t just after records of work, it’s building the infrastructure to decide what counts as a completed task. It has reportedly advertised paid roles for people who help define those success criteria, which means human judgment calls about “was this job done correctly” are quietly becoming training labels that shape what an agent learns to optimize for.
Can an email archive actually teach an agent to do knowledge work?
This is the real open question, and there’s good reason for skepticism. Consider a realistic scenario: an invoice doesn’t match a purchase order. The invoice lives in one system, the purchase order in another, a supplier emails about a credit but the right reviewer isn’t copied, and there’s a dispute over whether a partial delivery counts as complete. Resolving that might take 30 messages and three meetings. The actual value-creating moment might be one person noticing a credit was misapplied. Everything else, status updates, “just following up” messages, scheduling threads, is coordination overhead, not the work itself.
An agent trained on the full thread has no inherent way to distinguish the message that solved the problem from the 29 that didn’t. Worse, a thread can end with “invoice approved, ticket closed” and look like a success case, even if a duplicate payment surfaces six weeks later. Archives capture the performance of work (status updates, visibility, follow-ups) sitting right next to the actual substance of work, with no label distinguishing the two. If a buyer treats an entire sequence as an example of “competent accounting,” the agent risks learning to imitate the appearance of resolution rather than the judgment that produces it.
Why does it matter that the data comes from a bankrupt company?
There’s a structural bias worth naming: the companies most willing to sell their internal archives are often the ones failing. Spirit Airlines went bankrupt. The sellers in this emerging market are disproportionately businesses that went to the wall and have no better use for the data than a one-time payout. That means the record being used to teach agents “how knowledge work gets done” skews toward organizations where, by definition, the work didn’t ultimately sustain the business. A thriving company with proprietary insight into its market has a stronger incentive to build on its own data than to sell it off cheaply. The sellers in this market are, in a real sense, data of last resort.
Are workers pushing back?
Yes, in both public and private ways. In the Spirit case, the flight attendants’ union raised concerns about confidential employee information embedded in the records and challenged whether excluding certain systems actually protected sensitive data that might have been copied elsewhere, since real-world work is messy and information often ends up in the wrong place. That’s a concrete, visible form of resistance.
Less visible resistance is also happening inside companies rolling out AI tools. Some leaders privately estimate that a meaningful share of their teams, as much as a third in some cases, are quietly slowing down or sabotaging AI tool adoption. This often stems from employees sensing, correctly or not, that a tool framed as “doing your job for you” threatens how their contribution is valued, not just whether a task gets automated.
TL;DR
- Google’s $10 million bid for Spirit Airlines’ bankruptcy-sale records (reportedly around 100 million emails and roughly half a billion Teams messages) beat a $7.5 million offer from Merkor, with a sale hearing set for October 14.
- AI labs are betting that raw corporate archives, emails, chats, documents, can teach agents to do knowledge work the way code repositories taught models to write software, though this remains an unproven theory.
- Archives mix the performance of work with the actual work, so an agent trained on a full email thread can’t easily tell the status update that signaled progress from the single message that actually solved the problem.
- A broker economy is forming fast, including companies like GP.AI packaging small-business data for resale, and Merkor pairing data acquisition with its own Deep Tune training-environment business to define what “task success” even means.
- Failing or bankrupt companies are the likeliest sellers, which means the training data shaping future knowledge-work agents skews toward organizations whose approach to work didn’t ultimately keep the business alive.
- Employees generally have no stake in these sales, though unions and individual workers are starting to contest data terms and quietly resist AI rollouts framed as full job replacement.
Frequently Asked Questions
What exactly did Google bid on in the Spirit Airlines auction?
Google placed a winning $10 million bid for a portion of Spirit Airlines’ internal digital records during the bankrupt airline’s asset sale, reportedly including around 100 million emails, roughly half a billion Teams messages, plus code and business documents. The sale is subject to a bankruptcy court hearing scheduled for October 14.
Why do AI labs want old corporate emails and chat logs?
Labs are testing the theory that records of real workplace communication, emails, chats, documents, can be used to train AI agents to perform knowledge work tasks like managing invoices, coordinating approvals, or resolving disputes, the same way code repositories helped train AI coding assistants.
Is buying bankrupt companies’ data a reliable way to train good AI agents?
Plans first. Then code.
Remy writes the spec, manages the build, and ships the app.
It’s unproven and contested. Archives mix meaningful problem-solving with routine status updates and coordination overhead, with no clear label distinguishing the two. There’s also a selection bias: most sellers are distressed or bankrupt companies, so the data may not represent how successful organizations actually operate.
What is Merkor and why does it matter in this market?
Merkor is a company that bid on the Spirit Airlines data and separately announced plans to acquire Deep Tune, a firm that builds training environments (simulated tasks and success checks) for AI agents. That combination positions Merkor to both supply raw work records and define what counts as a successfully completed task, a judgment call that shapes what agents actually learn.
Can employees or unions stop their work data from being sold?
They can raise objections during bankruptcy or sale proceedings, as the Spirit Airlines flight attendants’ union did regarding confidential employee information. Outside formal proceedings, workers have limited legal power over employer-owned communication systems, though some respond with quiet resistance to AI tools they see as devaluing their work.