AI data market
Scale data partnerships: what should an operational data owner ask?
By SourceX Editorial · Reviewed by Noah Loul ·
9 sections
Short answer
Ask Scale which signed agreement governs the partnership, what exact records leave your environment, who de-identifies them and how that is tested, what uses and onward transfers the license allows, how long it runs, and what happens to delivered data when it ends. Scale's partnership page, checked October 18, 2026, describes in-environment anonymization and per-package approval; those are provider claims until a contract and a sample package confirm them.
Value drivers
Scale data partnerships
Key takeaways
- As of October 18, 2026, Scale's own page describes an ongoing data partnership built on structured, de-identified records from everyday business systems.
- Scale says anonymization happens inside your environment and that nothing is shared without per-package approval; treat both as provider claims until the contract and a sample package show them.
- Scale publishes several agreements with different data-use terms, so the first question is which signed document governs a data partnership.
- De-identification lowers but does not remove re-identification risk, and it does not by itself give you the right to license records.
- A buyer of partnership-sourced data needs the chain from the original record owner to the license it receives, not just a vendor's assurance.
- Stop or escalate when records contain third-party confidential terms, regulated data or customer contracts that restrict AI use.
What should you ask about Scale's data partnerships?
Ask for the governing agreement first, then for proof of each operational promise: what leaves your environment, who removes identifiers and how that is tested, what the recipient may do with the data, how long, and how it ends. Everything else on a program page is context until a signed document says it.
This page is a dated evidence review, not a recommendation for or against Scale AI. Every statement about Scale comes from Scale's own pages as they read on October 18, 2026, and is attributed as the company's claim. Programs change, so recheck the pages and the date before relying on any line here.
The page serves two readers. An operating company approached about a data partnership needs to know what to ask before any package leaves its systems. An AI team considering data that a vendor sourced through partnerships needs to know what evidence travels with it. The questions overlap more than either side expects.
What does Scale say its data partnership program is?
Scale's data partnership page describes an ongoing arrangement shaped around a business's operations, not a one-time file sale, and says it takes structured, de-identified records from everyday systems rather than access to the systems themselves. The table records each statement, where it appears and how it should be treated.
Status uses three labels. Provider claim means Scale says it on a marketing or program page and nothing public substantiates it. Provider-published record means Scale documents it in its own technical or legal publication, which is stronger but still not independent. Unknown means the public pages do not answer it.
| What Scale says (checked 2026-10-18) | Where it appears | Status | What would substantiate it |
|---|---|---|---|
| The partnership is ongoing and adapts to the business, rather than ending at file transfer | scale.com/data-partnership | Provider claim | Term, renewal and termination clauses in the signed agreement |
| It draws structured, de-identified records from everyday systems, such as document, spreadsheet and work-tracking records with their history | scale.com/data-partnership | Provider claim | A written data specification listing systems, objects and fields per package |
| It looks for years of records from real workflows, not demos or synthetic data | scale.com/data-partnership | Provider claim | Acceptance criteria in the agreement or statement of work |
| Anonymization happens inside the company's environment before extraction, and PII is stripped | scale.com/data-partnership | Provider claim | A described method, a re-identification test on a sample, and who runs the tooling |
| Nothing is shared without explicit, per-package approval | scale.com/data-partnership | Provider claim | An approval step written into the contract, with a record of each approved package |
| Employee and customer identities, raw systems and credentials stay with the company; audit trail, encryption and on-prem or VPC options are offered | scale.com/data-partnership | Provider claim | Security exhibit, access design and a sample audit log |
| Private startup codebases acquired through partnerships were used to build an evaluation set | labs.scale.com SWE-Bench Pro leaderboard | Provider-published record | Shows one documented use of partnership-sourced data; terms of those deals are not public |
| Who owns delivered data, permitted uses, onward licensing and price basis | Not stated in a public partnership agreement | Unknown | The signed agreement and any order form |
Which Scale documents could govern the deal?
Ask which document binds a data partnership, because Scale publishes several agreements whose data-use terms differ and none of them is labeled as the data partnership agreement. Until Scale names the governing document, assume none of the public ones applies as written.
The differences are material. Scale's Main Services Agreement lets Scale use aggregated and anonymized data, and de-identified data derived from a customer's use of the Services. Scale's Product Terms say that on the pay-as-you-go tier of its GenAI Platform the customer agrees Scale may publicly display and distribute Customer Materials. Scale's GenAI Portfolio page says customers retain full ownership of their data, business logic and custom solutions. Those are three different answers to the question of what happens to your data, written for three different relationships.
| Scale document | Written for | Data-use point it makes | Question to ask |
|---|---|---|---|
| Data partnership page | Companies supplying records | In-environment anonymization, per-package approval | Which signed agreement turns these statements into obligations? |
| Main Services Agreement | Customers buying Scale services | Scale may use aggregated, anonymized and de-identified usage data | Does any of this apply to a supplier, or only to customers? |
| Product Terms, GenAI Platform pay-as-you-go tier | Self-serve platform users | Scale may publicly display and distribute Customer Materials | Confirm this tier is not part of the partnership workflow |
| GenAI Portfolio page | Enterprise solution customers | Customer retains ownership of data and custom solutions | Is ownership of delivered records stated the same way for suppliers? |
Which statements are provider claims, and which are substantiated?
Almost everything public about the program is a provider claim; the one provider-published record is Scale's own description of an evaluation set built from partnership-sourced code. Nothing in the public record is independently audited, which is normal for a commercial program and is exactly why the questions below ask for documents.
Scale's SWE-Bench Pro leaderboard says its commercial subset contains 276 instances sourced from 18 private, proprietary codebases from startups, acquired through partnerships and not publicly accessible. That shows partnership-sourced data being used for model evaluation and kept unpublished. It does not show the price, the license scope or how those startups' agreements read, so it supports the existence of the route, not any particular term.
A useful discipline is to rewrite each claim as an obligation and ask where it is signed. 'PII is stripped inside your environment' becomes: who runs the tool, on which fields, tested how, and what happens if an identifier is found after delivery. If no document carries the obligation, the claim cannot be relied on, however credible the company.
Which questions settle buyer fit, and what evidence answers each?
Ten questions cover the decision, and each one names the document that settles it, so a meeting can end with a list of documents rather than impressions. Use the table as an evidence request: send it before the first call and mark each row answered, partly answered or unknown.
De-identification needs particular care. NIST's report on de-identifying personal information notes that as long as any utility remains in the data, some information might be linked back to the original identities, and it suggests data use agreements as an additional control. So the question is not whether identifiers are removed but how residual risk is tested and what the contract forbids.
| Question | Why it matters | Evidence that answers it | If unanswered |
|---|---|---|---|
| 1. Which signed agreement governs the partnership? | Public terms differ by relationship | Draft agreement and order form | Do not prepare data |
| 2. Exactly which systems, objects and fields are in each package? | Scope creep starts with vague package definitions | Data specification per package | Limit to one named package |
| 3. Who runs de-identification, with what tool, inside whose environment? | Control of the tool is control of the data | Method description and access design | Treat as raw data leaving your systems |
| 4. How is re-identification risk tested before release? | Removing names does not remove linkage risk | Test report on a sample package | Ask counsel before any release |
| 5. What uses are permitted, and can the data pass to model developers or other third parties? | Training, evaluation, resale and onward transfer are separate rights | Permitted-use, sublicensing and onward-transfer clauses | Restrict to named uses and the named recipient |
| 6. Is the license exclusive, and for how long? | Exclusivity limits your future options | Exclusivity and term clauses | Do not sign |
| 7. What happens to delivered data and derived models at termination? | Deletion rarely reaches trained weights | Termination and survival clauses | Assume delivered data persists |
| 8. Who warrants that you have the right to license the records? | Privacy clearance is not ownership | Warranty and indemnity sections | Run your own rights review first |
| 9. How is each package approved and logged? | Per-package approval only protects you if it is recorded | Approval workflow and sample audit log | Approve nothing without a written record |
| 10. How is compensation calculated, and can terms change without a signed amendment? | Recurring arrangements can be repriced or amended quietly | Pricing schedule, entire-agreement and amendment clauses | Fix terms for the first package and require signed amendments |
Worked example: a property manager reviews the program (illustrative)
This example is illustrative, not a real company or deal. A U.S. property management firm with eight years of maintenance work orders in a work-tracking tool and lease files in shared drives is approached about a data partnership. Its operations lead sends the 10 questions before the first call.
After two calls the firm marks questions 2, 3 and 9 answered: a draft specification names work orders and their status history, Scale's tooling would run in the firm's cloud account, and each package needs a signed approval. Questions 1, 5 and 7 are partly answered: a draft agreement exists but the permitted-use clause allows sublicensing and is silent on trained models after termination. Question 4 is unknown.
The decision follows the table. The firm narrows the first package to work orders only, asks for a re-identification test on 500 sample records, and asks counsel to rewrite the sublicensing and termination clauses before anything is prepared. Lease files stay out of scope.
The exception: tenant lease files contain negotiated terms and, in some cases, confidentiality clauses with commercial tenants. Even fully de-identified, those records may be covered by contracts the firm signed with third parties, so they need a rights review that de-identification cannot replace. That is why they were excluded, not because the partnership page says anything about them.
What should a buyer of partnership-sourced data ask?
A buyer should ask for the chain from each original record owner to the license it receives, because a vendor's assurance does not transfer rights the vendor never obtained. Ask which partner supplied each slice, under what agreement, with which permitted uses, and whether the vendor may sublicense to you.
Ask how promises made to the original data subjects are kept. FTC staff have written that it may be unfair or deceptive for a company to adopt more permissive data practices, such as using data for AI training, and disclose that only through a surreptitious, retroactive change to its terms. In a separate post, staff said companies can be liable for breaking promises not to use customer data to train models, and noted past orders requiring deletion of models built on unlawfully obtained data.
Finally, ask what the supplying company approved. Per-package approval, if it exists, should produce a record you can see in redacted form: package name, date, approver role and scope. If the vendor cannot show that record, you cannot show it to your own auditors either.
- Supplier agreement reference and permitted uses for each slice
- Whether sublicensing to you is expressly allowed
- De-identification method and the test result for the delivered version
- Package approval record from the supplying company
- Termination terms that would affect your copy or your trained models
What should end, narrow or escalate a Scale partnership review?
Stop when no signed document carries the promises you are relying on; narrow when only some record types are clean; escalate to counsel when the records involve third parties, regulated data or contracts that limit use. A partnership that can only proceed on marketing-page terms is not ready.
At a glance
What should end, narrow or escalate a Scale partnership review
Get specialist review when records include health, financial or children's information, employee files, or customer contracts with confidentiality or no-AI-use clauses. Narrow the first package to one system and one record type when de-identification has not been tested on your actual data. Revisit the whole review when Scale changes its program page, its terms or the governing agreement.
- Stop: no named governing agreement, or the draft allows unlimited sublicensing without your consent
- Stop: you cannot confirm you hold the right to license the records, whatever their privacy status
- Narrow: some systems hold third-party confidential terms; exclude them from the first package
- Narrow: de-identification is untested; release a small sample package first
- Escalate: regulated or sensitive records, cross-border transfer, or employee data
- Recheck: any change to the provider's pages or terms after your checked date
Do the same questions apply to SourceX?
The same questions apply to any data licensing provider, including SourceX as a direct licensee. SourceX's role is to take selected business records through rights review, contracting, data preparation, delivery and payment as licensed AI datasets, and its process has the supplying company approve scope and terms and sign before any record is delivered.
AI teams get data the same way in reverse: SourceX sources it from U.S. businesses for each request instead of keeping stock, so the request itself states scope, volume, language, format, timing and licensing needs and can carry the evidence questions above.
Frequently asked questions
Is a Scale data partnership the same as being a Scale customer?
Not necessarily. Scale publishes customer-facing agreements with their own data-use terms, but none is labeled as the data partnership agreement. Ask Scale which signed document governs a partnership before assuming any public term applies.
Does de-identification mean my company can license the records?
No. De-identification addresses privacy risk; it does not grant ownership or a right to license. Records may still be covered by customer contracts, confidentiality clauses or third-party rights, so a rights review is a separate step that counsel should cover.
Who owns data delivered under a data partnership?
The public partnership page does not state it. Scale's GenAI Portfolio page says enterprise customers retain ownership of their data, but that page is written for a different relationship. Ownership, permitted uses and onward licensing must be read from the signed agreement.
What does per-package approval protect against?
If written into the contract, it stops data leaving your environment without a specific sign-off for each defined package. It only protects you if packages are precisely defined and each approval is recorded with date, approver and scope.
How often should this review be repeated?
Recheck whenever the provider's program page, terms or governing agreement changes, and before each new package. This review is dated October 18, 2026; anything later must be confirmed against the current pages and the signed documents.
Sources
- Scale's data partnership page describes an ongoing arrangement using structured, de-identified records from everyday systems, anonymization inside the company's environment before extraction, per-package approval, and identities, raw systems and credentials staying with the company. Source
- Scale's Main Services Agreement allows Scale to use aggregated and anonymized data and de-identified data derived from a customer's use of the Services. Source
- Scale's Product Terms say that on the pay-as-you-go tier of the GenAI Platform the customer agrees Scale may publicly display and distribute Customer Materials. Source
- Scale's GenAI Portfolio page says customers retain full ownership of their data, business logic and custom AI solutions, while Scale owns the platform infrastructure. Source
- Scale says the SWE-Bench Pro commercial subset has 276 instances from 18 private, proprietary startup codebases acquired through partnerships and not publicly accessible. Source
- NIST notes that as long as utility remains in de-identified data some information might be linked back to identities, and suggests data use agreements as an additional control. Source
- FTC staff wrote that adopting more permissive data practices, such as AI training, and disclosing them only through a surreptitious retroactive terms change may be unfair or deceptive. Source
- FTC staff wrote that companies can be liable for breaking promises not to use customer data to train models, and noted orders requiring deletion of models built on unlawfully obtained data. Source
Tell us what data you need
Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.