When owners first hear that AI companies are paying for business data, most assume it means customer lists or financial statements. It almost never does. What buyers want is the everyday record of how work gets done.

The record of real work

AI models have read most of the public internet. What they have rarely seen is how a real company operates over years: how a dispatcher reroutes a late load, how an account manager handles a renewal that is slipping, how a project lead answers an RFI, how a support team escalates a ticket. That record lives in your email, chat, CRM, ticketing system and shared documents.

Teams building AI agents need exactly this kind of material to train and test systems that can do workplace tasks. Public data cannot supply it, which is why buyers license it directly from companies.

What is usually in scope

• Communications. Email threads, calendars, Slack or Teams channels and meeting notes.
• Sales and CRM. Pipelines, account history, quotes and proposals.
• Support. Tickets, help desk conversations and knowledge bases.
• Operations. SOPs, scheduling and dispatch logs, project boards.
• Engineering. Repositories, issues and technical documentation, for software firms.

Buyers are not buying your secrets. They are buying the pattern of how competent people do their jobs.

What is usually out of scope

You choose what is included. Owners commonly exclude HR and payroll files, privileged legal communications, records covered by client NDAs and anything regulated, such as health or payment card data. Names and identifying details are removed by the buyer before any records leave your control.

Why the ordinary is valuable

The more ordinary and consistent your records are, the more useful they tend to be. Years of the same systems, used by a real team, give buyers a clear view of how work evolves. That is the asset, and it is one most established companies already have without realising it.

This article is general information, not legal, tax or financial advice.

Once a buyer makes a written offer, the details of the contract decide how comfortable the deal will be. This is not legal advice, and every owner should have their own lawyer review the terms. But these five clauses are where most of the important decisions sit.

If a term matters to you, it belongs in the contract, not in a conversation.

Before you sign

Nothing moves until you sign, and you can walk away at any point before then. Take the time to get these five clauses right.

This article is general information, not legal, tax or financial advice.

Many owners assume that once a company closes, its records are worth nothing. For AI buyers, the opposite can be true. A business that ran for ten years and then closed holds a complete record of how it operated from start to finish.

Question one: do you still have the records?

Records need to exist and be accessible: an archived email account, an exported CRM, a retained Slack workspace or backups of shared drives. If accounts were shut down and nothing was kept, there may be nothing to license. If they were archived, there may be a lot.

Question two: did the rights stay with you?

If the company simply wound down, the records usually remain with the owners or the entity. If the business was sold, the purchase agreement decides who owns them. In many asset sales, records transfer to the buyer; in others, the seller keeps them. Your lawyer can confirm which applies.

A closed company can hold the most complete record of all: the whole life of a business, from first hire to last invoice.

What buyers look for

The same things as with an operating company: several years of history, a team of 30 or more at your peak, and systems that were used consistently. Industries that are rarely represented online, like freight, trades, insurance and distribution, are especially sought after.

The next step

A fifteen-minute call is enough to tell whether your records are likely to interest a buyer. There is no cost and no obligation.

This article is general information, not legal, tax or financial advice.

Professional services firms, agencies and contractors often have confidentiality agreements with their clients. It is right to take them seriously. The good news is that they rarely stop a deal; they shape what goes into it.

Start with an inventory

Before any offer, list the client agreements that include confidentiality or data-use terms. Your lawyer can review which records each one covers and whether licensing de-identified copies is permitted, restricted or prohibited.

Exclude what is covered

Records tied to clients with restrictive terms can be left out entirely, by client, mailbox, channel or date range. Many deals go ahead with a handful of client accounts excluded and everything else in scope.

Exclusions are normal. Buyers would rather have a clean, narrower dataset than a broad one with uncertain rights.

Make it explicit in the contract

The license should list the excluded records clearly, and the buyer should confirm they will not be transferred. De-identification removes names and identifying details from what is included, but it does not replace the need to respect your agreements.

Talk it through early

Raise confidentiality on the first call. It helps match you with buyers whose programs fit, and it avoids surprises once an offer is on the table.

This article is general information, not legal, tax or financial advice.

In September 2026 Mercor’s CEO said the company had partnered with more than 300 enterprises to license operational data, with some payouts reaching $10 million. The same note listed the systems that appeared most often across those partners.

These percentages are more useful than the headline partnership count. They show buyers are not primarily hunting exotic or single-purpose datasets. They are looking for the ordinary stack that most established companies already run, and especially for the connections between those tools.

Why cross-system traces matter

Language models have already trained on the public web. The scarcer input for agent-style systems is a record of how work actually moves: a request that starts in email or Slack, becomes a ticket or Jira issue, pulls context from Drive or Figma, and ends in a decision or code change. A single-tool export, only CRM or only Slack, is less valuable than the linked sequence across tools. Mercor’s system list is effectively a map of where those sequences live in real companies.

A 200-person firm and a 5,000-person firm often share the same core tools. What differs is depth of history, cleanliness of rights, and how completely the handoffs are captured. That is why published ranges across the broader market still cluster in the $100K to $2M band for typical established companies, while outlier archives can clear higher.

A practical note

For a company reviewing its own records, the relevant question is not “Do we have Slack?” but “Can we produce a coherent, rights-clean extract that shows how work moved across two or more of these systems over multiple years?”

This article is general information, not legal, tax or financial advice. Figures are as publicly reported at the time of writing.

Most buyers publish ranges rather than closed prices. Across micro1, Polyshares, Handshake AI, Endless Data and others, the commonly cited band for an established company is $100,000 to $2 million, with some programs listing higher ceilings for rare or very large archives. These are the buyers’ own figures, usually labeled typical or illustrative.

Disclosed dollar amounts remain rare. The clearest public price-discovery event is the Spirit Airlines archive.

The package was described as roughly 100 million emails and 500 million Teams messages, plus related business documents. Passenger and loyalty data were excluded. As of early November 2026 the sale still required court approval.

Why the gap exists

Most voluntary licenses include confidentiality provisions. Bankruptcy and distressed sales are simply more likely to become public. Spirit is also an outlier on volume; few operating companies will match hundreds of millions of messages. The bids confirm that buyers will compete aggressively for large, multi-system operational history, but they do not reset the typical mid-market range.

Other scattered public or semi-public figures for smaller company archives have tended to land in the low-to-mid six figures. Polyshares, for example, has referenced individual deals between roughly $110,000 and $450,000. These remain the exception rather than a full dataset.

Scale and cleanliness move the number. Most companies sit inside the published band until proven otherwise.

A practical note

Treat published ranges as the starting reference for qualification conversations, not as a quote. The variables that actually shift an offer are years of history, the number of systems that can be linked, domain rarity, and how cleanly exclusions such as client NDAs, HR and privileged material can be applied.

This article is general information, not legal, tax or financial advice. Figures are as publicly reported at the time of writing.

The Spirit process remains the most detailed public look at how a large operational archive is scoped. The estate offered employee emails, Teams messages, spreadsheets, calendars, marketing material, operations data and productivity records. Passenger and loyalty profiles were kept out from the start.

Even with those exclusions, the Association of Flight Attendants-CWA objected, arguing that residual content could still expose employee grievances and internal investigations. The objection did not stop the bidding, but it showed that removing the most obvious personal-data categories does not automatically eliminate every downstream concern.

What this maps to in voluntary deals

Operating companies are in a stronger position than a bankruptcy estate, because they choose the scope before any introduction. Common exclusions in ordinary licenses include:

• Client work under NDA
• HR, compensation and performance files
• Privileged or legally sensitive communications
• Raw regulated personal data that cannot be adequately de-identified
• Anything the company simply prefers to withhold

Originals stay in the company’s systems. Only a de-identified copy moves, and only after the company and its counsel have reviewed the terms. The Spirit objections do not change the existence of the $7.5 million to $12.5 million bids; they do reinforce why scope definition and legal review are standard rather than optional.

A practical note

The useful pre-conversation inventory for any company is short: which systems, how many years, who can sign, and which categories must stay out. Those four items determine whether a buyer conversation is even relevant.

This article is general information, not legal, tax or financial advice. Figures are as publicly reported at the time of writing.

When a company licenses operational records to an AI buyer, it grants permission to use a defined, de-identified copy for specified purposes. It does not transfer ownership of the underlying data, the systems that generated it, or the business itself. The originals remain in the company’s control. The license is bounded by the scope the parties write down: systems, date ranges, permitted uses, term and any exclusivity.

This structure appears consistently across the programs that have published terms. micro1, Mercor, Polyshares, Handshake AI and others describe the transaction as a license in which the company retains ownership. What moves is a copy from which personal identifiers have been removed. The company continues to operate on its own systems exactly as before.

License: defined permission for a de-identified copy. Ownership and originals stay.

Why the distinction matters

First, a license does not appear on the cap table, create a repayment obligation, or require the company to stop using its records. Second, because ownership stays with the company, the decision about what is included remains with the company and its counsel up to the moment of signing. That is different from an asset sale in bankruptcy, as in the Spirit process, or from a transaction in which raw records leave the company’s environment before terms are final.

Why both sides prefer it

Buyers want the license structure because it lets them obtain training material without taking on the legal and operational burden of owning another company’s systems or employee records. Companies accept it because the exchange is limited to a defined copy and a defined use. The remaining questions, such as how long the license runs, whether it is exclusive, and what happens if the de-identification method is later challenged, are contract questions, not questions about whether the underlying model is a sale.

Where Buzz Data Desk fits

Buzz Data Desk is built around this boundary. The fifteen-minute call establishes only whether the company’s headcount, years of history, systems and signer put it inside the scope buyers are purchasing. If it does, the introduction is direct. The company then reviews the actual license terms, including the de-identification method, with its own counsel. Originals never leave the company’s systems. Buzz Data Desk does not copy or store the records, is paid by the buyer and never from the company’s payout, and the company can decline at any point before signing.

This article is general information, not legal, tax or financial advice.

In the programs that have published descriptions, the originals stay with the company. The only thing that moves is a de-identified copy of the records the company has chosen to include. Categories that commonly remain out are HR and personnel files, material covered by client NDAs, privileged communications, and regulated personal data that cannot be reduced to the standard the buyer and the company’s counsel accept.

The public example

The Spirit Airlines filings remain the most detailed public illustration of the same principle at scale. Passenger and loyalty profiles were excluded at the outset. A third-party de-identification step was required. Even then, the Association of Flight Attendants raised residual concerns about employee grievances and investigations. The bids still occurred. The episode shows both that exclusions are normal and that excluding the most obvious categories does not automatically resolve every downstream question.

Your scope. Your originals. Your counsel. Your decision.

Why operating companies are in a stronger position

Operating companies are better placed than a bankruptcy estate because they define the exclusions themselves, before any introduction, and they review the method with their own counsel before anything is transferred. The practical protection is contractual and sequential: scope is written down, samples can be reviewed, the de-identification standard is described, and the company keeps the ability to stop before signing.

Where Buzz Data Desk fits

Buzz Data Desk does not alter that sequence. The fifteen-minute call covers systems, years and who can sign. It does not include documents or logins. If the profile fits, the introduction is direct. From that point the company controls what is in and what is out. Every offer and every description of the de-identification method is reviewed by the company’s own lawyer. Originals remain exactly where they are. Buzz Data Desk never copies or stores the records, is paid by the buyer rather than from the payout, and the company can walk away at any point before signing.

This article is general information, not legal, tax or financial advice.

Public web data trained the current generation of language models. The scarcer input for systems that are expected to act inside real organizations is a record of how work actually moves. A request arrives by email or Slack, becomes a ticket or Jira issue, pulls context from a document or CRM record, and ends in a decision, a code change or a customer outcome. That sequence is spread across ordinary tools. A single-tool export captures only one slice of it.

The stack buyers keep naming

This is why program descriptions and partnership data repeatedly list email, chat, CRM, ticketing, document stores and code repositories together. Mercor’s published system frequencies (Slack, GitHub, Drive, Gmail, Figma, Jira) are one visible example. The value is not in the volume of any one system; it is in the linked trace that shows state changes, handoffs and outcomes. A 200-person company and a 5,000-person company often run on overlapping stacks. What differs is the length of the history and how completely the handoffs were captured.

Value sits in the links between tools, not in any single export.

The question to ask of your own records

For a company, the evaluation question is concrete: can it produce a coherent, rights-clean extract that shows work moving across at least two of these systems over multiple years, after the categories that must stay out have been removed? That question can be answered from an inventory. It does not require handing over data.

Where Buzz Data Desk fits

Buzz Data Desk asks exactly that inventory question on the fifteen-minute call: systems, years, headcount and signer. If the profile matches the criteria buyers publish, the introduction is direct. The company then decides which systems and date ranges are included and which categories are excluded. The license, if signed, covers only the de-identified copy of what the company approved. Originals stay. Buzz Data Desk is paid by the buyer, never from the payout, and the company can decline before signing.

This article is general information, not legal, tax or financial advice.