What AI labs pay for company data: published deals and honest ranges

What AI labs actually pay for company email, chat and project records, from published deals. Honest ranges by company size, and what moves the number.

Travis Wheeler

By Travis Wheeler

Co-Founder, Commercial and Partnerships

Published · Updated

Short answer. A US company with 50 to 100 office employees and five or more years of email, chat and project records has been paid between $110,000 and $650,000 in the deals buyers have published, with most landing between $275,000 and $450,000. Headline ranges on buyer websites run to $2 million and above, and the published deals that reach those numbers involve hundreds of employees or decades of records.

Figures in this guide come from court filings, press reports and buyer websites. Each one carries its source and the date we read it. None of them is an offer, and none is a Vaquero quote.

How we sorted the numbers

Three kinds of figure appear below, and they are not equal.

  • A published deal is a price attached to a specific transaction: a court docket, a press report that names the seller, or a buyer's own case study with a company type and headcount. A buyer's case study is the buyer's statement. We mark it as unverified unless a second source confirms it.
  • A headline range is the range a buyer prints on its site to describe what it pays. It is marketing, not a close.
  • An estimator output is what a buyer's calculator returns for stated inputs. It is a formula, and most of the formulas use headcount and years only.

Vaquero has no closed deal of its own to report. Every figure here belongs to someone else, and every row links to where it was published.

What has actually been paid for company records?

Most pages on this subject quote ranges. Fewer quote deals. Here is every deal with a published price and a named source that we could find as of November 12, 2026.

Seller Buyer Price What was licensed Source
Spirit Airlines (in bankruptcy) Google $10,000,000 Internal email, chats, documents and spreadsheets, de-identified by a third party at Google's cost. The latest filing describes about 1.9 billion emails across 91,000 accounts, 664 million Teams items and about 30 million lines of code. Mercor was alternate bidder at $7,500,000. Sale hearing set for November 17, 2026; not yet approved. Axios, Aug 17, 2026; auction results, Dkt. 1463, via chapter11cases.com; Skift, Nov 10, 2026
cielo24 (shut down) Undisclosed "Hundreds of thousands of dollars" Every Slack message, internal email and Jira ticket, per former CEO Shanna Johnson Fast Company, Apr 2026
Roofing company Polyshares $450,000 "Perpetual exclusive license to the operational record. Anonymized before delivery." polyshares.com/track-record
General contractor Polyshares $300,000 Same structure polyshares.com/track-record
Solar company Polyshares $275,000 Same structure polyshares.com/track-record
Global marketing agency Polyshares $110,000 Same structure polyshares.com/track-record
100-person freight broker Telegraph Lab $650,000 Rate quotes, shipment documents, dispatch records, delivery exceptions and claims. Buyer case study labeled "Paid to the business." Not independently verified. telegraphlab.com/data-partnerships
70-person engineering firm Telegraph Lab $450,000 CAD drawings, calculations, technical reports, revisions and specifications. Buyer case study. telegraphlab.com/data-partnerships
60-person accounting firm Telegraph Lab $350,000 Five years of workpapers, reconciliations, financial models and tax files. Buyer case study. telegraphlab.com/data-partnerships
340-employee freight and logistics company, Ohio Corpus Data $835,385 Closed example on buyer site. The homepage rounds the same deal to $850,000. Company identity withheld. datacorpus.ai/payouts
760-employee industrial manufacturer, Michigan Corpus Data $1,100,000 Closed example on buyer site datacorpus.ai/payouts
1,180-employee healthcare administration company, Texas Corpus Data $1,400,000 Closed example on buyer site datacorpus.ai/payouts
2,400-employee financial services company, New York Corpus Data $2,000,000 Closed example on buyer site datacorpus.ai/payouts

Three more sets of figures are self-reported by intermediaries rather than tied to a named deal.

Grepped's founder Erik Newsham posted on October 24, 2026 that the broker's average deal had moved from about $250,000 to $400,000 up to roughly $1.1 million to $1.3 million, and that its manufacturer deals paid around $500,000 to $1.2 million for a two-year data license, as reported by FourWeekMBA. The source is the founder's own post.

SimpleClosure, which sells the assets of companies that are shutting down, told Fast Company in April 2026 that it had processed 100 such deals in the past year with payouts from $10,000 to $100,000. By October, MIT Technology Review reported that its Asset Hub had handled over 150 deals and expected to return over $10 million to startups by the end of 2026. The one deal that article priced: a small crypto startup netted roughly $25,000 for its code base, with about $15,000 more pending for workspace data.

Ooak Data says on its partner page that "Deals to date range from $20K, the smallest closed, to several hundred thousand dollars" (ooakdata.com/partner). Data Licensing Desk, which buys documents behind investment decisions rather than operating records, says its smallest partnership was $6,000 and its largest $500,000 (datalicensingdesk.com).

Four things stand out in the table.

  1. The one eight-figure deal involved an airline with records back to the 1980s and a bankruptcy auction with two bidders. As of November 10, 2026 it still awaits court approval. It is not a comparable for a 60-person company.
  2. The four Polyshares deals, the only buyer-published closes with company types and a stated license structure attached, sit between $110,000 and $450,000.
  3. Telegraph Lab's three case studies and the three deals on Pewter Data's deals page are the same three deals. Pewter labels them "reported by our licensing partner" and says it did not advise on them. A reader who sees "$450K engineering firm" on five pages is looking at one deal.
  4. Pewter's deals page also lists $1.2 million for an industrial distributor and $500,000 for a third-party logistics company, attributed to an unnamed licensing firm. We could not trace those to a buyer's own page, so they are not in the table.

What do buyers say they pay?

Buyer websites publish ranges, tiers or calculator outputs. These are marketing numbers, not closes, but they show where each buyer wants to play. All read November 12, 2026.

Buyer Published range Basis Source
Polyshares "Typical deals run $100K to $2M and above" Range polyshares.com
micro1 "$100K-500K" for qualified enterprise data partnerships; "$500K-3M" for large-scale operational datasets; "$3M+" for highly unique proprietary data. Floor of 30 or more full-time employees. Tiers micro1.ai/data-partnerships
Handshake AI "Earn $100K to $4M+"; its calculator at default inputs says "up to $100,000" Range joinhandshake.com/ai/data-partnerships
Telegraph Lab "$100K–$4M" headline; estimator returns "$385K – $895K" for 100 employees, 10 years, USA Headline and estimator default telegraphlab.com/data-partnerships
Pewter Data Same "$385K to $895K" for 100 employees and 10 years, labeled "A scenario, not a valuation"; "Deals published in 2026 ran from about $110,000 to $1.2 million" Estimator default pewterdata.com
Replay (Sunsets HQ Corp.) "$10K - $100K" for 20 to 50 employees; "$100K - $1M" for 50 to 250; "$1,000,000+" for 250 and up. Paid for historical data first, then a perpetual revenue share on each license. Tiers by headcount replay.ai
SourceX "$50K–$300K for 50–99 employees"; "up to $4M is an upper bound for the largest enterprise datasets, not a typical or guaranteed outcome" Indicative by size sourcex.si/faq; sourcex.si
Corpus Data Typical close $350K for 25 to 100 employees; $800K for 100 to 500; $1.4M for 500 to 2,000; $2M+ above that. "Engagements start at $100,000 and run up to $2 million." Typical closes by band datacorpus.ai
Data Licensing Desk (Longitude Labs) "$25k–$500k" per partnership; smallest $6,000, largest $500,000; "c. $1m paid out to partners in August and September 2026 alone" Deal history datalicensingdesk.com
Labelbox Prism Smaller datasets "typically start around $50K"; large enterprise partnerships "can reach $1M+"; upfront fee at signing plus revenue share Range prism.labelbox.com
Scale AI "$10K–$1M+ illustrative value per data partnership, scaling with cadence" Illustrative scale.com/data-partnership
Actual Data "$285,896 to $604,463" homepage example, "Indicative estimate only." Inputs were 50 employees, 10 years, United States when we read the live form on November 10, 2026. Estimator actualdata.ai
Montlake Licensing "$500K–$2.6M" headline; example card "$650K to $970K in total"; "not an offer" Headline montlakelicensing.com
Korra "Your business records could be worth $200K to $2M" Headline korra.io
Mode "Your company's data could be worth $100K–$5M" Headline data.mode.inc
Ooak Data "Earn up to $1M"; "Deals to date range from $20K, the smallest closed, to several hundred thousand dollars" Headline and deal history ooakdata.com/partner
Nyne "$38,697-$90,608" for 100 employees, 10 years, USA; "indicative and not an offer" Estimator default nyne.ai/data-partnerships
Sanctus Data Labs "A license pays $100K to $3M+" Range sanctusdatalabs.com
Troveo "$850K–$925K" estimated first-year net earnings given four or more licensing deals, "Illustrative only" Estimator sample troveo.ai/business-data-assessment

Every published estimator we have tested reduces to the same three inputs: office headcount, years of records, and country. Industry rarely moves the output. The same inputs produce very different numbers by buyer. Telegraph Lab returns $385,000 to $895,000 for 100 employees and 10 years. Nyne returns $38,697 to $90,608 for the same two inputs. The calculators describe the buyer's budget, not your records.

Quoted versus realized: why the headline and the check differ

The honest comparison is a buyer's own headline against its own published closes.

Buyer Headline Published closes
Polyshares $100K to $2M and above $110K, $275K, $300K, $450K
Telegraph Lab $100K to $4M $350K, $450K, $650K
Corpus Data $100K to $2M $835K (340 employees), $1.1M (760), $1.4M (1,180), $2.0M (2,400)
Data Licensing Desk $25K to $500K $6,000 smallest, $500,000 largest
Ooak Data Up to $1M $20K smallest, "several hundred thousand" at the top
SimpleClosure Not published $10K to $100K, per Fast Company, April 2026

The pattern is consistent. Published closes cluster in the lower third of the headline range, and the deals above $800,000 involve companies with several hundred employees. A 50-person company reading "$100K to $2M" should plan around the first number, not the second.

Two other reasons the check comes in below the headline:

  • Scope shrinks in diligence. Buyers price the records they can legally take. Client-owned deliverables, records under NDA, and anything with personal health or payment data come out before the final number. The Spirit sale shows this at scale: the parties pulled the passenger databases, refund records, inflight purchase and Wi-Fi systems out of the deal, and the unions won exclusion of medical, payroll, drug-testing and biometric records. The price stayed at $10 million, but the scope that price buys is smaller than the original schedule. A company that starts with eight years of everything often licenses five years of some of it.
  • Exclusivity is priced in. The larger headline figures usually assume the buyer gets an exclusive license. Keep the right to license elsewhere and the per-deal number drops, though the total can rise. More on that below.

What moves the number up or down?

Buyers evaluate the same handful of factors. The table shows the direction each one pushes, drawn from the published estimators, buyer statements, and our conversations with buyers.

Factor Pushes the number up Pushes it down
Office headcount 100 to 500 desk employees is the sweet spot for most buyers. Over 500 adds procurement friction but not much price. Under 20 desk employees is below most buyers' floors. Shop floor, field and warehouse staff do not count.
Years of records 10 or more years of continuous history. Under 3 years. A tool migration that dropped history, such as a Slack to Discord move, cuts the value of everything before it.
Richness Five or more kinds of record with history: email, chat, project files, estimates and change orders, drawings or CAD with revisions, support tickets with narrative, code. One or two kinds. Ticket stubs with open and close dates but no narrative are discounted by every buyer.
Ownership Records about your own products, projects and decisions. Manufacturers, distributors, contractors, software product companies. Records that belong to clients. Agencies, consultancies, job shops working to customer prints, law firms.
Demand this quarter Buyers rotate by lab project. In late 2026 the named asks are mechanical and circuit-board CAD, architecture and BIM, heavy construction, and 3D game assets. Marketing agencies, staffing, call centers and education vendors are excluded by several buyers outright.
Operating or closed An operating company with a reachable owner. A closed company adds chain-of-title work and removes the buyers that only purchase from operating companies. Some buyers run separate programs for closed companies.
Country and language US-based team, records mainly in English. SourceX asks that at least 70 percent of the data be in English for companies outside its priority regions. Records mostly outside the US or in another language. Published estimators apply a discount when the country input changes from USA.
Deal structure Exclusive license, termed. Non-exclusive, or exclusive in perpetuity, which some buyers ask for and which blocks every future deal.

What buyers describe as valuable, in their own words, is "how work actually happened": the decision, the discussion, the approval, the revision, the outcome, traceable across systems over years. Volume without that thread does not price well.

What makes company data worth nothing?

Some records that look like assets price at zero. The common cases:

  • Ticket stubs. A helpdesk export with opened and closed timestamps and no written exchange. One buyer's own solar company had exactly this and it had no market.
  • Client-owned work product. An agency's deliverables, a consultancy's reports, a job shop's work to customer drawings. The client owns it, so you cannot license it. Internal playbooks and SOPs survive, the client work does not.
  • Records under privilege or regulation. Patient data under HIPAA, cardholder data under PCI, attorney-client communications, records of minors, biometric data. Buyers exclude these and will not pay you to scrub them.
  • Export-controlled technical data. ITAR and EAR drawings and specifications stay in the building.
  • A migration that dropped history. Years of records that were deleted when the company changed tools are gone, and a buyer prices what exists today.
  • Privacy promises you already made. A privacy policy that says "we will never sell your data" cannot be amended retroactively for records already collected. The FTC has said so since 2024.

If most of what you hold falls in one of these buckets, a sale now is unlikely. The useful move is to fix the contracts and policies going forward so the next three years of records can be licensed later.

How does exclusivity change the economics?

Buyers prefer exclusive licenses and several make them the default. Polyshares' published case studies describe a perpetual exclusive license. SourceX says "Licenses are exclusive for AI training," with the term set in each Dataset Deal Confirmation. Labelbox Prism says licenses "are structured as exclusive by default." Telegraph Lab says "Any exclusivity must be explicitly agreed."

The trade works like this:

  • Exclusive, termed. The buyer pays more per deal, typically the upper part of its range, and you agree not to license the same records to anyone else for a fixed period. Two years is the term Grepped's founder described for its manufacturer deals. After the term you can license again.
  • Exclusive, perpetual. You can never license those records again. This also becomes a diligence item if you ever sell the company, because an acquirer inherits the restriction. Our view is that perpetual exclusivity should be refused or priced well above the termed number.
  • Non-exclusive. Lower per deal. Some sellers have stacked several non-exclusive licenses. Troveo illustrates this on troveo.ai/business-data: a first sale of $200,000 to a frontier AI lab, then $250,000 from a second buyer, $200,000 from an AI startup and $200,000 from a large tech company, for $850,000 in total. Troveo labels it an illustrative example with "actual terms scoped per partnership." It is not a deal record.

A reasonable rule: perpetual use of the files the buyer already has is normal and cannot be clawed back. Exclusivity is what you sell, and it should have an end date.

How does a broker change what you net?

Going direct to one buyer gives you one quote and that buyer's standard paper. A seller-side broker shops the same records to several buyers and handles the contract review. In Vaquero's case the seller pays nothing. The buyer pays our fee when the deal closes, and the arrangement is disclosed to you in the transaction documents before you sign.

Here is how the two routes compare on a $300,000 deal, which is close to the midpoint of the Polyshares published closes.

Route License fee What you pay Net to seller What you get
Direct to one buyer $300,000 $0 $300,000 One quote, buyer's standard terms, you handle contract review
Through Vaquero $300,000 $0 $300,000 Quotes from several buyers side by side, counsel review of your client contracts, termed exclusivity negotiated, payment tracked to your account

The broker route pays for itself when a second quote comes in higher than the first, or when a contract review removes a term you would have regretted. It does not pay for itself when only one buyer wants your records and the terms are already clean.

Find out where your company lands

The estimator on vaquerodata.ai takes office headcount at peak and years of records and returns a range from published deals. It takes two minutes, asks for no files, and a person replies within one business day with a view on whether your records are the kind buyers are paying for right now.

Estimates are from published deals and buyer-published ranges. They are not an offer, and buyer quotes can be higher, lower or declined. Nothing here is legal or tax advice.

Sources

All pages read November 12, 2026 unless noted.

Questions and answers

How much is my company's data worth to AI companies?

For a US company with 50 to 100 office employees and five or more years of records, published deals sit between $110,000 and $650,000, with most between $275,000 and $450,000. Smaller companies, companies with fewer years, and companies whose records belong to clients price lower or not at all. Larger companies with decades of records have reached seven figures.

Do AI labs buy data from small businesses?

Some buyers set floors. micro1 asks for 30 or more full-time employees for operating companies. SourceX says it is best suited to companies with 50 or more full-time employees at peak. A few buy from companies of any size if the records are deep. Under 20 office employees, expect fewer bidders and a lower number.

Is the money a sale or a license?

A license. You keep your records and your ownership. The buyer gets the right to use a de-identified copy for the purposes and term in the agreement. Perpetual use of that copy is standard. Exclusivity is negotiated.

Who pays to remove names and personal information?

In the published programs, the buyer does. Polyshares says identifying fields are stripped by a specialist PII partner before Polyshares or any lab sees the record. In the Spirit Airlines sale, Google agreed to cover the cost of de-identification by a third party. Confirm in writing who does it, when, and who carries the liability if something gets through.

How long until payment?

Polyshares publishes net 30 to 60 once the data is approved, anonymized and delivered, with an offer within one to two weeks of intake and terms papered right after. SourceX publishes payment within 60 calendar days of a valid invoice. Other buyers do not publish timelines. From first contact to cash, plan on six to twelve weeks.

Can I license the same records to more than one buyer?

Only if the license is non-exclusive or the exclusivity period has ended. Several buyers ask for exclusivity by default. Read for "sell to us first" clauses, which work as exclusivity without the word.

Are closed companies paid less?

Often. The buyers that purchase from closed companies are a smaller set, and someone has to establish who holds the rights. SimpleClosure told Fast Company in April 2026 that payouts on its wind-down data deals ranged from $10,000 to $100,000. Closed companies with large peak headcounts and clean rights have done much better, as the Spirit Airlines auction shows.

Is the payment taxable?

Yes. License income to an operating company is ordinary income in most cases. Treatment for a dissolved entity or for a sale rather than a license differs. Ask your CPA before signing.

What could your records be worth?

See my range