Why would AI companies want our emails?

Because, buyers say, AI systems built to do office and trade work need examples of how real work moves (requests, quotes, approvals, schedules, invoices), and that record sits in ordinary business email and software, not on the public web. Epoch AI projected in 2024 that, on current trends, models will fully utilize public human-written text between 2026 and 2032.

What do AI companies say they want our records for?

To train systems that do a company’s work, using examples of how that work was done. As of October 2026, the companies buying and brokering these records put it in similar terms:

• Mercor: “Frontier AI models are increasingly being trained to work inside companies, navigating real tools, workflows, and decisions. Public data and synthetic datasets cannot replicate that.”
• Scale AI: “AI companies want to improve their models for certain corporate personas such as marketers or financiers.”
• Handshake AI: “The next frontier is understanding how businesses actually operate, how decisions get made, how teams use tools, and how industries move.”
• Avelence: “AI buyers license company data for the context behind real work”.

The words that recur are workflows, decisions and tools. A system meant to answer a customer, price a job or chase a late invoice has to see many examples of people doing exactly that, with the back-and-forth, the attachments and how it ended. Most of the companies above supply AI labs rather than building the models themselves; the buyers directory describes each one on the same fields.

Haven’t they already used everything on the internet?

Much of what is public, yes, and the public web never held much of this kind of record. Epoch AI’s researchers projected in 2024 that, if trends continue, language models will fully utilize the stock of human-generated public text between 2026 and 2032 (Epoch AI). That is a conditional projection, not a date when progress stops.

Some kinds of work were well covered online. TIME reported in August 2026 that coding models improved quickly partly because “there was plenty of coding data already out there on the internet”. Service businesses rarely publish the inside of their work, such as a bid that was revised twice or a schedule that moved for weather. That record exists only inside companies, which buyers such as Mercor give as the reason they pay for it. Google is one of them: VentureBeat reported on September 30, 2026 that Google was expanding a pilot paying developers and small businesses in about 100 countries for proprietary code and other non-public content, without disclosing what it pays.

We put roofs on warehouses. Why would anyone want our emails?

Because your email is a written record of how a job moves from first call to final payment, made at the time by the people doing it.

For example, trace one job at a hypothetical 60-person commercial roofer:

1. A property manager emails photos of a leak over a distribution warehouse and asks for a bid.
2. The estimator visits, shoots photos and drone footage, and builds the takeoff in estimating software.
3. The bid goes out as a PDF. The general contractor replies with questions about the insulation spec and asks for an alternate price.
4. The revised number is approved by email and a purchase order follows.
5. Material is ordered from a distributor, and the delivery date moves twice in a chain of replies.
6. The superintendent schedules a crew in chat and reschedules for rain.
7. Inspection photos, a change order for rotted decking, the invoice, and a follow-up when payment runs late.
8. Three years later a warranty call refers back to the original job.

Each step is a decision made with specific information, by a named role, with an outcome written down. A system being trained to help an estimator, a project manager or the property manager on the other end learns from thousands of jobs like this one: what was asked, who decided, what changed and how it ended. A roofer that has done this for twenty years may hold thousands of them, if its systems kept them.

Trade businesses are on providers’ lists. Handshake AI’s example sources include ServiceTitan and AutoCAD, Telegraph Lab (affiliated with this site) lists construction and property among its target industries, and Polyshares’ track-record page lists an anonymized roofing company at $450,000 for a perpetual exclusive license, a figure that is self-reported, undated and not independently verified (as of October 2026). Whether your company clears each provider’s rules is covered in Does my company’s data qualify?

Is any of this better than the rest? What’s the good stuff?

Yes. Records where people ask, decide and explain count for more, especially when they link up across systems over a long, unbroken stretch of years. Providers describe it this way, as of October 2026:

What counts for more	How providers put it	What it looks like in a trade business
Threads with replies	Polyshares: “The value is not in volume. It is in reasoning that was captured while the work happened.”	An email chain where the estimator explains why the price changed
Records linked across systems	Appen prices on “the number and connectedness of systems”; Miro Advisory values records that follow work “from the initial request through the information reviewed, the decision made, the action taken and the final outcome”	The bid email, the job in field-service software and the invoice in accounting, all carrying the same job number
Long, continuous history	Scale AI looks for years of records in established systems	Ten years of mail on one platform, with no gap at the year you changed systems
Real work, not demos	Scale AI: “Day-to-day processes — not demos or synthetic data”	Live jobs and real customers, not templates or training material

What counts for less is mail nobody answered: supplier newsletters, automated alerts, receipts and system notifications. They add bulk but show no decision. A mailbox that is mostly notifications holds less than its size suggests, which is why a buyer looks at a sample before pricing.

You can see which kind you have before talking to anyone. Data Licensing Report’s free sample tool reads a mailbox or Slack export in your browser and counts threads and people, and messages by month, so a thin year or a migration gap shows up in the inventory it builds.

Does that mean they want our customer names?

No. The programs say identities come out before the data is used, and the work stays in. Troveo says “People’s names, contact details, credentials, and client identities are scrubbed under one documented standard before your data is put to work.” Polyshares says its anonymization partner removes names, emails, phone numbers, account numbers and addresses while keeping roles, timestamps and work content, and Scale AI says employee and customer identities stay with the company. How well that holds in a small company, and what to check, is covered in what “anonymized” means; the wider trade-offs are in what’s the catch.

Has anyone actually paid for emails and chat?

Yes, with the largest public figures coming from companies that had closed. Bloomberg Law reported that Google was named the winning bidder, at $10 million, in an August 2026 bankruptcy auction for de-identified Spirit Airlines records that included about 100 million emails and 500 million Microsoft Teams chats, with Mercor’s $7.5 million as the backup bid (Bloomberg Law). As of late September 2026 the sale still needed the bankruptcy court’s approval (Moneywise). Forbes reported in April 2026 that wind-down firm SimpleClosure had handled nearly 100 data deals for defunct companies in the prior year, typically $10,000 to $100,000 each (Forbes).

For operating companies the public record is thinner. micro1 said in a court filing, Fortune reported in September 2026, that it had completed more than 50 data transactions in the previous 45 days, without naming sellers, prices or buyers or saying whether the sellers were still operating (Fortune). Are AI companies really paying businesses? sets out the evidence, and how much work it takes covers what licensing asks of you once a buyer is interested.