Definitions and comparisons
Exportable versus licensable data: a field-level test for investors
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
Exportable data can technically leave a system; licensable data is data someone has authority to let a buyer use for a stated purpose. The two overlap less than access claims suggest. Investors should review them separately, export evidence with an engineer and rights evidence with counsel, then join the results field by field, because only fields that pass both tracks can be delivered.
System export
Exportable
Key takeaways
- Export access measures what a system lets out; licensability measures who may approve a buyer's use. A pitch that counts integrations or API connections is counting the first.
- Exportable data fails the license test when it is held for someone else, when the access route forbids the use, when collection-time promises exclude it or when no license was ever attached.
- Licensable data can also fail to export intact: volume caps, missing history and dropped attachments impair value even when rights are clean.
- The metric to ask for is deliverable field coverage: the share of claimed fields that pass both an export test and a rights test, over a stated denominator.
- A signed agreement or a polished sample proves neither track; an export re-run and an investor-drawn sample from the manifest do.
What is the difference between exportable and licensable data?
Exportable data is data that can be extracted from the system holding it, through an admin export, an API or a connector. Licensable data is data that some party has authority to let another party use for a defined purpose, such as model training, evaluation or reinforcement learning environments. The first is a property of software and access rights; the second is a property of contracts, ownership and the promises made when the data was collected.
For an investor, the gap between them is where supply claims inflate. A data company's deck may count connected systems, API integrations or terabytes within reach, and every one of those measures exportability. Revenue depends on licensability, and on records arriving with enough context to pass a buyer's acceptance test. The decision this page supports is how much of a company's claimed supply is deliverable, and what evidence would show it.
The distinction matters more as buyers turn to operational records. In October 2026 micro1 announced it would commit $1 billion, over the next 12 months, to licensing and acquiring the operational data of enterprises to build reinforcement learning environments. It remains a plan rather than spending and is silent about any particular company's data, but it describes demand for records a company can license, which is a narrower set than records a company can export.
Four ways exportable data fails the license test
Exportable data usually fails because it is held for someone else, because the access route forbids the use, because collection-time promises exclude it or because no license was ever attached. Each failure is invisible to an export test and plain to a rights review, which is why the two reviews have to be separate.
Business contracts add a fifth layer that has nothing to do with personal data: confidentiality clauses that make a customer's documents, prices and specifications off-limits for any outside use, however easily they export. The four main failures, each with the evidence behind it:
- Held for someone else. A software company can export its customers' records in minutes, but the GDPR requires a processor's contract to limit it to processing personal data only on documented instructions from the controller, and Article 28(10) treats a processor that decides the purposes and means of processing as a controller for that processing. Hosting data is not authority over it.
- The route forbids the use. Slack's API terms say a provider of an app offered outside its own organization may not use API data to train a large language model, and they bar bulk export of message and file data unless an additional agreement allows it. The workspace owner's own export tools still work; the third-party route cannot carry a training license.
- Collection-time promises exclude it. FTC staff wrote in February 2024 that adopting more permissive data practices, such as using data for AI training, while telling consumers only through a surreptitious, retroactive change to terms or a privacy policy may be unfair or deceptive. Records collected under narrower terms do not become licensable when the terms change.
- No license was ever attached. The Data Provenance Initiative's 2023 audit of more than 1,800 text datasets found license omission of 70 percent or more and error rates of 50 percent or more on widely used hosting sites. Downloadable is not licensed, even for public datasets.
When licensable data cannot be exported intact
The reverse gap is quieter but costs money: the rights are clean, yet the records cannot leave with the context a buyer needs. Export tools are built for backups and migrations rather than datasets, and they often cap volume, flatten relationships or drop history.
Volume caps are the simplest case. Atlassian's documentation says Jira Cloud supports exporting up to 10,000 work items through its asynchronous CSV export from the issue navigator, and suggests splitting larger searches into batches with JQL. A company with a decade of tickets needs a tested extraction route rather than a button, and every batch boundary is a place where records can be duplicated or skipped.
Context loss is more serious. Comment threads, status history, attachments and links between records often need separate API calls or are missing from flat exports, and records stripped of them can fail a buyer's acceptance test even though the license covers them. The guide to read-only archives versus full exports explains why relationships decide whether retired-system records stay usable.
Buyers weigh both halves. OpenAI's data partnerships page asks for large datasets that the web does not already make easily accessible, and it excludes sensitive or personal information along with anything belonging to a third party. That description fits records that take real work to extract and that the supplier has clean authority to license.
The exportability and licensability matrix
Place each claimed dataset, and ideally each field group within it, in one of four cells. Only the first cell is supply today; the others are work in progress or liabilities and should be valued that way.
The four cells and how to treat each in diligence:
| Cell | What it looks like | Typical evidence | How to treat it in diligence |
|---|---|---|---|
| Exportable and licensable | Fields that leave intact and sit inside a signed grant for the buyer's use | Export manifest plus a grant clause covering the same fields | Count as deliverable supply |
| Exportable, not licensable | Customer-owned records, API-only access, data collected under narrow terms | Connector logs and API keys with no matching grant | Count as zero; a liability if already delivered |
| Licensable, not exportable | Rights are clear but history, attachments or volume cannot be extracted | A signed grant with no tested extraction route | Value only after an extraction test; cost it as engineering work |
| Neither | Systems the company can see but can neither extract nor license | Integration listings and screenshots | Exclude from supply claims |
The two-track evidence request list
Split the request list into an engineering track and a counsel track, have different reviewers run them in parallel and compare only at the end. Keeping them apart prevents the most common diligence error in this market: reading a working export as proof of permission, or a signed grant as proof that the fields exist.
Each request names the record to ask for, the test the investor's reviewer runs and the finding that should change the decision:
| Track | Evidence to request | Test to run | Red flag |
|---|---|---|---|
| Engineering | Export manifest per system: fields, row counts, date range, file checksums | Re-run the export on a date you choose and diff counts and checksums | Counts that cannot be reproduced |
| Engineering | Extraction route for each system: admin export, API or connector, with the terms version in force | Read the route's terms for training and bulk-export limits | Training data pulled through a route whose terms bar it |
| Engineering | Context map: which history, attachments and links survive extraction | Rebuild five complete records end to end from the export | Threads, attachments or status history missing |
| Counsel | Role per dataset: controller, processor or licensee, with the governing agreement | Confirm the company decides purposes for records it plans to license | Processor-held records inside licensed scope |
| Counsel | Grant clauses naming permitted uses: training, evaluation, RL environments | Match each buyer use to a clause, field group by field group | Uses with no matching clause |
| Counsel | Terms, notices and confidentiality clauses in force when records were created | Compare effective dates with record dates | Retroactive terms changes relied on for older records |
| Join | Field-level map marking each claimed field as passing or failing each track | Recompute deliverable field coverage from the map | Coverage stated without a denominator |
| Join | A stratified sample drawn by the investor from the manifest | Qualified domain reviewers score usability and rights fit | A sample selected only by the company |
How to join the tracks field by field
Join the two reviews at the field level, because export behavior and rights both vary inside a single system. One help desk record can hold a licensable ticket body, an internal-notes field restricted by customer confidentiality, attachments containing third-party documents and a status history that only an audit API returns.
Deliverable field coverage is the share of claimed fields that pass both tracks. State the denominator: the fields in the scope the company pitched, not the fields it later chose to show. Where some fields matter more to the buyer's task, weight them and publish the weights. A company can truthfully report that 90 percent of fields export and 90 percent are licensable and still deliver far less, if the failures fall on different fields.
The illustrative ledger in the figure shows one help desk dataset with six field groups. Ticket body and resolution codes pass both tracks. Internal notes export but fail the rights test. Attachments export only partly and fail on third-party content. Customer names export but are excluded. Status history is licensable, but the audit API it needs is untested. Two of six field groups are deliverable today, and the status history is the cheapest to move into the first cell.
Does a signed agreement or a polished sample prove deliverable value?
No. A signed agreement proves a grant exists; it does not show that the granted fields can be extracted with their context, that they fall inside the grantor's own authority or that they pass a buyer's acceptance test. A polished sample proves that some records are good, not that the population resembles them.
The test that separates the two is an investor-drawn sample. Ask for the export manifest first, commit it before any draw, select a stratified sample by system, year and record type, and have qualified domain reviewers score each record for usability and rights fit. A company that cannot produce a manifest to draw from has not yet shown what it would deliver.
Reproducibility is the last check. If the company re-runs the export a month later and the counts or checksums move without an explanation such as new records, the manifest describes a moving target, and every coverage figure built on it inherits the drift.
What could make the exportable-licensable gap matter less
Several developments could narrow the gap or reduce what closing it is worth, and an investor should weigh them before treating licensing authority as a durable advantage:
- Law could reduce the need to license some uses. NIST's Generative AI Profile (NIST AI 600-1, 2024) describes intellectual property risk as arising where the use of copyrighted works is not a fair use, a boundary courts are still drawing. Rulings that widen permitted uses of content a developer already holds would lower the premium on some licensed content, though they give no route to private records.
- Buyer demand could stay concentrated. Reddit's Form 10-Q for the quarter ended March 31, 2026 states that 'substantially all of the contract value associated with our licensing revenue is derived from two of our partners.' A supplier with clean, deliverable data can still depend on a few buyers.
- Platform terms can move in either direction. Slack's current API terms close the third-party training route; a later revision of those or any platform's terms could open or close a route a company's supply depends on.
- Stated demand may not convert. micro1's $1 billion figure is a plan for the next 12 months, not spending, and it is no evidence that any given company's records will be bought.
- Review cost grows with complexity. Every added system, acquisition and customer contract family adds engineering and counsel hours, which shows up both in a data company's margins and in the investor's own diligence timeline.
Questions to ask any data company, SourceX included
Put the same questions to every company that claims licensable business data, and expect each answer to arrive as a record rather than a description.
SourceX's published model places the rights step before delivery. It carries a company's selected operational records through sourcing and qualification, rights review, contracting, preparation, delivery and payment to AI developers, and data reaches SourceX only after the company has approved scope and terms and signed; from that point SourceX is permitted to use, sublicense and distribute the de-identified dataset within what the agreement allows. SourceX does not publish deal, supply or commercial metrics, keeping company and commercial metrics for direct sharing with investors. Its Rule 506(c) raise is with accredited investors, and investor materials are shared on request through sourcex.si/investors.
The questions:
Checklist · 6 items
- Which claimed datasets are exportable and licensable today, and which sit in the other three cells?
- For each dataset, who is the controller, and which agreement gives you authority to grant training, evaluation or RL-environment use?
- Which extraction route does each dataset use, and do that route's terms permit the use you sell?
- What is your deliverable field coverage, over what denominator, and how did it change across your last three exports?
- May we choose the system, date and sample for an export re-run and a reviewer-scored sample?
- Which records were collected under terms older than the ones you rely on now, and how are they excluded?
0 of 6 ticked. Tick each item as you go.
Methods note: how the two-track list was built and what it leaves out
The matrix, the two-track request list and deliverable field coverage are editorial frameworks built for this article on October 27, 2026 from the sources cited: the GDPR's processor provisions, Slack's API terms, FTC staff guidance from February 2024, the Data Provenance Initiative audit, Atlassian's Jira Cloud export documentation, OpenAI's data partnerships page, NIST AI 600-1, Reddit's March 2026 quarterly report and micro1's October 2026 announcement. All nine were confirmed through searches restricted to the publisher's own site, and a reviewer should open each before release.
The help desk ledger is illustrative and describes no real company. Nothing here measures any company's supply, rights position or coverage, SourceX's included, and no score is assigned to anyone. The legal points are general information; counsel should review the agreements behind any specific dataset.
Frequently asked questions
Is data a company can export automatically data it can license?
No. Export access shows that software lets records out; licensing needs authority over the records for the buyer's specific use. Records held as a processor for customers, pulled through a route whose terms bar training, or collected under narrower promises can be exported but not licensed.
Does an API connection give a data company the right to train on the data?
Not by itself. API terms often restrict use: Slack's API terms, for example, bar apps offered outside their own organization from using API data to train a large language model. The right to use data for training comes from the data owner's grant, and the route has to permit that use too.
What evidence shows a dataset is licensable, not just exportable?
A signed grant naming the permitted uses, proof that the grantor controls the records rather than holding them for someone else, and the terms and notices in force when the records were created. Each should be matched to the fields being delivered, not to the system as a whole.
How should an investor measure the gap between exportable and licensable data?
Use deliverable field coverage: the share of claimed fields that pass both an export test and a rights test, with the pitched scope as the denominator. Recompute it from a field-level map and check it against an export re-run and an investor-drawn sample.
How can an investor put these questions to SourceX?
Request investor materials through sourcex.si/investors and ask the questions above. Rather than publishing company and commercial metrics, SourceX shares them with investors directly; any offering is made through its offering documents, to verified accredited investors and no one else.
Sources
- micro1 announced in October 2026 a commitment to spend $1 billion over the next 12 months acquiring and licensing enterprise operational data for reinforcement learning environments; a stated plan. Source
- GDPR Article 28 requires a processor's contract to limit it to processing personal data on the controller's documented instructions, and treats a processor that determines purposes and means as a controller. Source
- Slack's API terms bar providers of apps offered outside their organization from using API data to train a large language model or bulk exporting message and file data without an additional agreement. Source
- FTC staff wrote in February 2024 that adopting more permissive data practices such as AI training through surreptitious, retroactive terms changes may be unfair or deceptive. Source
- The Data Provenance Initiative's 2023 audit of more than 1,800 text datasets found license omission of 70 percent or more and error rates of 50 percent or more on widely used hosting sites. Source
- Jira Cloud supports exporting up to 10,000 work items with its asynchronous CSV export; larger searches are split into batches with JQL. Source
- OpenAI's data partnerships page seeks large datasets not already easily accessible online and excludes sensitive or personal information and information belonging to a third party. Source
- NIST AI 600-1 states that intellectual property risks from generative AI may arise where the use of copyrighted works is not a fair use. Source
- Reddit's Form 10-Q for the quarter ended March 31, 2026 states that substantially all of the contract value associated with its licensing revenue is derived from two partners. Source
Request SourceX investor materials
SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.