AI Companies Offer Up to $2 Million for Archives of Emails and Work Documents

Emails, tickets, documents and call transcripts have suddenly become an asset that can be sold to AI developers. Intermediaries are sending companies offers worth hundreds of thousands, even millions, of dollars for a copy of their old work archives. The reason is simple: the public internet is good at teaching a model to continue text, but poor at showing how decisions are really made, mistakes corrected and projects carried to completion.

A simple example shows the difference. A training dataset might contain a neat instruction for returning a product. A corporate archive shows the whole chain: the customer's complaint, a dispute between support and accounting, an exception to the rules, approval of a discount and the final reply. For a model meant to do office work, the second set is far more useful. It contains context, uncertainty and the consequences of decisions.

Why archives have become expensive

According to WIRED, intermediaries offered between $350,000 and $2 million for corporate archives. Handshake separately buys professional documents and student work. Mercor and Micro1 collect data on real work processes to train and evaluate models.

The scale of the market is still hard to gauge. Micro1's chief executive claims companies spend $40–50 billion a year on data, but the WIRED author could not independently confirm that figure. The forecast that the market will grow to $1 trillion a year is likewise a claim by a market participant, not an established fact.

A telling case arose during the bankruptcy of Spirit Airlines. A digital archive was put up for sale that included 80,000 email accounts, 500 million internal messages, 20.6 million SharePoint files and a million time-tracking records. Micro1 offered $12.5 million for it. The deal was suspended after objections from the flight attendants' union: employees fear that along with work materials, the buyer would obtain their personal information.

Anonymization doesn't solve everything

Data sellers insist they remove personal information before handing over archives. For a table where the name sits in a separate column, this is a relatively straightforward task. Emails and messages are harder. A person may describe a diagnosis, a family problem or a conflict with a manager without ever giving a full name. The combination of job title, project, date and conversation details can still identify the author.

There is also a second question: who owns an employee's work trail. A company owns the email system and the documents, but it doesn't follow that it has obtained people's informed consent to train a model. It is especially unpleasant if the archive is used to automate the very professions whose members created it.

For developers, this means a simple promise of "we removed the PII" is not enough. You need verifiable rules on data provenance, a list of permitted uses, retention periods and a mechanism for deleting a specific person's materials. Independent verification of anonymization, memorization tests for text fragments and an access log for the source archive are all useful.

What a product team should do

Before buying or selling work data, build a map: who created each type of material, which contracts applied at the time, and whether it contains information about customers, health, finances or labor disputes. Check attachments, backups and quotations from old emails separately; they often survive automatic cleanup.

Don't load such an archive straight into training. First select a limited sample, remove unneeded fields, run a manual review and try to re-identify people from the remaining context. If your team can do it, an outsider probably can too.

The main takeaway for founders is simple: a corporate archive can indeed become a new line of revenue, but it also creates legal, reputational and product debt. Its value can't be judged by the size of the check alone. Before any deal there must be employee consent, a right to opt out, a technical audit and a precise limit on the purposes of use. Otherwise the company is selling not dusty files, but other people's working lives.