Published programs use five delivery models: files uploaded or shipped, an export by the company, an export by the program, a read-only connection and a continuing API feed. The buyer's review that follows decides when the invoice can be issued.
In AI training data licensing, a standard export can drop comments, attachments or message content, so the format and the checks before the first transfer matter. Under micro1's terms the buyer has 30 business days after delivery to reject, and silence counts as acceptance (s.1.3).
Program statements below are the named party's own words, checked 2026-11-09 unless another date is shown, and the method is set out in how we source, date and check every figure.
What Delivery Covers in an AI Data License
Delivery covers five things the contract should fix before anything moves: the scope of the records, their format, the transfer method, the schedule and the evidence of what was handed over.
- Scope: systems, date ranges and record types, and exclusions.
- Format: the file type and the export tool.
- Method: upload, shipped media, program export, connection or API.
- Schedule: one handover, batches or a refresh.
- Handover evidence: a record of what was sent and when.
Evidence matters because micro1's terms let the buyer reject delivered data "in whole or in part" (s.1.3).
Copy or Access: Five Delivery Models
Programs that publish their process use five delivery models, from a one-time file transfer to a connection the program keeps open to company systems. The table runs from the most closed to the most open, as of 2026-11-09 unless a later date is shown.
| Model | How the data moves | Who prepares and de-identifies | Published example (party, role) | As of | Source |
|---|---|---|---|---|---|
| Files sent by upload, cloud transfer or shipped media | the company sends files once | the receiving party processes | Troveo (marketplace): drag-and-drop upload, cloud transfer or shipped media; processing 1 to 2 weeks | 2026-11-09 | troveo.ai |
| Company export, uploaded | the company exports an agreed copy and transfers it | the company, or the receiving party where it says so | McCarthy's AI (broker): the business exports a de-identified copy, keeps its originals and never shares passwords; micro1 (buyer program): transfer via micro1 APIs or approved methods, micro1 de-identifies; Spirit alternate bid: secure upload at data.mercor.com | 2026-11-09; Dkt. 1463 2026-08-14 | mccarthys.ai; micro1.ai; CourtListener |
| Program-run export of an agreed copy | the program exports a defined copy once | the program | Mode (buyer program): exports and de-identifies the agreed copy; Collinear (buyer program): export and anonymization, one-time snapshot, access revocable after transfer | 2026-11-09 | data.mode.inc; collinear.ai |
| Read-only connection to source systems | the program connects and extracts the approved scope | the program | Mercor (buyer program): connect, automated extraction and de-identification; Centific (buyer program): read-only, background extraction; Magma (marketplace): read-only encrypted connection; Pangea (marketplace): read-only code hosts; Gradient Health (data vendor): EHR, PACS or RIS cloud access | 2026-11-09 | mercor.com; centific.com; magmahq.ai; pangea.ai; gradienthealth.io |
| Continuing API feed | the licensee pulls data through an API during the term | the licensor | Reddit: continuous API access under its January 2024 licenses; Wikimedia Enterprise: on-demand, snapshot and realtime API | 2024-02-22; 2026-01-15 | sec.gov; enterprise.wikimedia.com |
When the Program Extracts the Data Itself
Several programs connect to the company's systems with read-only access and extract the agreed scope themselves, so the company never builds a file.
Mercor states "No AI lab buyer or third-party vendor ever sees your raw data", and Centific states "You license your data. You never hand it over." Magma limits its connection to the approved scope, and Pangea.ai connects GitHub, GitLab or Bitbucket read-only.
Access after signing differs from the in-environment access used to prove data quality, which lets a buyer test a dataset before any offer.
When the Company Exports and Uploads a Copy
In the copy model the company, or the program on its behalf, exports a defined set of records once and transfers that copy, and the source systems stay closed.
McCarthy's AI describes a business that exports a de-identified copy, and micro1 states that transfer runs through its APIs or approved methods, with micro1 de-identifying sensitive information itself.
Mode exports and de-identifies the agreed copy, and Collinear states a one-time snapshot with access revocable after transfer.
Continuing Feeds Instead of One Handover
Some licenses deliver through an API that keeps running for the term, which turns delivery from a one-time event into a service.
Reddit's Form S-1 (filed 2024-02-22) states an aggregate contract value of USD 203.0 million for its January 2024 licenses, two to three year terms and delivery by continuous API access, without naming the counterparties, and Wikimedia Enterprise's partner announcement of 2026-01-15 describes commercial API access (on-demand, snapshot and realtime).
Choosing a Delivery Format: What Exports Keep and Drop
The export format decides what reaches the buyer, because several systems' standard exports leave out comments, attachments or message content, or cap the number of records.
Mechanics per system are covered under export rights and mechanics by source system, and the table shows what vendors' own documentation says each export omits or caps, as of 2026-11-09 (Purview 2026-11-16).
| System | Standard export format | What it leaves out or caps | As of | Source |
|---|---|---|---|---|
| Intercom | Dataset CSV | conversation and message content not included | 2026-11-09 | intercom.com |
| Linear | CSV; Markdown copy | CSV excludes attachments and its field list has no comments; the Markdown copy includes comments | 2026-11-09 | linear.app |
| Microsoft Purview eDiscovery | PST packages | package size 1, 2, 5 (default) or 10 GB | 2026-11-16 | learn.microsoft.com |
| Notion | HTML or Markdown and CSV, plus files | PDF with subpages only on Business or Enterprise | 2026-11-09 | notion.com |
| ServiceNow | XML, CSV, XLSX, XLS, JSON | default export limit 10,000 records, Zurich release | 2026-11-09 | servicenow.com |
| Slack (Enterprise) | JSON or TXT | Org Owners contact Support to enable self-serve export | 2026-11-09 | slack.com |
| Zendesk | JSON; XML | JSON omits comments when a ticket exceeds 1 MB; XML capped at 500 MB, about 200,000 tickets, not available above | 2026-11-09 | support.zendesk.com |
None of the programs cited here publishes a required file format, so the format is agreed per deal.
Stating the Delivered Format in the Manifest
The format, the export tool and any known gaps belong in the written description of the delivery, because a buyer can reject data that does not match it and, under some terms, claw back the price later.
In micro1's terms the Data Manifest is the licensor's written description approved by the licensee (s.1.1), and a refund can follow if it is materially inaccurate, even after acceptance (s.4.3).
A data manifest for AI buyers lists format and known gaps for that reason, and Nyne states that it formats the data itself.
Checks Before Data Leaves the Company
Seven checks before the first transfer catch the problems that are hardest to fix once a copy has left.
- The delivery matches the scope in the order form.
- Excluded categories are removed: micro1's s.5.7 lists seven, from health data and biometrics to export-controlled data, unless the order form says otherwise.
- De-identification is done, or its owner is named in writing, as micro1 and Mode state they do it themselves.
- Record counts and date ranges match the manifest.
- Exports are complete for comments and attachments.
- The export-control screen is done.
- A log records what was sent, when and how.
AfterQuery states a re-identification test before delivery and adds that "no process removes all risk on its own". None of the published terms cited here names a checksum or count reconciliation. Methods for the removal step are covered under PII removal and de-identification before licensing.
Delivering in Stages: Sample, First Delivery, Full Set and Refresh
Most published processes deliver in stages, so the full archive moves only after a sample has been priced and a first delivery has been reviewed.
- Sample under NDA: Patronus AI states a sample reviewed in 5 business days, and Sieve a rights review of samples.
- First delivery: Mercor's referral page ties the referrer's payment to the company's first delivery.
- Full delivery: Patronus AI states full delivery followed by a review of up to 30 days; Pangea.ai states batches to buyers over "normally between 2 and 16 weeks", subject to demand.
- Refresh or further batches: Centific offers an optional quarterly refresh after a first accepted contribution typically in six to ten weeks.
The published terms do not say whether a review window restarts for each batch, so that point belongs in the contract.
From Handover to Invoice: The Buyer's Review Window
Handover starts the buyer's review, and under published terms the invoice or payment waits until that review ends. micro1's terms give 30 business days to reject, a license that attaches after acceptance and payment (s.1.2) and invoices only after acceptance (s.3), while its program page says that after transfer "micro1 has 30 days to review and approve it".
Patronus AI states a review of up to 30 days and payment on acceptance. Clauses on rejection notices, cure and deemed acceptance are covered under acceptance, rejection and deemed acceptance of delivered data.
Protecting Data While It Moves
The license, not the transfer tool, sets the security duties during delivery, and micro1's terms require encryption in transit and at rest (s.2.4) and notice of a security incident within 2 business days after becoming aware (s.2.5). Magma states an encrypted connection, and Mercor states SOC 2 Type II and AWS KMS keys.
Controls, access rules and breach duties are covered under security during an AI data sale.
Copies Left Behind Once Delivery Ends
Delivery leaves copies on both sides, the transfer copy and raw archive with the buyer and the export files with the company, and the contract decides when each is deleted. micro1's s.4.4 deletes raw data at the earlier of 120 days after completion of delivery or termination, with no duty to delete or retrain Licensee Materials, and Ooak Data states raw data deleted within 30 days of delivery with written confirmation.
Terms on deletion, audits and reuse limits after the sale are covered separately.
Questions to Settle in Writing Before the First Transfer
Eight answers turn "delivery by agreed method" into terms both sides can check. They belong in the order form of an AI data licensing agreement, and on cost micro1's s.1.1 puts export and delivery costs on the licensor, while Mercor states it covers extraction, anonymization and transfer costs.
- Which method is used, and who operates it?
- Which format, and which export tool produces it?
- Which scope and which exclusions apply?
- Who de-identifies, and to which standard?
- What is the batch schedule?
- What counts as the delivery date?
- Who keeps transfer copies, and for how long?
- What happens to a rejected batch?
Copy, Connector or No Delivery: Benefits, Risks, Downsides and Alternatives
A read-only connector saves the company work but gives a program ongoing reach into its systems, a one-time copy keeps the systems closed but puts the export work on the company, and not delivering at all is a valid outcome.
Benefits
- A connector means the company builds no files, as Mercor and Centific describe.
- A one-time snapshot, as Collinear states, ends access after transfer.
- A program-run export moves de-identification to the program, as Mode states.
Risks
- A connector reaches everything the connected account can see unless the scope is limited in writing.
- A copy that leaves cannot be recalled: micro1's s.4.4 sets no duty to delete or retrain Licensee Materials.
- A refund can follow acceptance when the manifest is materially inaccurate (s.4.3).
Downsides
- The company builds and checks the export in the copy model.
- Standard exports drop content, as the table above shows.
- Staged delivery adds review steps.
Alternatives
- A narrower one-time snapshot.
- Batches with a review between them.
- A description or sample only, with no delivery of the full set.
- Not licensing at all, as covered under alternatives to selling data to AI labs.
Where Delivery Sits in the Process to Sell or License Company Data
Delivery follows the signed contract and comes before payment, so the steps on either side decide what is delivered and when it is paid. The full sequence is in the guide to sell or license company data to AI companies:
- Before signing: term sheets, order forms and LOIs.
- Before signing: due diligence in an AI data deal.
- Preparing the files: data room for an AI data licensing deal.
- After acceptance: when and how AI data buyers pay.
- Overall duration: how long an AI data licensing deal takes.
Export-Controlled Engineering Data in a Delivery
Engineering archives can hold export-controlled technical data, and for that data the transfer method matters legally, not only technically.
Under 15 CFR 734.18(a)(5), end-to-end encrypted sending or storing of unclassified technology, not stored in Country Group D:5, is not an export, and 22 CFR 120.54(a)(5) sets a similar rule for ITAR technical data, both in force. micro1's s.5.7(g) excludes such data unless the order form allows it.
The wider topic is covered under export controls and licensing technical data to AI buyers. This is information, not legal advice.
Delivery in a Court-Supervised Data Sale
In a bankruptcy sale the delivery terms sit in the sale agreement the court is asked to approve, and the Spirit Airlines record shows them in writing.
Mercor.io Corporation's alternate bid of USD 7,500,000 at the 2026-08-14 auction names secure upload at data.mercor.com as the delivery method (Dkt. 1463, Exhibit C), and the successful bid by Google LLC of USD 10,000,000 puts de-identification costs on the buyer, with the sale not approved as of 2026-11-09. See data assets in insolvency and selling data from a defunct company, and the Spirit Airlines data archive sale.
Programs Cited on This Page
The delivery terms above come from these programs' own pages, each checked on the date shown in the tables.
- AfterQuery profile
- Centific profile
- Collinear AI profile
- Gradient Health profile
- Magma profile
- McCarthy's AI profile
- Mercor data buyer profile
- micro1 data buyer profile and micro1 Data License Master Terms
- Mode
- Nyne profile
- Ooak Data profile
- Pangea.ai profile
- Patronus AI profile
- Reddit's AI data licensing deals
- Sieve profile
- Troveo profile
- Wikimedia Enterprise AI deals
- AI data partnership payouts and program terms compared
Frequently Asked Questions
How is licensed data delivered to an AI buyer?
By one of five methods: files uploaded or shipped, a company export, an export the program makes, a read-only connection or a continuing API feed, as the license agrees. micro1's terms call for "a mutually agreed upon delivery method" (s.1.1).
What file format do AI buyers want?
None of the programs cited here publishes a required file format; the format is agreed per deal and should keep the comments, attachments and content the standard export drops. Zendesk JSON, Intercom CSV and Linear CSV each omit some.
Can personal data be delivered to a buyer in the US?
Personal data from the EU needs a transfer mechanism before delivery. The page on transferring training data to US buyers covers it.