Know a company with years of records? Refer it and earn up to $50K
Reference map

Who buys which kind of company data for AI training (2026)

AI training data buyers by data type: who buys chat, code, call recordings, CAD, claims, imaging and video, with published price ranges and deal examples.

Largest public bid
$10MGoogle's winning bid for Spirit Airlines' records; not yet approved by the court as of November 6, 2026
Highest published company range
$100K to $1.5MMercor, manufacturing and industrial
Program headline ceilings
$3M+ to $4M+micro1, Handshake AI, Endless
Per-record health price
$100Tempus 8-K, Oct 2026
Data types mapped
19
Three folders side by side with a ruler laid across them
Short answer

AI training data buyers fall into three groups: general programs such as micro1, Handshake AI, Mercor and Labelbox Prism that buy many kinds of company records; specialists that buy one type, such as imaging, driving video or screen recordings; and AI labs that buy directly in rare large deals. Published company ranges run from about $10K for small R&D archives to $1.5M for manufacturing records, with headline ceilings of $3M to $4M, and the largest public deal is Google's $10M winning bid for Spirit Airlines' records, not yet approved by the court as of November 6, 2026. The catch: most figures are ranges or ceilings set by the buyers themselves, and records with personal data, client confidences or protected health information must be cleared or removed first.

  • Linked operating records from regulated, physical and engineering work carry the highest published ranges: manufacturing up to $1.5M, CAD up to $1.3M, law firms up to $1.2M (Mercor listings).
  • Video, dashcam and screen recordings are bought by the hour or in bulk; published rates are much lower per unit.
  • Healthcare has the only public per-record price: $100 per de-identified record in a Tempus filing.
  • Insurance claims and mortgage files have buyers but no public price yet.
  • Customer-facing personal data is what buyers leave out: Google's Spirit bid excluded passenger profiles and customer chat and call recordings.

Who buys which kind of company data? The map

Each row is one kind of company data, the buyers or buyer types that publicly ask for it, the published price signals with their source, what makes it valuable, and Origin's guide to it. Mercor ranges are its published company ranges, read on its listings page on November 6, 2026 (data.mercor.com/refer). Every other figure links to its source. A range is what a buyer publishes, not what a given company will be offered. See what your data is worth for a figure based on your own records.

Data typeWho buysPublished price signals (source)What makes it valuableOrigin guide
Chat, email and documentsmicro1, Handshake AI, Mercor, Labelbox Prism, Polyshares; Google and Mercor bid in a bankruptcy auctionGoogle's $10M winning bid for Spirit Airlines' data, including about 100M emails and 500M Teams chats, not yet approved by the court as of November 6, 2026 (Axios); micro1 $100K to $3M+ per approved package (micro1); Handshake AI $100K to $4M+ (Handshake)Shows how real work and decisions happen across a company, years deepSlack and email archives
Support ticketsmicro1, Labelbox Prism, Endless, PolysharesNo ticket-only price published. micro1 lists support tickets in its $100K to $3M+ track (micro1); Prism connects to Zendesk and ServiceNow, with targeted datasets from around $50K to $1M+ (Prism)A problem, the steps taken and the outcome in one threadSupport tickets
Source codeMercor, SimpleClosure buyers (foundation labs, workspace-simulation vendors)Mercor $50K to $500K for private code, 15 to 500 staff; a closed crypto startup netted about $25,000 for its code base (MIT Technology Review)Private code with its review and change historySource code
Call recordingsMercor (call centers and BPOs)Mercor $20K to $250K for 50 to 5,000 agents. Google's Spirit bid left out the customer chat and call recordings, per a reading of the asset schedule (BenchmarkList)Multilingual conversations linked to case outcomesCall recordings
Field-service work ordersEndless, Handshake AI (lists ServiceTitan as a source), PolysharesEndless "up to $4M" ceiling for quotes, tickets and work orders, as a fixed, non-exclusive payment (Endless)Job, diagnosis, parts and fix tied togetherField-service records
Manufacturing recordsMercor, micro1, PolysharesMercor $100K to $1.5M for production, maintenance and inspection records, 20 to 2,000 staff; Mercor names work orders, QA logs and spec sheets (Mercor)Process steps linked to measured quality and failuresManufacturers
Shop-floor and first-person videoRobotics data programs that pay businesses: Shift, Amelia and othersShift up to $20 per hour of accepted footage (Shift); Amelia $500 to $2,000 a month per recording employee (Amelia)Skilled hands-on work that cannot be filmed from outsideBusiness video programs compared
CAD with revision historyMercor (hardware engineering), Handshake AI (lists AutoCAD)Mercor $45K to $1.3M for CAD assemblies, circuit designs and test results, 20 to 2,000 staffNative design history and engineering changes, rarely publicCAD and engineering data
Construction project recordsMercor; Polyshares (lists construction)Mercor $115K to $1M for project records, drawings and site imagery from completed projects, 50 to 2,000 staffEach job ends with a checkable cost and outcomeConstruction companies
Law firm work productMercor; AI-native law firmsMercor $50K to $1.2M for complete files from closed matters and deals; two AI-native law firms bid for Weil Gotshal's data (Law.com)Templates, playbooks and redlines that show legal judgmentLaw firm data ethics
Accounting workpapersMercor; Endless (excludes client books)Mercor $25K to $865K for tax returns, workpapers and bookkeeping records, 5 to 200 staff. Client confidentiality limits what can go (Journal of Accountancy)Repeatable judgment work across many kinds of entityAccounting firms
Mortgage and loan filesMercor (finance: underwriting files, credit files), Polyshares (financial services)No mortgage-specific price published. Nearest Mercor ranges: investment firms (closed deals, credit files) $50K to $800K; payments and fintech $80K to $900KUnderwriting decisions with conditions and outcomesMortgage lenders
Insurance claimsMercor (lists underwriting files and claims) (Mercor); Endless (agency workflows incl. claims assistance)No public price foundClaim, adjuster notes, reserve and payout in one fileInsurance claims data
Healthcare records and imagingSegmed, Gradient Health, Dandelion, OneMedNet, Protege, Mercor$100 per de-identified neurology record paid to Tempus (8-K); Mercor $80K to $820K (practices) and $80K to $620K (imaging centers); Gradient example: $240K partner share of a $3M deal (Gradient)Longitudinal, multimodal records with signed reportsHealth data programs compared
Aviation maintenance recordsGoogle, Mercor (Spirit auction bidders)Spirit's $10M winning bid covered MRO records from TRAX, listed at 459,828 scheduled tasks and 787,452 parts received, per a transcription of the filed asset schedule (BenchmarkList); Mercor bid $7.5M (Axios); the sale hearing moved to November 17, 2026Decades of fault-to-fix records under strict rulesAircraft maintenance records
Fleet dashcam videoWorld-model and autonomous-driving labsA six-figure agreement for 3,000 hours of driving data to a world-model lab (NATIX)Varied real roads, weather and driver behaviorDashcam footage
Desk-work screen recordingsFrontier labs building computer-use agentsMarkov says it has sold 15,000+ hours to frontier AI labs; no price published (Y Combinator)Screen, mouse and keyboard actions from expert software workScreen recordings
R&D lab recordsMercor (biotech, semiconductor, chemistry and materials)Mercor $50K to $875K (biotech labs), $100K to $750K (semiconductor and materials), $10K to $300K (chemistry and materials R&D records)Failed experiments and run records that never get publishedR&D experiment records
Aerial, LiDAR and 3DSimulation and digital-twin companies (Zeromatter, Neara)Two off-the-shelf licences worth a combined $1.1M to Aerometrex, an ASX-listed survey firm (Grafa)Calibrated, dated archives that are costly to re-flyAerial and LiDAR data

Source: Mercor Data: Refer a company (25 listings with company ranges)

Which companies buy data for AI training across many data types?

Most company data is bought by a handful of programs that take many record types and resell or license them to AI labs. Their published terms differ more than their headlines.

ProgramWhat it says it buysPublished amountKey termsSource
micro1Internal communications, SOPs, project histories, CRM records, support tickets$100K to $3M+ per approved data package; 30+ employeesEnterprise and defunct-company tracksmicro1
Handshake AISources such as AutoCAD, Gmail, NetSuite, QuickBooks, ServiceTitan and Slack$100K to $4M+ payouts; 20+ full-time employees and 3+ yearsIts calculator says estimates are indicative and not an offerHandshake AI
MercorContracts, clinical notes, claims, work orders, shipping manifests and more25 listings from $10K to $1.5M by industryPays by wire on delivery to the buyer, typically within 2 to 4 weeks of signingMercor
Labelbox PrismSalesforce, Jira, SAP, ServiceNow, Slack, Zendesk, GitHub dataFrom around $50K for targeted datasets to $1M+Exclusive by default; upfront fee plus a revenue sharePrism
PolysharesOperational data from any sector$100K to $2M and aboveSays it buys directly; net 30 to 60 after approvalPolyshares
EndlessQuotes, tickets, work orders, SOPs, handover notesUp to $4MFixed sum agreed upfront; non-exclusive; pools many companiesEndless

Source: micro1: data partnership tracks and company compensation

What data is most valuable for AI?

Published ranges point the same way: records that tie a decision to its outcome, from industries AI cannot see from the public web. Manufacturing ($100K to $1.5M), hardware engineering ($45K to $1.3M), law ($50K to $1.2M) and construction ($115K to $1M) carry Mercor's highest ceilings. Call-center recordings ($20K to $250K) and private code ($50K to $500K) sit lower.

Buyers say depth beats volume. Polyshares says depth of operating history and how much reasoning the record shows move the number far more than raw volume (Polyshares). Labelbox Prism lists human decisions, escalation paths, expert corrections and edge cases as the highest-value data (Prism). Surge AI told MIT Technology Review that one company's Slack history, if it captures varied, deep conversations, can be worth more than another's Slack, email and calendars combined (MIT Technology Review).

Video and screen recordings are valuable in bulk, but each hour is cheap. They work best as an addition to a records licence, not instead of one.

Source: Labelbox Prism: enterprise data licensing

How much do AI companies pay for data? Public deals and prices

Few deals are public. These are the ones with a stated amount, as reported or filed.

DealDataAmountSource
Google wins Spirit Airlines' bankruptcy auctionAbout 100M emails and 500M Teams chats plus documents, financial databases and operating records, de-identified$10M winning bid; Mercor next at $7.5M; not approved as of November 6, 2026; the debtors moved the sale hearing to November 17, 2026Axios; Chapter 11 Cases; Dkt. 1779
Tempus licenses neurology records to PathosUp to 500,000 de-identified patient records$100 per delivered record; related-party dealSEC EDGAR
Aerometrex licences 3D and LiDAR dataOff-the-shelf survey datasets to Zeromatter and Neara$1.1M combinedGrafa
NATIX sells driving footage3,000 hours to a world-model labSix figuresNATIX
OneMedNet supplies a foundation model companyMillions of de-identified clinical records$11.5M+ initial contract valueOneMedNet
A closed crypto startup sells its codeCode base, via SimpleClosureAbout $25,000MIT Technology Review

Source: Axios: Google wins bankruptcy auction for Spirit Airlines data (Aug 2026)

What data do AI companies not buy?

Buyers want the work, not the people in it. In the Spirit auction, Google bid for operational data but declined everything customer-facing, including customer profiles, loyalty data, surveys, and customer chat and call recordings, according to a reading of the filed asset schedule (BenchmarkList). Google said any data it receives is scrubbed of personal information by a third party first (Axios).

Programs set the same limits. Endless says it never licenses customer, employee or client personal data, protected health information, privileged client matter content, client books or proprietary pricing logic (Endless). For accounting firms, the Journal of Accountancy warns that client confidentiality obligations "do not disappear because the data is redacted" (Journal of Accountancy). In healthcare, identifiable patient data cannot be sold without authorization.

See what company data you can sell to AI and how to de-identify company data.

Source: Endless: license operational data to AI labs

Are these prices real? How to read published ranges

Most figures here are set by the buyers: ranges, ceilings and "up to" rates. Handshake AI says its estimate "is an indicative estimate only and is not an offer to purchase your data" (Handshake AI). Shift says its advertised rates are potential maximum earnings and are not guaranteed (Shift).

Terms matter as much as the number. Labelbox Prism licenses are exclusive by default, while Endless licenses are non-exclusive. Mercor pays on delivery to the buyer, micro1 within 30 days of invoicing after a review of up to 30 days, Polyshares net 30 to 60. Compare like with like before choosing.

Read exclusive vs non-exclusive licenses and data licensing pricing models before you sign.

Source: Handshake AI: get paid for your company data

How to use this map for your company

Find the rows that match what your company holds, then write down the systems, years of history and approximate volume for each. That short list is enough for a first conversation with any buyer, and it shows which programs fit. If your company is not a single type, the most valuable package is usually the linked set: records plus the video, photos or recordings of the same work.

Origin Data Partners is an independent data partner that helps companies get their data bought. We map your records against the buyers above and compare offers on price, exclusivity and payment terms. See what your data is worth.

Origin Data Partners works with companies that own data and helps get it bought. We are independent of the services on this page, and a buyer may pay us when a deal closes (how Origin is paid). See what your data is worth.

Common questions

Who buys company data for AI training?

Three groups: general programs such as micro1, Handshake AI, Mercor, Labelbox Prism and Polyshares that buy many record types; specialists that buy one type, such as Segmed and Gradient Health for medical imaging or world-model labs for driving video; and AI labs that occasionally buy directly, as Google did with its $10M winning bid for Spirit Airlines' records, not yet approved by the court as of November 6, 2026.

What data do AI companies buy?

Operational records that show how work gets done: chat, email and documents, support tickets, source code, call recordings, work orders, manufacturing and construction records, CAD files, legal and accounting work product, de-identified health records, and video or screen recordings of skilled work. They generally leave out customer personal data, client confidences and identifiable health information.

Which companies buy data for AI training?

Companies that publish terms include micro1 ($100K to $3M+ per approved package), Handshake AI ($100K to $4M+), Mercor (25 industry listings from $10K to $1.5M), Labelbox Prism (from about $50K to $1M+), Polyshares ($100K to $2M+) and Endless (up to $4M). Health data goes to Segmed, Gradient Health, Dandelion Health, OneMedNet and Protege on a revenue share.

What data is most valuable for AI?

Records that link a decision to its outcome in industries AI cannot see from the public web. Mercor's highest published ranges are for manufacturing ($100K to $1.5M), hardware engineering and CAD ($45K to $1.3M), law firms ($50K to $1.2M) and construction ($115K to $1M). Buyers say depth of history and visible reasoning matter more than volume.

How much do AI companies pay for training data?

Published company ranges run from about $10K to $1.5M, with headline ceilings of $3M+ to $4M+. Public deals include Google's $10M winning bid for Spirit Airlines' records (not yet approved by the court as of November 6, 2026; hearing moved to November 17), $100 per de-identified patient record in a Tempus filing, and a six-figure order for 3,000 hours of driving video. Business video programs mostly publish "up to" rates, such as Shift's up to $20 per accepted hour.

Do AI companies buy customer data?

Mostly not in identifiable form. Google's Spirit bid excluded passenger profiles, loyalty data and customer chat and call recordings, and programs such as Endless say they never license customer, employee or client personal data. Records are de-identified before delivery, and some data, such as identifiable health records, cannot be sold without consent.

Can a small business sell data to AI companies?

Sometimes. Many programs set minimums: micro1 asks for 30+ employees and Handshake AI for 20+ full-time employees and 3+ years, while Mercor lists some firm types from 5 staff and imaging centers of any size. Video programs such as Shift and Amelia pay businesses for accepted footage of their work, with no headcount minimum stated on the pages we read.

Is there a list of AI training data buyers by industry?

Yes: the map on this page lists 19 data types with the buyers that publicly ask for each, their published price signals and sources. Mercor also publishes 25 listings by company type, each with a company range, on its referral page.

Sources and further reading

Origin Data Partners is independent and not affiliated with any buyer named on this page.

What is your company’s data worth?

AI companies pay $20K to $5M for company records. Six questions, no files.