Data Inventory for an AI Data Sale: What You Have, Where It Lives, How Much

A data inventory for an AI data sale is a list of every source of company data a buyer might license, recording what each holds, where it sits, the date range and how much, counted in the unit a buyer uses. The Spirit Airlines asset schedule filed in bankruptcy court is a public example: it counts 100,000,000 emails and 500,000,000 Teams items, and the sale is not approved as of November 9, 2026.

The inventory is the first working document in AI training data licensing for a company holding data. It stays internal, it is built before any buyer sees a sample, and it can end in a decision not to license anything.

Fact-check note: every figure on this page is dated and sourced as described in how we source, date and check every figure.

What a Data Inventory Records Before Any AI Buyer Sees the Data

The inventory answers three questions per source, what it is, where it lives and how much, and adds the facts that decide whether the source can go further: who administers it, which contracts cover it and whether people appear in it.

A data catalog, a data map and a data dictionary are different documents. A catalog helps staff search and understand data, a data map shows how data flows between systems for compliance work, and a dictionary defines the fields inside one dataset. The governance sense of "data inventory" serves compliance records; the sense used here serves a licensing decision.

The inventory is internal and complete, while a description sent to a buyer is outward and selective. That outward description is the data manifest for AI buyers.

The Inventory Row, Field by Field

One row per source with ten fields is enough to answer every intake question in this guide. The table lists each field with an example and the reason buyers ask for it (as of 2026-11-10).

The status field takes one of three values: include, exclude or hold, and a hold row stays out of any offer until its open question is answered.

What You Have: Sorting Company Records by Data Type

Sort records by what they are before where they sit, because buyers price data types, not systems. Six families cover most companies:

• Communications: email, chat, calls and calendar entries.
• Documents and wikis: files, knowledge bases and handover notes.
• Tickets and project histories: IT, support and engineering trackers.
• Code: repositories, commits and pull requests.
• CRM, sales and support records.
• Finance, operations and industry records: billing, work orders, patient or matter files.

Buyer programs name these families in their own words, each statement checked 2026-11-09. micro1 states it wants engineering documentation, support tickets, CRM records and internal communications; Nyne states chat, email and calendar, documents, and CRM, HR and support systems; Turing states chat, tickets, wikis, source code and legacy stacks such as Lotus Notes, Novell and file shares. HR and payroll lines get a hold flag, since the Spirit schedule counts payroll records among the data offered.

Sorting detail, and what each type sells for, sits with data types AI companies buy.

Where It Lives: Source Systems, Administrators and Date Ranges

Each source belongs to a system with a plan tier, an administrator who can count and export it, and a retention rule that sets how far back the records go. The administrator is often one person, and the plan decides whether private channels, archives or old tickets can be exported at all.

Retention can shorten the range. Slack deletes history older than one year in Free workspaces since 2024-08-26 (Slack help, checked 2026-11-09), so the date range that exists can be shorter than the company's history. Industry systems appear in buyer listings: Mercor names Epic and athenahealth for medical practices and QuickBooks, Xero and CCH for accounting firms (checked 2026-11-09), and the Spirit schedule names UKG, Kronos and ServiceNow. SimpleClosure, a shutdown service, states that 95 % of the workspaces it offered include Slack and 82 % include GitHub (article of 2026-09-03, checked 2026-11-09).

Plans, export routes, API terms and vendor deadlines are covered under export rights and mechanics by source system, and this page only records the result per row.

How Much: Counting Volume in the Units AI Buyers Use

Volume means nothing until it is stated in the unit a buyer counts, and the unit changes with the data type.

Units of Count by Data Type

Communications are counted in messages and accounts, code in repositories, commits and lines, recordings in hours, images in items and structured data in records or gigabytes. The table gives one public example per type, from a court filing, a filing with the SEC or a buyer's own page (as of 2026-11-10).

Prices appear here only as evidence of the unit: Pangea.ai states USD 100 per million tokens of source code (checked 2026-11-09), and the Tempus order form of 2026-10-22 prices clinical data per delivered record. The size of a model's training set is stated in other units again, as when the Grok 4.7 summary of SpaceXAI LLC (formerly xAI LLC) of 2026-09-21 gives text in tokens and audio and video in hours.

Published Volume Floors by Data Type

The table shows twelve of Mercor's 25 referral listings that set a minimum volume, each Mercor's own statement, checked November 9, 2026. The table lists them alphabetically with the other conditions in the same listing and what Mercor says its deal range depends on (as of 2026-11-09); the ranges themselves are not shown here.

A floor is one buyer's entry rule, not a market standard, and listings whose floor is a headcount (Construction Company, Game Studio) or years alone (such as Investment Firm) are not shown.

Gross Volume vs Licensable Volume

The count that matters is what is left after exclusions, duplicates and third-party material are removed, and it is often far below the raw count. The Spirit schedule shows the gap in one line: 97,500,000 passenger profiles are marked Not Included, and the ombudsman report adds systems excluded in any form.

Code shows it too. SimpleClosure states that roughly 37 % of raw code is removed in cleaning (article of 2026-09-03, checked 2026-11-09), and Mercor states that forks, mostly AI-written code and single-author repositories are not a fit. Records that barred sources contribute are subtracted from the count; see company data that cannot be licensed for AI training. The inventory therefore carries two numbers per row, gross and licensable, with the gap explained.

A Court-Filed Inventory: How the Spirit Airlines Asset Schedule Is Built

The asset schedule attached to the Spirit Airlines auction results is a public example of a data inventory: one line per source, a unit, a count and a column saying Included or Not Included. The table reproduces it (Bankr. S.D.N.Y. 25-11897 (SHL), Dkt. 1463 Exh. A-2, filed 2026-08-14; sale not approved as of 2026-11-09, hearing 2026-11-17, Dkt. 1779).

Consumer transaction systems (transactions, PNR), refunds, inflight purchases and Wi-Fi sales are excluded in any form (ombudsman report, Dkt. 1684, 2026-10-05).

Three lessons carry over to any company: give a count per system, name the unit beside the count, and mark what stays out of the sale.

What Buyers and Intermediaries Ask About Your Inventory at Intake

Most programs ask for an inventory in all but name within the first call, before any sample or contract. The table lists seven parties alphabetically with what each asks for first; each cell is the named party's own statement, checked 2026-11-09 (as of 2026-11-09). Brokers place data with buyers without acquiring it; Endless licenses data non-exclusively and places it with labs.

Each request maps to a field of the row table: systems, date ranges, volumes and exclusions are all inventory fields.

How to Build the Inventory Without Moving the Data

An inventory is built from admin consoles, billing records and the people who run each system, and no record needs to leave its system to complete it.

1. List systems from billing, single sign-on and expense records, retired ones included.
2. Name the administrator and the plan for each system.
3. Record the earliest and latest date held, noting retention rules such as the Slack Free limit.
4. Count in the units of the units table above, and mark every estimate as an estimate.
5. Flag people and contracts that cover the records.
6. Note the export route and the deletion deadline, for example the 60-day life of a Google export bucket.
7. Set the status, include, exclude or hold, and date the inventory.

Counts and date ranges come from admin reports and system owners, not from exports. Rows flagged for people feed the work described under PII removal and de-identification, which happens before any record leaves the company.

Inventory Work: Benefits, Risks, Downsides and Alternatives, Including Not Selling

An inventory costs staff time before any offer exists, and it is useful even if the company never licenses anything. No public figure for the effort is recorded here.

Benefits

• One answer to every intake request in the intake table above.
• Retention losses are found early, while the records still exist.
• A ready exclusion annex for any agreement.

Risks

• The inventory reveals business detail, so it stays internal.
• Counting by exporting creates copies of the data.
• A buyer's published floor can anchor the count, although a floor is one buyer's rule.

Downsides

• No buyer pays for the inventory itself.
• Counts age and need a refresh date.

Alternatives

• Not selling or licensing at all.
• Licensing one data family only.
• Waiting for rights and consent to be settled.
• Asking an independent adviser to review the rows.

The full range of choices, with their costs, is set out under alternatives to selling data to AI labs.

Where the Inventory Sits in the Process to Sell or License Company Data

The inventory is step two of the seller process: it follows the readiness check and feeds the rights audit, consent review and every later step. The full sequence is in the guide to sell or license company data to AI companies.

1. Data sale readiness
2. Data inventory (this page)
3. Rights audit
4. Customer consent and confidentiality clauses
5. Employee notice and consent
6. How much company data is worth to AI companies
7. Choosing a buyer for your data

Inventory for a Closed or Insolvent Company

A closed company inventories what still exists, which depends on export deadlines that start at cancellation, and the person who signs may be a trustee or administrator. Microsoft 365 data is deleted no later than 180 days after cancellation (Microsoft page dated 2026-06-08), and Google excludes core service data from export once a subscription is cancelled or suspended (checked 2026-11-09). The route for such companies is described under selling data from a shut-down or defunct company, and vendor deadlines under exporting company data before shutting down.

Program Pages Cited on This Page

The lists, floors and intake steps above come from these programs' own pages, each checked November 9, 2026.

• AI data buyer match by published acceptance criteria
• Company data partnership programs
• Endless
• Marin Data
• Mercor data buyer profile
• micro1 data buyer profile
• Mode
• Nyne
• Ooak Data
• Pangea.ai
• Pewter
• Scale AI
• SimpleClosure
• Turing Project Lazarus profile
• Who buys AI training data

Frequently Asked Questions

What is a data inventory?

A data inventory is a record of the data an organization holds, where it is stored and how much there is. Before an AI data sale it adds the unit of count and whether each source can be licensed.

How do I create a data inventory for an AI data sale?

List every system, name its administrator, record date ranges and counts in buyer units, flag people and contracts, and mark each source include, exclude or hold. The seven steps above give the order.

What is the difference between a data inventory and a data catalog?

An inventory lists what exists and where; a catalog describes data for the people who use it, usually with search and definitions.

Do I have to export data to build the inventory?

No; counts and date ranges come from admin reports and system owners, and exports wait until a contract defines the scope. Scale AI states it anonymizes inside the company's environment before extraction, with per-package approval gates (checked 2026-11-09).