Google has bid $10 million in bankruptcy court for Spirit Airlines’ internal records: roughly 100 million emails, 500 million Teams messages, plus spreadsheets and documents written by employees doing their jobs. The first wave of large language models learned from the public web, so they can explain a business without knowing how one actually runs. That knowledge lives in reply-all threads and Teams arguments about who should call the customer, and it is what developers building AI agents now want. More than 120 US lawmakers have raised employee privacy concerns, and Google says an independent third party will strip out identifying information first, which is harder than it sounds in conversations full of specific incidents and relationships. The part that should worry anyone running enterprise technology is the fine print on vendor promises. Microsoft and Slack both commit that customer data will not train shared foundation models, and Spirit shows that protection lasts as long as the company does. Once it fails, the archive becomes an asset creditors can sell.
Google has offered $10 million for Spirit Airlines’ internal data. Not the aircraft or the airport assets. The emails and Teams chats the airline’s employees wrote while running the place.
It’s a lot of material. Roughly 100 million emails and 500 million Teams messages, plus millions of documents and spreadsheets. Google says it could use it to improve its products and train AI models. The bankruptcy court still has to approve the sale, and more than 120 US lawmakers have already raised concerns about employee privacy.
I think this is the clearest signal yet of where AI development is heading, and it should make anyone who runs enterprise technology a bit uncomfortable.
Why an airline’s inbox is worth money
The first wave of large language models learned from the public web. That made them good at explaining things and drafting a decent business plan. It didn’t teach them how a business actually runs, because most of that never gets published. When a supplier misses a delivery, the real fix lives in a reply-all thread and a Teams argument about who should call the customer. The procedure manual says one thing and the inbox says what people actually did.
That is what developers building AI agents want. The value isn’t in any one email so much as in following a problem from the first message through the cost estimate to the decision, and doing that millions of times. Corporate records are messy and contradictory, and training on them won’t automatically produce better AI, but synthetic examples have a hard time faking real organizational mess.
Spirit isn’t alone either. Business Insider reported this month on a growing market for internal workplace data, from companies still operating and companies that aren’t. Meta suspended an internal program after employees objected to it collecting keystrokes and mouse movements so AI could learn how people do digital work. Not every deal feeds a shared model, but the demand underneath is the same. Developers want examples of real work.
The people who wrote it
Bankruptcies have sold data before. Toysmart, Borders and RadioShack all ran into privacy fights over customer databases. Spirit is different because the buyer has no interest in running an airline. It wants the records to build AI.
The employees who wrote those messages were doing their jobs, handling an angry customer or sorting out a dispute with a colleague. None of them agreed to become training material. Google says an independent third party will strip out personally identifiable information before the transfer. I suspect that’s harder than it sounds. Workplace conversations are full of specific incidents and relationships, and you can often work out who someone is long after their name is gone. The archive will also hold confidential details about customers and suppliers who were never asked.
Then there’s the question the privacy debate mostly skips. If these records help build systems that automate the work those employees did, should they get a say in the sale? Should they share in what it’s worth?
The promise vendors make
This is the part that bothers me most. Microsoft tells Copilot customers their business data won’t be used to train shared foundation models without authorization. Slack makes a similar commitment about third-party AI training. Those promises close enterprise deals, because companies want AI that understands their operations without handing their secrets to a competitor. (I ran one of the first enterprise AI pilots in the YMCA federation on Copilot and ChatGPT, so I’ve been on the buying side of that promise.)
Spirit shows how far that protection goes, which is as long as the company stays in business. Once it fails, the archive can become something its creditors want to sell.
For years the practical advice on records was to keep what compliance and legal discovery need and let the rest go. Now there’s a buyer for the rest. I think boards are going to start treating that history as something they own and can sell, and staff are going to want it in writing who gets to decide that before a creditor does it for them.
Frequently Asked Questions
What exactly is Google buying, and what is excluded?
The internal side of the business. Court filings describe roughly 100 million emails across about 80,000 accounts, 500 million Microsoft Teams chats, some 3.4 million payroll records, and employee data going back to 1986, along with spreadsheets, calendars, financial databases, audits, presentations, HR files and project management records. Passenger and loyalty data is carved out, which covers 97.5 million passengers, 52.4 million loyalty members and 740,000 co-branded cardholders. Credit card information is excluded too. So the thing being sold is the record of how the company was run, not the record of who flew on it.
Has the sale actually happened?
Not yet. Google won the auction at $10 million after bidding against AI data company Mercor, but the bankruptcy court still has to sign off, and the hearing has slipped to 6 November after Congressional Democrats asked for stronger privacy safeguards first. A rival bidder, Micro1, has since put forward a higher offer. The flight attendants’ union has formally objected unless employee protections improve. None of that guarantees the deal dies, and the auction result tells you the market price for this kind of archive regardless of who ends up with it.
Can workplace data really be de-identified?
Names and account details come out cleanly enough. The context does not. A thread about a specific flight cancellation on a specific date, between the person who made the call and the manager who questioned it, identifies both of them to anyone who was there, and often to anyone who can cross-reference a public record. That is the union’s argument and I think it is the right one. Removing identifiers from structured fields is a solved problem. Removing them from a decade of conversation is not.
Could my company’s Copilot or Slack data end up in a sale like this?
The vendor promise and the insolvency question are two different things. Microsoft and Slack commit that your tenant data won’t train their shared models, and that commitment holds while the relationship holds. What Spirit demonstrates is that your own archive is an asset on your own balance sheet, and in insolvency a trustee sells assets. The protection you bought from the vendor says nothing about what your creditors can do with the same content.
What should a business do about this now?
Three things worth doing before it matters. Know what you are actually retaining and why, because “keep everything, storage is cheap” now has a cost attached that nobody priced in. Put the question of who can authorize a transfer of employee communications in writing, in policy and ideally in employment agreements, rather than discovering the answer in a creditors’ meeting. And if you are a buyer of enterprise AI, ask your vendor what happens to tenant data if you go under, not just what happens while you are a customer. The second question is the one contracts rarely answer.