Old emails won’t teach an AI your job

Google has more data than almost anyone. So why bid million for a bankrupt airline's old work emails? In August, Google won the auction for Spirit Airlines' archive with a million bid, ahead of Mercor at .5 million. Spirit's inventory listed around 100 million emails and half a billion Teams records, plus code and business documents. The sale still needs approval at a hearing set for October 14.

The theory behind deals like this is that with enough records of work (messages, call transcripts, documents) AI can learn to do knowledge work. There are good reasons to doubt it, and it's worth understanding why, because the answer says a lot about what our work is actually made of.

From verifiable work to judgement

Much of AI's progress so far has been in verifiable work. If the code runs and the tests pass, it's good; it either works or it doesn't. Most knowledge work isn't like that. How do you know a document is good? Which sentence, or which of dozens of drafts, made it good? The value people create at work is often not visible in the records being sold.

Work versus the performance of work

Take an invoice that doesn't match its purchase order and needs fixing before the month closes. The invoice is in one system and the order in another. The supplier emailed a credit, but the reviewer wasn't copied. There's a disagreement about whether a partial delivery counts as complete. Someone who knows this supplier remembers that it combines two shipments on one invoice, and that an exception approved last month would be a mistake this month.

Say it takes 30 messages and three meetings to resolve. The valuable moment is one person noticing that the credit was never applied. Everything else is coordination: status updates, follow-ups, meetings because two systems disagree, and messages that mainly show someone was involved. An archive stores the message that prevented a costly mistake right next to the one that just says "following up", and an agent reads both the same way.

The ending can mislead too. The thread closes with "approved" and the ticket is marked done, which looks like success. Six weeks later a duplicate payment turns up. Trained on records like that, an agent risks learning the performance of work rather than the work.

Who decides what "done" means

Records alone aren't enough, so buyers are also building training environments: the software an agent works in, an assignment, and checks for whether it succeeded. Mercor, the losing bidder, announced in July that it would acquire a company that builds such environments, and it pays experts to write the assignments and success checks.

That's where the real decisions happen. For the invoice, is closing the ticket enough? Whoever writes the check turns one person's view of a job into the target an agent is trained to hit, and into the evidence a vendor later shows an employer when it says an agent can handle a role.

Data of last resort

The incentives in this market are lopsided.

Sellers are often companies with nothing better to do with their data, like an airline in bankruptcy. New brokers offer to package small-business archives for anywhere from ,000 to million. So the examples agents learn from may come disproportionately from companies that failed.
A healthy company holding an archive worth a million dollars to a lab may be able to do more with it itself: it knows its market, its customers and its product.
Buyers want clean results to resell, which pushes them to turn the open-ended mess of knowledge work into narrow targets: one environment for invoices, another for seat assignments. That loses the ambiguity that makes the work hard, which may be part of why AI's effect on jobs has been slower than predicted.
Workers created the records but have no seat at the table. Spirit's flight attendants' union has raised concerns about confidential employee information, including sensitive details that end up copied into email even when whole systems are excluded.

What agents should learn instead

Train an agent on chat logs and you'll get agents that are very good at stand-up messages saying there's nothing to report. Where agents add value is different: repeatable, high-drudgery work where the agent reads the data, processes it and produces a defined output, with no need to go through email or chat at all.

The same goes for documents. A model can lay out a polished slide deck, but a deck exists to align people, and if the thinking and the story aren't right, the rest doesn't matter. Generating it isn't the work; the alignment is.

The setups that work best are the unglamorous ones: a clear, structured record of truth that agents can read, such as well-kept markdown files that both people and agents use. People then spend their time on careful product documents and on evaluations that help the agents do better. That's real work, and it isn't what's being sold in these archives.

What to do about it

Explain your value in outcomes. Your contribution isn't any single email or document; it's how you help the business earn more or spend less. If you can describe that clearly, you can use as much AI as you like to get there faster.
If your company is considering a sale, ask about the scope of use, how sensitive records are handled, and whether personal information is included. Pick a few representative tasks and spell out what a correct result depends on, or the data loses value and the agents don't help.
Give workers a say. People should be able to say what counts as success and which records to leave out, and the expertise needed to turn records into training material should be recognised and paid as high-value work.
Be wary of the view from the top. Tools that search across all of a company's chat, email and code give leaders more visibility, but only of what's in the records, which are usually partial, messy and contradictory.

Before records of work get bought, sold and turned into agents, the people involved need a credible answer to a simple question: what is the work? It's rarely what the inbox shows.