How to Sell Your Company's Data to AI Labs (2026)

A holder's guide to licensing a company's records for AI training: the buyers, the prices, the rights work, the process and the fees

Google bid $10 million for a bankrupt airline's emails, chats and tickets in August 2026, and a rival offered $12.5 million three weeks later. The airline was Spirit, the bidding opened at $5 million, and three qualified bidders fought over a corporate archive that held about 100 million emails and 500 million Microsoft Teams items - Chapter 11 docket summary. Two months later, the data vendor micro1 announced it would spend $1 billion over 12 months buying and licensing the operating records of ordinary companies - The Next Web. A market that barely existed in 2024 now has published price tiers, referral bounties and court precedent.

But most holders meet this market through one buyer's estimator, and that is the weakest position to sell from. Published figures run from a $5,000 baseline for a single graded code repository to $4 million or more for a large operating company's records, and the same company can see indications that differ by a factor of twenty depending on which form it fills in. Price is rarely what kills a sale, though. Rights do: customer contracts, software vendors' API terms, employee messages, personal data and old privacy promises. A holder who finds those problems after a buyer has seen a sample has lost both time and leverage.

This guide is written for a founder, CEO, CFO or the person winding a company down. It explains why AI labs pay for company records now, who buys and on what published terms, what each kind of record fetches, the rights questions that end deals, the step-by-step process with realistic timelines, how the answer changes for an operating, closing or closed company, why a competitive process beats a single offer, how intermediaries are paid, and the red flags to watch. It ends with a decision framework and answers to the questions holders ask most.

A note on who is writing. AIData.Broker acts as the holder's agent in this market: the holder keeps ownership of its data and licenses it to the buyer itself. We value data against public prices, ask the holder to settle its rights file before any buyer sees anything, and put the data before every buyer that fits at the same time, without naming the holder until a buyer is interested and has signed an undertaking. We do not buy a holder's data, take ownership of it, or license or sell it in our own name, and we are not the agent of any buyer. Only the holder pays us, a fee fixed in the mandate, and only if a sale closes: we take nothing from any buyer, so we earn the same whichever buyer the holder chooses. Where this guide describes our terms, it says so plainly. Everything else is sourced to the buyers' own pages, court records and named press.

This guide is research from public sources, not legal, tax or investment advice: take your own advice before you sign anything.

Contents

Start with At a glance: what each buyer tells a holder before the first call, then:

1. Why AI labs pay for company records now
2. Who buys, and how each buyer works
3. What records sell and what they fetch
4. The rights questions that kill sales
5. From inventory to payment: the process and its timeline
6. Operating, closing or closed: what changes
7. One buyer or several: what the Spirit auction shows
8. Fees: how intermediaries get paid

At a glance: what each buyer tells a holder before the first call

There is no price list in this market, so the first useful question is not "who pays most" but "who tells me enough to decide whether a call is worth my time". The table below scores the ten organisations that publicly buy or license company records and say something checkable about their terms. It is a disclosure score, not a quality score and not a prediction of the price you will be offered. A buyer that publishes little can still pay well; it simply makes you do more work, under NDA, before you find out.

Each buyer is scored from 0 to 10 on four criteria, using only what it published on its own pages (or, for one wind-down service, what named press reported) as read on 11 October 2026. AIData.Broker is not in the table because it is not a buyer: it does not purchase or license data for its own account, so scoring it against buyers would compare unlike things. The detailed profiles and the sources for every cell are in section 2.

How to read the criteria. Price published rewards a concrete range or tiers on the buyer's own page, because that is the only anchor a holder has before a call. Payment terms rewards a stated trigger and timing, since "when do I get paid, and on what event" is where one-off deals go wrong. Eligibility rewards published size, region, language or status rules that let you rule yourself in or out in a minute. Holder control rewards terms that let the holder decide what leaves its systems, who de-identifies, and what happens to the raw copy, which the Spirit sale showed is itself a priced term. The weights are our own judgment, not a market fact: they favour price and control because those two decide whether a sale is worth doing at all.

The scores move as programs change, and they change monthly. The live version of this comparison, with each buyer's source and the date it was read, is the buyer board, and the buyer match tool runs your company's size, region, language and status against each buyer's published rules.

1. Why AI labs pay for company records now

For most of the last decade, AI models learned from the public internet, and nobody paid the companies whose internal records sat behind a login. That changed because the public supply ran low and because the way models improve changed. Epoch AI estimated the stock of quality public human text at about 300 trillion tokens and projected it would be fully used somewhere between 2026 and 2032 - Epoch AI. By mid-2026 the argument had moved further: former OpenAI researcher Will DePue wrote that even verifiable maths and coding tasks are being exhausted, that labs now lean on proprietary datasets and custom tasks, and forecast (as an opinion, not a budget) that leading labs could spend more than $100 billion a year on data by 2030 - Ground Level AI.

The second shift matters more for a holder. The gains now come from reinforcement learning, where a model attempts a task thousands of times and a grader scores each attempt. The practice space is called an RL environment: a sandbox, a set of tasks, and a grader for each, usually a unit test or a rubric. In interviews with 18 people who build and buy them, Epoch found that enterprise workflows (expense reports, spreadsheets, CRM updates, ERP work) were the fastest-growing category of demand - Epoch AI. To train an agent to "close this ticket and update the account", a lab needs a fake company that behaves like a real one. Your records are the blueprint for that fake company.

What the buyer is actually buying

Start from first principles. A lab is not buying your Slack history to read it like a book. It is buying the structure of real work: a request comes in, people discuss it, someone opens a ticket, code or a document changes, a customer is told, the ticket closes. Each link in that chain lets an environment builder reconstruct a realistic task with a checkable outcome. Surge's head of RL environments, Nick Heiner, put it in one line: "RL environments are only as good as the data that populates them" - TIME. When Mercor agreed to buy the environment builder Deeptune in July 2026, it wrote that "the constraint has shifted to the environments themselves" - Mercor.

That framing explains almost everything else in this guide. It explains why linked records (a ticket tied to the commit that closed it and the thread that argued about it) are worth far more than a loose pile of documents. It explains why raw volume is the weakest argument a holder can make. And it explains why buyers care so much about clean rights: a lab that trains on your records has to defend where they came from, and the regulations that now ask it to describe its training data flow down to you through the contract.

The ceilings on what records are worth

Two forces cap prices, and a holder should understand both before believing any estimator. The first is synthetic data. micro1, one of the largest buyers, "increasingly generates synthetic data without human involvement" - TechCrunch. One practitioner summed up the rule bluntly: "Whatever a machine can verify, machines will eventually learn without you" - Thoroughly Intrigued. If a vendor can generate your kind of record cheaply, it will not pay you much for it.

The second ceiling is replication. Vendors build working replicas of websites and software for agents to practise in, at around $20,000 per website replica and around $300,000 for a replica as complex as Slack - Epoch AI. Where a replica can be built from the outside, nobody pays the original rights holder. What cannot be replicated from outside is the private history of how a specific business actually ran: its exceptions, its judgment calls, its industry quirks. That is the scarce part, and it is where the money goes.

A third force is decay. One analysis argued that "AI data is a flow business, not a stock business", and that environments lose value once models learn what they teach, with a half-life of roughly six months - Next Signal. For a holder this has a practical meaning: the buyer paying a premium this quarter for your category may pay much less next year, once similar records from rival companies have reached the same labs. Waiting is a decision with a cost.

Who is spending, and how concentrated it is

Demand is concentrated in a handful of frontier labs, and most of the money reaches holders through vendors that sell to those labs. Mercor reported $614 million of revenue in the first half of 2026, more than 90% of it from foundation-model companies, as reported by The Information - BigGo Finance. Forbes reported that the top six Chinese labs spend about $500 million a year with American data vendors - Forbes. Labs also buy directly: Google won the Spirit auction itself, and Bloomberg reported that SpaceX held internal discussions about buying customer and operational data from failed startups - Bloomberg.

The vendors have money to deploy. micro1's $1 billion commitment is financed with capital from Citi and Hercules Capital - Business Wire via FinancialContent. In a court filing, micro1 said it had completed more than 50 data transactions in the preceding 45 days, without naming holders or prices - Fortune. Compute economics push labs toward quality: Mechanize estimated that a frontier RL task burns about $2,400 of compute over its life, so buying tasks for less than about $500 wastes compute - Mechanize.

Why this matters: the buyer is paying for raw material that it will turn into graded practice tasks, and that raw material depreciates as models learn. How to apply it: describe your records the way a buyer will value them (which systems, how they link, how many years, which industry), not by terabytes, and treat timing as part of price. The rest of this guide builds on that one idea.

2. Who buys, and how each buyer works

The word "buyer" hides three different kinds of business, and the difference decides how your deal is priced. Most of the organisations that publicly take company records are principals: they buy or license your records for their own account, turn them into training environments or datasets, and sell the result to labs, sometimes to several labs. A smaller group are brokers and marketplaces that license your records onward to labs and keep a fee or a spread. The third group are wind-down services that list a closing company's code and workspace data as part of a dissolution. Labs themselves buy directly mainly at very large scale, as Google did with Spirit.

The most important structural fact is that the best-funded buyers act for themselves, not for you. A principal's job is to buy as cheaply as the holder will accept and resell at the highest price the labs will pay. micro1's off-the-shelf datasets, sold to more than one customer, carry gross margins reported as high as 80% to 90% - TechCrunch. That is not a criticism; it is how a principal business works. It does mean that a buyer's published figure is the start of its negotiation, not the value of your records to the market.

The buyer comparison table

The table below sets out what each buyer publishes, in its own words where short, as read on 11 October 2026. "Introduction fee" is what the buyer says it pays whoever introduces a company. It matters because a person who recommends a buyer to you may be paid by that buyer, and you should know when.

Two cautions before you read it. First, every figure is the buyer's own statement of what is possible, not a record of what holders were actually paid: few buyers publish completed deal prices, and the large principals do not name their holders. Second, these programs are young and change often. micro1 launched its referral program in June at up to $25,000 per company - micro1 forum, and showed $50,000 by October. A pricing tracker recorded Mercor showing up to $100K per referral on its data page from late July and removing the figure in a September rebuild - UsagePricing. Treat every cell as a dated reading, and re-read the buyer's page before relying on it in a negotiation.

Read the table from right to left as well as left to right. Six of the ten buyers publish a referral payment, and four of those reach $50,000 or more. Only four publish how soon the holder is paid. And the widest published price range ( Grepped's) comes from a broker that publishes no eligibility rule and calls it what a company's data "could be worth", so the range tells you what is possible, not what is likely.

The chart below shows the published floor and ceiling for the seven buyers that publish both, in thousands of US dollars. Where a buyer writes " $1M+" or " $4M+", the chart uses the stated figure as the ceiling, so the real top of each range is open-ended.

Sources: Grepped, Handshake AI, micro1, Scale AI, Turing, Ooak's floor and ceiling, SimpleClosure.

The large principals: micro1, Mercor, Handshake AI and Scale AI

micro1 publishes the clearest tiers in the market: $100K+ for a qualified partnership, $500K+ for large or multi-team datasets and $1M+ for highly unique data, for companies of 30 or more employees that work and document in English, with the US first and then the UK, Canada and other Western markets - micro1. It says companies keep ownership, that personal information is removed "when relevant", and that some datasets may be rewritten synthetically. Two details deserve attention. Its own payout estimator, run with default inputs, shows $50,284 to $92,776, below its stated entry tier - micro1 estimator. And its referral terms, updated on 8 October 2026, leave payout timing to its "sole discretion" and forfeit unpaid referral payouts on termination - micro1 referral terms. In July it said it was partnering with 50 companies, paying $100K to $2M+ each - micro1 forum. Our micro1 profile tracks its program and deals.

Mercor runs the opposite model: no published price at all, but the clearest payment terms. Its page opens with one question, which tools your team runs on, and says "Connecting more tools generally results in a higher payout" - Mercor. It names eight target industries (legal, healthcare, finance and insurance, professional services, manufacturing, retail and e-commerce, real estate, logistics), covers all extraction and anonymisation costs, says nothing leaves your systems until you approve it, and pays by wire on delivery, typically within 2 to 4 weeks of signing. It also sells into private-equity portfolios: its sponsor page at an operating-partners event offers AI agents "with an option to offset deployment costs through data monetization" - PEI Operating Partners. Its referral app pays 5% of the deal, up to $100,000 - Mercor referrals. See our Mercor profile.

Handshake AI, best known as a supplier of expert work to labs, now buys operational data from businesses "across every major industry" and publishes the widest range among the large principals: partnerships "can range from $100K to $4M+ depending on the data shared" - Handshake AI. It takes companies with 20+ full-time employees and 3+ years in operation. Its estimator shows $383,083 to $896,969 for a US company of 100 people with 10 years of operation and calls that "an indicative estimate only". Handshake de-identifies on its side, but its page adds a sentence every holder should read twice: data partners "remain responsible for ensuring that their transfer" complies with the law. It pays $10,000 to whoever refers a company, when the deal closes - Handshake referral page. Separately, it has paid individuals $6 per accepted page, up to $30,000, for professional documents they own - Business Insider via AOL. See our Handshake profile.

Scale AI, a long-standing data vendor to labs, offers the most seller-controlled process on paper. It shows $10K to $1M+ as an illustrative value per partnership, "scaling with cadence", prefers companies of 40 or more employees, anonymises everything "inside your environment, before any data is extracted", and requires the company's sign-off on every data package - Scale AI. It designs partnerships for recurring revenue rather than a one-time transfer, and pays up to $50,000 to an introducer only if it signs. The trade-off is that Scale publishes nothing about when or how the holder is paid. See our Scale profile.

Single-payment and small-company buyers: Turing, Nyne and Ooak

Turing's Lazarus program buys a company's whole working record for "a single payout of up to $1M", from operating and winding-down companies - Turing Lazarus. Its process is the fastest published: an intake call scheduled within one business day under NDA, a review of a representative sample "usually within a week", then one valuation and one contract. It says "You decide what's in scope, what's redacted, and what never leaves your environment." Turing started with code: its CEO said it had bought five to ten failed-startup codebases at an average in the tens of thousands of dollars each - Fortune. The ceiling is a ceiling, set only after sampling. See our Turing profile.

Nyne aims lower in company size and higher in specificity. It takes US businesses with at least 20 full-time W-2 employees and three years in business, wants accounting records (QuickBooks, Xero, NetSuite), CRM, support tickets and scheduling or dispatch data, and says partnerships "are commonly worth $100K or more" - Nyne. It is the only buyer that publishes a price match: it "will beat a qualifying written offer for the same eligible data". Read that sentence as a market signal: a buyer that promises to beat a rival's written offer is telling you that written rival offers move its price. Nyne pays "once your dataset is delivered and accepted" - Nyne data partnerships.

Ooak, a Y Combinator company, builds an anonymised copy of a company from its tools and licenses it to labs as a training ground. It takes companies of 20 or more employees, operating or closed, and says it works with "multiple frontier labs" and has data from 20 companies - Ooak launch post. Its partner page is unusually specific: deals so far range from $20K, the smallest closed, to several hundred thousand dollars; it advertises 2 to 4 weeks and "about one hour" of the holder's time; exclusions are annexed to the contract; payment goes to the entity, not to individuals; and the raw archive is deleted within 30 days - Ooak.

Marketplaces and brokers: Troveo and Grepped

Troveo began as a video licensing marketplace and now takes business records too. It licenses on the holder's behalf and pays a revenue share on every sale, saying "Typical full-company deals start at six figures"; exclusivity is time-limited, "has to be earned", and comes with exit terms - Troveo business data. It reports more than 7,000 licensors, 95% of them exclusive, and over $20 million paid to rights holders - Troveo. The revenue-share model can earn more over time than a one-off payment if your records sell repeatedly, and less if they do not. Troveo does not publish its percentage.

Grepped calls itself a broker: it signs with the holder, packages the records and sells them to labs. Its homepage says a company's data "could be worth $20K to $5M"; its estimator uses team size, founding year and revenue and shows a sample of $29K to $41.5K; it says most companies go from form to payment in "two to three weeks", with payment within 7 days of data passing review - Grepped. Its buyer page says every source "is licensed under a signed agreement with the data owner" - Grepped buyers, so it contracts with both sides; it does not publish how it is paid. It pays introducers 10% of what it pays the owner, up to $100,000 - Grepped referrals.

Wind-down services: SimpleClosure and its peers

SimpleClosure is a company-dissolution service that lists the code and workspace data of the startups it helps shut down. Its early test recovered just over $1 million across nearly 100 deals, typically paying $10,000 to $100,000 per company, which puts most deals near the bottom of that range - Gizmodo. By September it reported 350+ companies listed, AI buyers up 9x on 2025, more than $10 million on track to be returned by year-end, and that its Asset Hub was open to operating companies too - Business Wire via Webull. It excludes sensitive employee, compensation and health information before licensing - SimpleClosure Asset Hub. A competitor, Sunset, buys defunct-company data at similar prices and prices on size, age and "data richness" - Forbes.

What the estimators really tell you

Almost every buyer now has a one-minute estimator, and holders understandably treat the output as a valuation. It is not. Each one says so in its own small print, and the outputs disagree with each other by more than an order of magnitude because they measure different things with different inputs. Handshake's default ( 100 people, 10 years) shows $383,083 to $896,969. micro1's default shows $50,284 to $92,776. Grepped's sample shows $29K to $41.5K. Each is honest about being indicative; none is a price for your records.

The useful way to read an estimator is as a lead magnet that also leaks information: which inputs a buyer asks for tells you what it prices on. Headcount and years appear everywhere because they are proxies for how much linked history exists. Language and region appear where a buyer sells mainly to English-language labs. Tool count appears at Mercor because its pipeline gets more from more connected systems. For an independent first reading built on published prices rather than a single buyer's funnel, use the valuation tool, and compare it with the full price board. The market tracker shows how these programs and prices have moved month by month.

Why this matters: each buyer prices from its own model and its own resale plans, and the published numbers are anchors, not appraisals. How to apply it: shortlist the buyers whose published rules you meet, note which ones disclose payment terms and holder control, and never treat one buyer's estimator as the market.

3. What records sell and what they fetch

Holders usually ask "what is my data worth?" as if there were one number. There is not, and the evidence explains why: published prices for company records span more than three orders of magnitude, from a few thousand dollars for one code repository to ten million dollars and more for an airline's entire archive. Fortune concluded in September 2026 that "there is no reliable market average for a failed company's data" - Fortune. What exists instead is a scatter of public anchors, some from court records, some from buyers' own pages, some from named press, each with its own conditions.

The right way to use those anchors is as a map of what moved the price in each case, not as a menu. A 13-year archive and a two-year one, or linked tickets and loose PDFs, sit at opposite ends of the same category. The table below collects the anchors that matter for a company's records and code, with the basis of each figure. The full set, including video, audio, images, text and robot data, is on the price board.

Three patterns stand out. First, code alone is the cheapest company record: a repository without the tickets, reviews and discussion around it fetches thousands or tens of thousands, because a vendor can find or generate a lot of ordinary code. Second, age and linkage separate the low end from the high end of shut-down archives: most of SimpleClosure's first hundred deals sat near $10,000, while a 13-year archive fetched hundreds of thousands. Third, the operating-company tiers at $100K and up come with eligibility gates ( 20, 30 or 40 employees, English, years in business) that most very small companies do not meet.

What moves the price, in order

The evidence supports a clear order of price drivers, and it is not the order most holders assume. Volume, which is what holders usually lead with, comes last. The first five drivers are all about whether a buyer can turn your records into realistic, defensible, exclusive training material.

1. Linkage. A ticket tied to the commit that closed it and the thread that argued about it lets a buyer rebuild a whole task: intent, steps, tools and outcome. SimpleClosure found that code linked to Jira, Linear, Figma, email or support logs, and code in less common languages such as Ruby or C++, fetched more - Upstarts Media. Sunset's CEO put it plainly: a Jira ticket tied to a specific code commit carries more value than a standalone document - Forbes.

2. Scarcity. If a vendor can recreate your records by hiring contractors or generating them, it will. Niche industries, unusual tools and uncommon languages cannot be recreated cheaply. SimpleClosure reports buyer demand highest in manufacturing, energy, banking, insurance, professional services and logistics - SimpleClosure. A 15-year history in a regulated industry is scarce; a two-year SaaS startup's Slack is not.

3. Clean rights. A buyer has to defend where its training data came from. Troveo's own recipe excludes customer and employee personal data, third-party confidential material and contractually restricted data before anything is valued - Troveo. Records with a complete rights file move through a buyer's legal review; records without one stall there, and a stalled deal loses its price as the market moves.

4. Exclusivity. Two environment founders independently told Epoch that exclusive deals cost roughly 4 to 5 times non-exclusive ones - Epoch AI. That is the largest single lever on price, and it is the one a single-buyer offer never shows you, because the only way to price exclusivity is to know what the non-exclusive alternatives would pay.

5. Who controls de-identification. In the Spirit auction, Mercor offered $10 million on condition that it ran the scrub itself with its own tools, and lost to a $10 million bid with an independent third-party scrub - docket summary. Control of that step is a priced term.

6. Volume, last. About 37% of raw code disappeared once SimpleClosure removed duplicates, third-party code and generated code across the startup datasets it examined - SimpleClosure. Terabytes measure storage, not value. A buyer counts what survives cleaning.

That list has a practical consequence for how you prepare. The cheapest thing that raises your price is usually not more data but better description: showing which systems link to which, how many years of continuous history exist, and which exclusions you have already made. Each of those answers a buyer's question before it becomes a discount.

Which records are easiest to sell

Kinds of records differ less in demand than in how much rights work they carry. The simplest test is whose words and whose data they mostly contain. Material your company wrote about its own work (procedures, runbooks, technical documentation, engineering files) is the easiest to clear. Material that is mostly other people's words or data (support conversations, CRM notes, internal chat) carries the heaviest rights load, even though buyers ask for it.

This is not a ranking of value. Internal communications are what buyers like Mercor ask for first, and they carry the richest picture of how work happens. The point is that the heavy categories need the most work before they can be offered, and some of that work (consent, exclusions, contract checks) cannot be done in a weekend. The company records and source code pages break down demand, prices and rights for each category.

Why this matters: a buyer prices your records by what it can build from them and defend, so the same archive can be worth $10,000 or several hundred thousand depending on linkage, scarcity, rights and exclusivity. How to apply it: lead with the light-rights, highly linked material, describe linkage and years explicitly, and price exclusivity only once you know what shared rights would fetch.

4. The rights questions that kill sales

Owning the files is not the same as having the right to license them. A company archive bundles other people's rights: customers who handed you their information under a contract, software vendors whose terms govern how data leaves their platforms, employees who wrote the messages, outside parties whose confidential material sits in your inbox, and the people you once promised to treat their data a certain way. Each layer can stop a sale on its own, and each is invisible in a terabyte count. This is where most deals die, which is why the rights work should come before any buyer sees anything.

Lawyers who advise holders say the same thing in sharper terms. A sell-side checklist from Frankfurt Kurnit Klein and Selz tells companies that legal, privacy, security, HR and communications should all be involved, that these deals "should not be run by business development alone", and warns: "Do not share sample data before this review, since a sample itself could be a disclosure" - Legal 500. The same guidance warns that once data trains a model, it cannot practically be pulled back out. A licence can end; a trained model does not forget.

Customer contracts

If your company serves customers, much of what sits in your systems is theirs in a contractual sense. Master service agreements, NDAs and data processing agreements usually limit customer information to performing the contract, which rules out licensing it to a third party for AI training. An analysis of micro1's program published on 10 October 2026 listed customer NDAs, MSAs and privacy purpose limits as the first category of third-party rights a holder must check - BERI.

The contracts are also getting stricter. Across 16,140 signed agreements, the share of standard cloud contracts that prohibit AI training on customer data rose from under 1% in 2024 to 14% in 2026, and 53% of those signed in the year to June 2026 mention AI at all - Common Paper. Those clauses were written about the vendor's own models, not about resale to a lab, but they show which way customer expectations are moving. A buyer's lawyer will ask how you squared customer content with customer contracts, and "we de-identified it" is not always an answer the contract accepts.

How to apply this: list your largest customer agreements and read their confidentiality, data-use and AI clauses. Where they limit use to performing the contract, exclude that customer's content, or license only what you wrote about your own work. Support tickets and CRM notes are the categories most often caught.

Software vendors' terms: the Slack problem

The second layer is the platforms your records live on. Two separate questions apply, and holders often merge them. The first is whether you can export your own data under your own contract; on almost every platform, an administrator can. The second is whether a third-party app, such as a buyer's connector that pulls data through an API, may use what it pulls to train AI. On a growing list of platforms the answer to the second question is now no.

The best-known example is Slack. Salesforce revised Slack's API terms in 2025 to prohibit bulk export of data accessed through its APIs and its use in large language models - Hunton Andrews Kurth. The current terms say a developer may not "use API Data to train a large language model" - Slack API Terms. Slack's own export tool is a separate route, but full exports of private channels and direct messages need a paid plan and an application approved by the workspace's Primary Owner, and Slack frames them as a compliance tool - Slack Help.

The pattern is consistent: the restrictions bite on API connectors, not on a holder exporting its own data. That matters because several buyers' processes are built around connecting to your tools. It does not make those buyers wrong; it makes the question worth asking. How to apply this: prefer native administrator exports transferred under your own licence, and ask any buyer that wants to connect through APIs how its connector complies with each platform's developer terms, in writing, before you grant access.

Employee messages

Internal email and chat are the most requested records and the most contested. They contain every employee's words, their complaints, their health disclosures, their discussions of pay. In the US, the National Labor Relations Act protects employees' discussions of pay and working conditions, and lawyers flag that as a reason to exclude such material - Legal 500. In the Spirit case, unions for the flight attendants, pilots and mechanics filed objections - docket summary, and more than 100 members of Congress wrote to Google and Spirit asking them to halt the transfer of former employees' information - The Record.

Even the court-appointed privacy ombudsman, who recommended approval, said her review did not cover whether the sale posed privacy risks to employees - Yahoo Finance. Europe is stricter still. The Dutch data protection authority's position is that personal data can be sold out of a bankruptcy only with the prior consent of the people concerned - TvI 2026/7, and German lawyers expect regulators to view a sale of email and chat for AI training as "extremely critical" - Börsen-Zeitung.

How to apply this: exclude HR, payroll, disciplinary, medical and union material by default. micro1's late Spirit bid offered exactly such covenants, excluding disciplinary, medical, equal-employment and union records - docket summary, which shows buyers will accept the exclusion. Decide early whether to tell staff before a sale; finding out from the press is how employee objections become public ones.

Personal data and de-identification

Almost every company archive contains personal data, so almost every sale depends on de-identification, and the standard depends on the law that applies. For anything containing EU personal data, the European Data Protection Board adopted guidelines on anonymisation on 8 July 2026 that set three cumulative tests: no singling out, no linkage and no inference, with negligible re-identification risk - Pearl Cohen. By that standard, a Slack export with names swapped for tokens is still personal data, because writing style and project context identify people. For US health information, HIPAA offers two routes: removing 18 categories of identifiers, or an expert's determination that re-identification risk is "very small" - 45 CFR 164.514.

Re-identification is not hypothetical: one widely cited study found 99.98% of Americans could be re-identified from 15 demographic attributes, a figure the October analysis of micro1's program used to question weak scrubbing - BERI. The buyers' own pages are candid about limits. Grepped says "No de-identification method is perfect" - Grepped security. micro1 removes personal information "when relevant". Handshake says data partners remain responsible for the lawfulness of the transfer.

The Spirit case produced the clearest public model. The estate hired Tonic.ai to de-identify unstructured data, a certification must be filed with the court before each transfer, and the buyer must publicly commit not to re-identify anyone - The Next Web. The Frankfurt Kurnit checklist adds that the holder should select or jointly select the de-identification provider, approve the protocol and require certification. How to apply this: decide who scrubs, to which standard, who pays, and who certifies, and put all four in the contract. If removing the identities would remove the value (a list of who knows whom, a set of named experts' opinions), we do not take the data to market: in most regimes it can be sold lawfully only with the consent of the people in it.

Two more regimes are worth knowing. Several US states now run data-broker registries (California, Oregon, Texas and Vermont, with Connecticut and New Jersey following), and their definitions differ - Husch Blackwell. And the US Justice Department's bulk sensitive data rule restricts transfers of sensitive personal data to countries of concern, including China - US Department of Justice, which matters in a market where Chinese labs are reported to be large buyers.

Third-party content inside your archive

Your inbox and drives hold material that belongs to others: counterparties' confidential documents, licensed research and databases, suppliers' data, and open-source code. In the Spirit sale, the software vendor Springshot claimed it owned service-generated data, and the engine maker IAE cited confidentiality obligations - docket summary. Neither claim was about personal data, and both slowed the sale.

The courts are also drawing lines around licensed content. On 29 to 30 September 2026, the Third Circuit held in Thomson Reuters v. ROSS that copying Westlaw headnotes to train a competing legal AI was not fair use, the first federal appellate ruling on AI training - LawNext. Open-source licences travel with code too: contract claims over training on open-source code continue in Doe v. GitHub after the Ninth Circuit disposed only of a separate copyright-notice theory - Haynes Boone. Rewriting or anonymising someone else's material does not change who holds the rights. How to apply this: exclude third-party confidential folders and licensed databases, and run a licence scan over any code you offer so vendored and open-source files are identified before a buyer finds them.

Promises you made before

The last layer is your own privacy policy and terms. If you told users or customers their data would never be shared or sold, that promise does not expire because the company is closing. The US Federal Trade Commission warned in February 2024 that quietly changing a privacy policy to permit AI training may be unfair or deceptive - National Law Review. The rule that privacy promises travel with data sold in a wind-down goes back to the FTC's 2000 settlement with the bankrupt retailer Toysmart, which allowed its customer list to be sold only as part of the whole website to a qualified buyer - FTC. In Spirit, passenger databases were excluded from the AI sale entirely.

Why this section matters: each of these six layers can end a deal late, after time, samples and goodwill have been spent. How to apply it: build a rights file first, category by category: who created it, which contract governs it, who appears in it and whether they agreed. The rights check walks through those questions. As a matter of policy, AIData.Broker declines to take certain records to market at all: data whose value is the identity of the people in it, recordings of people who never agreed, customer data your contracts reserve for serving that customer, health, children's or financial account data, anything offered by someone who cannot sign for it, and material scraped or copied from others.

5. From inventory to payment: the process and its timeline

Buyer and broker sites describe roughly the same sequence: a free estimate, an intake form, a scoping call under NDA, a rights check, a signed agreement, read-only extraction, de-identification, a sample or acceptance review, and payment. What they describe less clearly is the order in which a holder should do things to protect its position. The buyer's sequence is designed to get a signature; a holder's sequence should be designed to arrive at the first buyer conversation with the rights questions already answered and the description already written.

The published timelines are short, and they are real, but they mostly count from the signed agreement. Mercor pays "typically within 2-4 weeks of the signed agreement" - Mercor. Grepped says most companies go from form to payment in two to three weeks. The slow part is everything before signature, which no buyer quantifies, and which a brokerage that does publish its caveats describes honestly: a transaction "may take weeks or months, or may not close" - FileYield. The steps below are in the order that keeps leverage with the holder.

Before any buyer: inventory, valuation and the rights file

Step 1: Inventory. Write down what exists before anyone asks for a sample. For each system (email, chat, tickets, code, documents, CRM, accounting), note the years covered, who holds administrator access, how it can be exported, and how it links to the others. Buyers' own forms show what they will ask: Grepped's intake asks for headcount, the number of employee inboxes, total data across storage, years of records, where data lives, time to export, and a checkbox attesting that the submitter "can speak for the company's data" - Grepped form. That last item is not a formality. Only someone with authority over the records can license them; an employee without that authority, or a former employee, cannot.

Step 2: Valuation against public prices. Before talking to a buyer, know what the public anchors imply for records like yours, so the first offer is measured against something other than itself. The valuation tool reads your answers against the price board and takes about a minute, with nothing uploaded. Treat the result as a range to test, not a number to defend.

Step 3: The rights file. This is the step no buyer will do for you, and the one where most sales die. Category by category, record who created the material, which contracts govern it (customer agreements, vendor terms, employment and union agreements), who appears in it and whether they agreed, and what you will exclude. It takes days for a simple company and longer where customer contracts are many or the records include communications. The rights check sets out the questions in order.

These three steps share a principle: they produce information that belongs to you and costs nothing to keep confidential. Nothing has left your systems, no sample has been shared, and no buyer has an exclusive conversation with you. If the rights file shows the records cannot be sold, you have lost a few days rather than a few months. If it shows they can, you walk into every conversation knowing the answers to the questions that would otherwise become discounts.

With buyers: the datasheet, the approach and the offers

Step 4: The datasheet. Package the description into one document a buyer's legal and research teams can approve: source, contents and how the systems link, verified size after cleaning, rights, personal data and how it will be handled, format and delivery, the licence offered, and known limits. The research template is "Datasheets for Datasets", which covers motivation, composition, collection and recommended uses - arXiv. Regulation makes the document non-optional for buyers: California's AB 2013 requires generative-AI developers to publish summaries of their training data, including sources, IP status and whether personal information is included - CAIDP, and the EU AI Act has required model providers to publish a training-content summary since August 2025, with fines enforceable from 2 August 2026 - Wilson Sonsini. The datasheet tool builds one from your answers.

Step 5: The approach. Send the datasheet, under confidentiality, to every buyer whose published rules you meet, at the same time. Expect a scoping call: Turing schedules an intake call within one business day and reviews a representative sample usually within a week - Turing; Nyne starts with a 15-minute call. Share samples only after the rights file is done, and prefer buyers that let you approve each package, as Scale does.

Step 6: Offers and terms. Compare offers on more than price. The standard licence shape gives the buyer a perpetual, worldwide right to use models already trained on your data, plus a finite licence to the dataset itself, often limited to a field of use; the holder warrants ownership or the right to license, non-infringement, lawful collection and that no individual opt-outs have been invoked, and indemnifies third-party claims - Terms.Law. That is why "deleted on request" almost always means the dataset copy, not what a model has learned. Price structures run from a flat fee to recurring licences and revenue shares, and the trend is away from flat fees - Troveo.

When offers arrive, put them side by side on the same handful of terms: price and how it is paid, exclusivity and its length, field of use, who de-identifies and to what standard, what happens to the copy when the licence ends, and what warranties and indemnities you are asked to give. A higher headline price with an uncapped indemnity and buyer-controlled scrubbing can be worth less than a lower one with a capped indemnity and an independent certifier. The buyers' own pages hint at the spread of terms on offer: some buy outright, some license and leave ownership with you, some pay once, some pay on each resale.

Transfer and payment

Step 7: Extraction and de-identification. Access should come only after signature, be read-only, and end with the project. Grepped connects read-only after the agreement is signed and revokes access at project end - Grepped security; Ooak deletes the raw archive within 30 days with written confirmation. If a buyer runs the scrub, its standard and the certifier belong in the contract. If you run it, the cost and the certification are yours.

Step 8: Acceptance and payment. Know the trigger. Mercor pays by wire on delivery of the anonymised dataset to the buyer; Nyne pays once the dataset is "delivered and accepted"; Grepped pays within 7 days of data passing review; Corpus Bridge, a broker, describes buyers paying on contract milestones, for example half on signature and half on delivery - Corpus Bridge. Acceptance-based triggers put the risk of a failed review on you, so define acceptance criteria in the contract.

Adding the published steps together, a clean operating-company sale with a short rights file can plausibly run from first description to payment in one to three months, with most of the variance before signature. That is a reading of the published steps, not an audited cycle time, and it assumes nothing goes wrong in the rights review. A public sale takes much longer: Google signed a confidentiality agreement with Spirit in June, won the auction on 14 August, and no sale order had been entered by 11 October after repeated adjournments - docket summary.

Why this matters: the published "two to four weeks" is the fast tail of the process, and the holder controls the slow part. How to apply it: do inventory, valuation and the rights file before any buyer conversation, send one datasheet to every fitting buyer at once, and never share a sample before legal review.

6. Operating, closing or closed: what changes

The same archive can be licensed for years, sold once, or lost entirely, depending on where the company is in its life. An operating company usually licenses its records and keeps both ownership and use of them; an outright sale is mainly a shutdown transaction - Troveo. A company in the middle of closing has the most valuable version of its records (complete, explainable, still in live systems) and the least time. A closed company is racing its software vendors' deletion schedules. A company in insolvency no longer decides at all: a fiduciary does, under a court or a statute.

Status also decides which buyers are open to you. Most published programs take only operating companies. Turing takes operating and winding-down companies, Ooak takes operating and closed companies of 20 or more employees, and SimpleClosure was built for shutdowns. The table summarises the differences before the detail below.

Operating companies: the strongest position, if you use it

An operating company has advantages a closing one does not: the people who made the records can explain them, the systems are live, and there is no deadline forcing the first offer. Scale designs its partnerships for recurring revenue, and micro1's $500K tier is for "ongoing participation across multiple teams" - micro1. Some buyers also approach operating companies through the firms that own them: Mercor tells private-equity firms it can unlock revenue "through operational data monetization", including offsetting the cost of AI agents with data revenue - Mercor enterprise.

The constraints on an operating company are the people who are still there. Current customers can object, current employees can find out, and your own customer contracts are live. The practical approach is to start with light-rights material (procedures, documentation, engineering files, code with tickets), keep communications and customer content for later or never, and treat exclusivity as something to sell for a limited period rather than forever.

Closing companies: the best records and the least time

A company that is closing or being acqui-hired holds its most complete, explainable records, and buyers know it. When the AI sales startup Warmly agreed to be acquired by HubSpot, Mercor emailed its CEO eight days later proposing to buy or license its codebase and task records - AI Weekly. Four offers arrived, up to $300,000, and the CEO declined them all; HubSpot said no data was included in the acquisition - FNNews, citing The Information. Acqui-hires often leave the data behind with the selling company, so the records may still be yours to sell, or not, depending on the purchase agreement.

Who must agree is the next question. Under Delaware's new assignment-for-the-benefit-of-creditors statute, a transfer of all or substantially all assets needs board and shareholder approval, and secured creditors must consent for encumbered assets - Troutman Pepper Locke. Cash is usually short: the median non-AI company closing through SimpleClosure in the first half of 2026 had $13,000 in the bank - SimpleClosure, which is why software subscriptions get cancelled early. And the money from a data sale does not simply go to the founder. Known liabilities come first, then shareholders through the liquidation preferences, and directors can be personally liable if they distribute before liabilities are settled - SimpleClosure distributions guide.

Closed companies: the deletion clock

Once subscriptions lapse, the records start to disappear on each vendor's schedule, and some schedules have no grace period at all. These are the windows as each vendor states them on its own help pages, read on 11 October 2026.

The practical rule follows directly: export before you cancel, or keep the paid plans alive until a buyer has sampled. A few months of subscriptions cost little against a sale of tens or hundreds of thousands of dollars, and a Google Workspace cancelled without an export cannot be recovered at any price. The deletion clock counts the days left for each system from the date you cancelled or plan to, and the estates page covers closed companies and their fiduciaries in more detail.

Companies in insolvency

In a formal insolvency, the trustee, assignee, liquidator or administrator controls the records and has a duty to maximise their value for creditors. In a US bankruptcy, a sale that involves personal information can bring in a consumer privacy ombudsman, and the sale needs the court's approval, which is why Spirit's process ran from June to October without an order. Bankruptcy lawyers now treat internal data as a standard estate asset to look for - Patterson Belknap.

Europe is a slower market for this. UK regulators jointly warned insolvency practitioners in 2020 that selling clients' personal data is likely unlawful, and that standard contract terms are "highly unlikely" to amount to valid consent - FCA. With the Dutch regulator's opt-in position and the German commentary described in section 4, the realistic European starting point is material that is not personal data at all: source code, product documentation, process documents, machine and sensor logs.

Why this matters: status decides the deal shape, the signatories, the buyers who will talk to you and the deadline. How to apply it: if you are operating, use the absence of a deadline to run a proper process; if you are closing, secure exports and board approval before anything else; if you are closed, check the deletion windows today.

7. One buyer or several: what the Spirit auction shows

Most company-data sales happen privately between one holder and one buyer, so their prices and terms are never seen. The Spirit Airlines sale is the exception, because a bankruptcy docket exposes what private deals hide, and it is the clearest evidence available of what competition does to price. Three qualified bidders took part in an auction run by the estate's advisers. Google opened at $5 million. Mercor bid $5.2 million on Google's terms, or $7 million if it could run the de-identification itself with its own tools, and finally $10 million, again on its own tools. Google won at $10 million, all cash, with an independent third-party scrub paid for by the buyer, and Mercor's $7.5 million bid was named the alternate - docket summary.

The bidding did not stop at the gavel. On 3 September, micro1 filed notice of a competing proposal, reported at $12.5 million all cash (the notice itself is described as stating no amount), a 25% premium over the winning bid, proposing covenants that excluded sensitive employee records. Its CEO called Google's price "actually quite low" - RuntimeWire. As of 11 October no sale order had been entered, and the ombudsman wanted to review micro1's in-house de-identification plan if its bid were considered.

Sources: Google's opening bid, Mercor's first bid, Mercor's alternate bid, Mercor's own-scrub bid, Google's winning bid, micro1's late offer.

Three lessons carry over to a private sale. Competition doubled the opening price in about one session, and a fourth party valued the same archive 25% higher again. Non-price terms decided the winner: Mercor matched Google's $10 million and lost, because the estate preferred an independent scrub to a buyer-controlled one. And the first number was not the market: had the estate negotiated only with its first bidder, it would have had no way to know that $5 million was half of what the archive would fetch.

A buyer's own marketing confirms the point. Nyne promises to beat a qualifying written offer for the same data - Nyne, which makes sense only if written rival offers raise its price. The form of the process matters too. Open public auctions of unique intangibles have a poor record: the first patent auction run by ICAP Ocean Tomo raised what IAM called a record low, and most lots in a weak 2009 sale sold in private talks afterwards - IAM. What works is a controlled, confidential process: the same datasheet to every buyer whose rules you meet, at the same time, with a deadline for offers.

Why this matters: a single offer gives you speed and nothing to measure it against, and exclusivity, the largest lever on price, cannot be priced without alternatives. How to apply it: unless one buyer is plainly right and its published terms are acceptable as they stand, run a short competitive process and compare offers on price, exclusivity, scrub control and payment trigger together.

8. Fees: how intermediaries get paid

Between a holder and a lab there are often one or two intermediaries, and the way each is paid tells you whose interests it serves. A buyer's introducer is paid by the buyer for introducing you. A broker may take a fee from one or both sides, or buy from you and resell, keeping the spread. A marketplace takes a cut of each sale it facilitates. A principal buys for its own account and earns its margin on resale. A holder's agent is paid by you, out of what you receive, to get the best outcome. None of these models is wrong in itself. The problem is not knowing which one you are dealing with.

The fine print matters most where the buyer pays the person across the table from you. micro1's referral terms, updated on 8 October 2026, set a fallback referral payout of 7.5% of the price paid to the company where no other payout is specified, leave timing to micro1's "sole discretion", and forfeit pending payouts if the introducer is terminated - micro1 referral terms. In English law, an agent who takes an undisclosed payment from the other side holds it on trust for its principal, the rule confirmed in FHR European Ventures v Cedar Capital - Maitland Chambers. Disclosure is the dividing line.

Agents that sell a one-off, hard-to-price asset they never own are almost always paid by the holder, on success, often on a sliding scale so the percentage falls as the deal grows; M&A advisers call the classic version the Lehman formula - Wikipedia. That is the model AIData.Broker uses, and since this guide is published by us, here it is in full so you can judge it against the alternatives above.

AIData.Broker's terms, stated plainly

Our fee is a success fee on what the holder is actually paid, band by band: 15% of the first $250,000, 10% from $250,000 to $1 million, and 6% above $1 million. Nothing is payable if nothing sells. Only the holder pays us, a fee fixed in the mandate, and we take nothing from any buyer: several buyers publish an introduction fee, and in a sale we run we decline it, show you beside each offer what that buyer pays introducers, and ask it to put that amount into your price instead. So we earn the same whichever buyer you choose, and more only when you are paid more. We do not buy your data or license it in our own name: you deliver it to the buyer, and the buyer pays you. As published, mandates are exclusive for 2 months: during those months our fee is due on any sale of the lot except one you make yourself to a buyer you listed at signing; after them, while the mandate runs and for a tail of 12 months after it ends, only on a sale to a buyer we introduced (in the tail, one on the list we send you in writing). We do not take mandates where the expected sale is below $50,000, because a self-serve route serves those holders better. The fees are on the pricing page; the binding rules are in the mandate, which our terms describe.

Sources: our three fee bands ( 15%, 10% and 6%) applied to each size of sale, as in the fee on a sale of half a million dollars.

A worked example shows how the holder-only rule plays out. On a $300,000 sale our fee is $42,500, and on a $500,000 sale it is $62,500, whichever buyer wins. If the winning buyer publishes an introduction fee of up to $50,000, we decline it, show it to you in writing beside that buyer's offer, and ask the buyer to add it to your price instead. Because our pay does not depend on which buyer wins, nothing in it steers you toward the buyer with the most generous introduction fee, and the choice of buyer is yours.

Why this matters: intermediaries' incentives follow whoever pays them, and the same introduction can be neutral advice or a paid referral. How to apply it: ask every intermediary who pays it, how much, and on what event, and get the answer in writing before you share anything.