Google has agreed to pay $10mn for what a dead airline knew. A judge considers the sale on Wednesday morning. Reuters reported the price on Monday.

Every account of the deal rests on one word. The data is deidentified, so nobody need worry.

The Spirit Airlines data sale agreement, filed with the bankruptcy court on 14 August, says three things about that process which have gone unreported. Read together they change what the word is worth.

Google picks the firm that does the scrubbing

Coverage has described an independent third party cleaning the data before Google receives it. The contract is more specific.

Spirit must deliver the material to one or more third parties “acceptable to or designated by Buyer”. Google chooses the agent.

Google also pays. The agreement makes the buyer solely responsible for every cost of deidentification, and states that those costs do not reduce the purchase price. The $10mn is the headline, not the bill.

Google then gets to inspect the work. Spirit must give the buyer a reasonable opportunity to review and comment on the process. It must give good faith consideration to those comments. The standard is a scrub reasonably satisfactory to Google.

None of that is improper. It is standard commercial drafting. It is simply not what an independent audit sounds like.

The scrub has to keep the records linked

This is the clause that matters most, and it is the one nobody quoted.

The agent must certify the work against the California Consumer Privacy Act standard. Health-related material goes to the federal health privacy rule. Both apply whether or not those laws would otherwise reach this data, which is a real protection.

Then the sentence ends with a condition. The certification must hold “while preserving referential integrity across the data set”.

Referential integrity means the joins survive. The joins survive by design. One pseudonymous person still runs from an email, to a support ticket, to a code commit, to a payroll record.

That is exactly what makes the archive valuable for training agents. It is also the property that makes any anonymisation fragile, and the contract requires it.

What is actually in the box

The schedule of Spirit Airlines data is more specific than the summaries suggested.

It lists 100 million emails across 80,000 accounts and 500 million Teams messages. Add 17,082,644 OneDrive files, 20,577,677 SharePoint files and 667,563 IT tickets. James Nani first reported the headline volumes for Bloomberg Law.

The engineering side runs to 516 repositories and roughly 30 million lines of code. It also carries 372,585 commits, 43,170 pull requests, and the pipeline logs around them.

The operational data is enormous. It covers 763,391 flights and 5,014,676 crew pairings. It holds 190,312,864 booking records and 7,510,221,520 transactions reaching back to May 2008. Disruption and reaccommodation add 3,000,347,472 rows.

Then the corporate interior: board presentations, budget walkthroughs, deal pipelines, due diligence reports, investment committee papers, lender materials and merger fairness opinions.

And the staff

The schedule lists 175,658 employee records, with the system of record running from August 1986.

It adds 3,426,618 payroll records and 148,018 employee tax forms. Then 1,092,000 time cards, training records, recruiting files and travel requests.

Roughly 17,000 people lost their jobs when Spirit stopped flying on 2 May. Their correspondence, pay history and tax paperwork are line items in a schedule now.

They signed employment contracts, not data licences. In a Chapter 11 estate the distinction does not arise.

What Spirit kept, and what it may still sell

The schedule marks the customer side not included throughout. The volumes are large: 97.5 million customer profiles, 50.2 million Free Spirit members, 740,000 card holders, 30,865,471 call recordings and 15,784,473 chat sessions.

Regulatory records sit on the same side of the line. That includes 2,491,715 disability service requests, denied boarding data and complaints to the Department of Transportation.

One clause is worth reading closely. Spirit may not sell the assets to anyone but Google, with a single exception: it may sell its customer data list, including individual traveller spend aggregated by year, to buyers in the hospitality or travel industries.

So this sale did not shield the passengers. It separated them, and the estate kept the right to market them elsewhere. Axios reported the exclusions on Monday.

The bidding, and what it revealed

Google opened at $5mn. Mercor, an AI data company, countered at $5.2mn, then offered $7mn if it could take the raw data first and anonymise it itself, Business Insider reported. Google closed at $10mn.

A bidder priced the unscrubbed version above its own bid for the scrubbed one. Mercor, which sought a $20bn valuation in July, remains the backup buyer at $7.5mn.

Google had been in diligence for a while. The agreement references a confidentiality agreement with Spirit dated 18 June.

Why an airline

The value is not text to pretrain on. It is a chain of consequence.

The archive connects a ticket to the emails about it, the commit that followed, the review thread and the operational result. Agents completing multi-step work need exactly that, and scraped text cannot supply it.

Google has airline-specific reasons too. Google Cloud signed a five-year partnership with Ryanair on 12 August covering fleet operations and maintenance scheduling.

It fits a pattern. Google is in talks to pay $1.5bn for a 35-person startup building coding environments, and China hit the same wall on training material this month.

Meta tried collecting this kind of material from live staff and paused the programme after a revolt. An estate has nobody left to object.

What would settle it

Almost nobody can object now. The deadline for written objections passed at 4pm on 17 August, and the court required anyone attending to register by 11am on 18 August.

Judge Sean H. Lane hears the Spirit Airlines data sale at 11am on 19 August, over Zoom.

Two questions outlast it. Who audits a deidentification that the buyer designed, paid for and approved. And who buys the customer list, given the estate kept the right to sell it.

Get the TNW newsletter

Get the most important tech news in your inbox each week.