Skip to content

Company and program review

Scale AI review: data partnerships, Outlier and buyer services

Compare Scale’s enterprise-data partnership, individual Outlier work and buyer engagements using current public evidence and an original workflow review.

Commercial basis
Enterprise partnership, buyer brief or Outlier project
Review updated
Provider website
scale.com ↗

The verdict

Start by choosing the right relationship. A business licensing operating records, an expert applying through Outlier, and an AI team buying evaluation services are pursuing different agreements.

What to check: Ask which structured records the partnership would accept and how permitted uses are documented.

How Fiund reviews companies and programs

Review basis. Fiund reviewed the cited public materials. Fiund also operates an AI data licensing business. This review includes no paid engagement or product rating.

Choose the relationship before comparing compensation

Scale has distinct public routes for AI buyers, individual experts and companies with operational records. A buyer can investigate data production and evaluation services. An individual can review Outlier opportunities. A company can investigate Scale’s Data Partnership. An offer or eligibility rule from one route does not determine the terms of the others.

The enterprise partnership page seeks structured, de-identified operating records from established businesses. It lists fields including professional services, manufacturing, energy, logistics and field services, and says approximately 40 employees is a preferred starting point with case-by-case assessment. It emphasizes real operating history. A fictional demonstration can explain a task without becoming an eligible deliverable.

What the operating-data page establishes

Scale’s published description includes records from business systems and a process with scoped package approval, de-identification and audit trails. Those are provider representations. A source owner should obtain the actual agreement and collection-specific explanation of how the process applies to its systems.

The public page displays illustrative partnership values. They are not a universal dataset price, a contributor pay rate or proof that a particular company will receive an offer. A firm should begin with the source types, history, ownership and useful decisions it can document. Access to internal files is a separate question from authority to license them.

An original field-service workflow example

Imagine an authorized fictional service ticket with an initial symptom, a diagnostic observation, a proposed action, an approved action and a recorded outcome. Add an attachment for a different job that uses similar wording. Ask an evaluator to identify which observation supports the final decision and whether the actual outcome is documented.

A useful specimen preserves the ticket-to-attachment link, event order, status and outcome reference. It should leave a missing completion record unresolved. Removing names while accidentally breaking those joins can make the record less useful; retaining a revealing job address can expose information outside the scope. The sample needs both task and disclosure review.

This is an independently authored exercise for explaining a proposed evaluation. It contains no real service records and reports no test on Scale data. A real collection must match the accepted source, rights and intended task. A synthetic package is not assumed to meet the partnership’s requirement for real operational work.

Expert work and published research serve other purposes

Outlier’s current offer and contributor terms are the relevant sources for an individual project’s eligibility, rate, paid activities, acceptance and payment timing. A completed assessment does not guarantee available work. Use the particular written offer when comparing compensation.

Scale Labs publishes research and evaluation work, including Humanity’s Last Exam and FineART. These demonstrate different uses for expert judgment and physical-action data under their documented setups. A research result is evidence for its stated experiment, rather than a purchase commitment or a measured result for another supplier’s collection.

What a buyer should request

Use a single brief that identifies the model task, source population, annotations, accepted unit and proposed rights. Request a specimen delivery, source and restriction record, review procedure and an explanation of which preparation is included. Compare collected examples, licensed existing records and expert-authored work as distinct scopes.

For an evaluation project, identify the reference answer, rubric version, disagreement process and held-out cases. For physical data, ask which signals are recorded and which are estimated, and whether the requested action representation is actually available. A company-wide capability menu does not settle those questions for the selected sample.

What a data owner should settle before transfer

Identify employer, employee, customer, vendor and other third-party interests in each source. Ask who can approve a package, what leaves the environment, which recipients may use it, how exclusions are enforced and what happens to existing licences after future participation ends. Confirm the commercial unit, deductions, reporting and payment event in the actual arrangement.

Fiund’s own standard contributor agreement defines its separate licensing authority and compensation. Compare the agreements directly for the same source and intended use. Neither a provider’s general governance statement nor a competitor’s review establishes asset-level clearance, acceptance or income.

Sources and further reading

Related industry exercises

Try a fictional worked example with records, a reference answer and a permission checklist.

Browse industry guides

Related reviews and guides

← Data licensing and catalogs