Selling company data for AI in 2026

Your company may have records that an AI developer wants to license. Several programs seek business databases, project histories and documents linked to professional decisions. Handshake and Mercor both publicly recruit companies for data partnerships. ^[1] ^[3]

The data must serve a buyer's intended use, your company must have permission to license it, and the agreed payment must justify the costs and obligations.

Buyer requirements

Some developers seek specialist content for model training. Others need realistic tasks for training or testing AI agents that carry out work across software systems.

Handshake says it uses employer data to build environments in which models learn to complete tasks. Mercor describes demand for records that connect tools, working procedures and decisions. An assessment should identify the task a buyer could train or test using a company's records. ^[1] ^[3]

OpenAI's 2023 data-partnership announcement sought material that was difficult to access publicly. It included text, images, audio and video, with an emphasis on human context. The announcement explains a type of buyer interest but provides no evidence of a current offer for an individual archive. ^[4]

Structured records and working histories

A documented database can help a buyer understand inventory movements, equipment readings, transactions or service outcomes. Its fields need explanations, and its records need to meet the buyer's quality requirements.

Working histories can include tickets, procedures, document revisions, messages and related decisions. Scale lists data warehouses, communication tools, documents, work-management systems and business platforms among its areas of interest. micro1 also names process documentation, project histories and human feedback on AI outputs. ^[2] ^[5]

Consider a hypothetical property-maintenance case. A reviewer can establish the final charge from an invoice. The original request, inspection notes, approved quote, scheduling changes and completion check provide more information about the work. A buyer testing a system's handling of maintenance requests may need those records to remain linked.

The proposed use affects both usefulness and permissions. A company-authored process guide and a borrower's application require different assessments. Real estate, title, mortgage and insurance records should be reviewed by category before they are included in a proposed license.

Advertised payments and reported revenue

On October 7, 2026, Handshake advertised a typical range of $100,000 to $4 million for qualified partnerships, depending on the data. Scale advertised $10,000 to $1 million-plus as an illustrative value per partnership, with value affected by delivery frequency. ^[1] ^[2]

These providers do not publish a representative set of completed operating-data deals from which to calculate a typical payment. Their advertised figures cannot establish an applicant's eligibility, likelihood of an offer or expected revenue.

Defined.ai separately reports average partner revenue of at least $600,000 and revenue of at least $1 million for at least five partners. The page does not disclose the period, dataset mix or calculation method. These provider-reported results therefore have limited use in estimating a new applicant's payment. ^[12]

Wiley reported $40 million in AI licensing revenue for fiscal 2025. This company-reported financial result concerns publishing content. Differences between that catalog and a business's operating records limit its use as a pricing comparison. ^[6]

An offer should specify the dataset, allowed uses, delivery obligations and payment milestones. Those terms determine what the seller must provide and when it is entitled to payment.

Factors in a data assessment

Scale emphasizes established systems and specialist expertise. micro1 includes process complexity, quality and uniqueness in its compensation factors. These are provider-specific criteria, so they can inform an assessment without supplying a general pricing formula. ^[2] ^[5]

A review can address the following questions:

Does the material support a task the buyer currently wants an AI system to perform?
Can the buyer follow a case from its initial information through decisions to an outcome?
Does the collection contain specialist knowledge that would be difficult to obtain elsewhere?
Are the records legible, consistent and linked, with explanations of the fields?
Can the company establish permission and prepare an approved dataset at an acceptable cost?
What uses, duration, exclusivity, updates and support would the payment cover?

Volume affects the number of examples available, but incomplete or restricted records may reduce the usable collection. A representative sample can help both parties define what can be delivered.

Licensing stages

A private license grants agreed rights to use a defined dataset. The agreement should address ownership, exclusivity, redistribution and permitted use of derived material. Retaining ownership can coexist with broad rights granted to the buyer, so both parties need to review the full scope.

Programs differ in how they reach an agreement. micro1 describes evaluation, discovery and agreement stages. Mercor describes scoping, connection, processing and payment. The seller and provider should agree the sequence, responsibilities and approvals before any access or transfer. ^[3] ^[5]

The seller can describe systems, record types, date coverage, approximate volume and business context without providing customer records.
The parties should confirm the buyer, intended use, confidentiality arrangements and authority to proceed. Restrictions need review before a sample is shared.
The agreement should define included and excluded data, permitted recipients, training and evaluation rights, fees and any ongoing work. It should also set acceptance criteria and payment dates.
Preparation should follow an agreed method for privacy protection, quality checks and review of the proposed delivery. The agreement should assign responsibility and costs for each step.
Delivery should use the approved transfer method. The parties should record acceptance and payment, then follow the agreed retention, deletion and future-delivery terms.

Rights and privacy

A company needs permission for the proposed AI use. A review should cover customer contracts, confidentiality commitments, employee arrangements, third-party content licenses and source-system terms. Counsel can assess which approvals the proposed use requires.

Existing privacy promises also apply. The FTC has warned that quietly changing terms or privacy policies to allow new AI uses of previously collected information may be unfair or deceptive. A proposed deal needs to comply with the obligations attached to the records. ^[7]

The FTC explains that the Gramm-Leach-Bliley Act limits certain disclosures of consumer financial information. It can also restrict a recipient's reuse and redisclosure. FTC security guidance expressly includes mortgage lenders and brokers among potentially covered businesses. ^[8] ^[9]

Insurance and title businesses need to establish which rules and contracts apply to their activities before including client records. The required review depends on the information and proposed use.

Working histories may identify customers or employees through emails, documents and case details. Where applicable, California privacy law gives consumers rights concerning sales of their personal information. Describing a collection as business data does not change those duties. ^[10]

Privacy preparation and access controls

Removing names may leave context or combinations of facts that identify someone. NIST recommends evaluating disclosure risks and warns that masking personal information may be insufficient. Although its guidance addresses government datasets, the same distinction is relevant when reviewing a provider's de-identification claims. ^[11]

The handling plan should identify where preparation takes place and who can see the original records. It should name any subprocessors, explain access logging and state whether the seller can approve the final dataset. The agreement should also cover incident handling and prohibited uses.

Deleting delivered files is a separate issue from ending use of derived datasets or trained models. The agreement needs to explain what happens at its end, which uses may continue and what the buyer must delete. Ending a license may leave models that were trained during the permitted period.

Initial eligibility

An initial assessment needs a description of the records and someone who can review the relevant restrictions. Some providers also state preferred company sizes. Scale gives a preferred starting point of 40 employees, assessed case by case, while micro1 describes a target of 30-plus employees. Those criteria apply to the individual programs. ^[2] ^[5]

The first review should answer these questions:

Which task or specialist knowledge do the records concern?
Where are the records stored, how much is available and what period do they cover?
Who controls the records, and which outside permissions or restrictions apply?
Which personal, confidential, privileged or regulated material needs exclusion or further review?
Who can approve the scope, preparation and commercial terms?
Would the expected payment justify the costs and continuing obligations?

Further work should depend on the buyer's requirements and the restrictions identified. A company can assess potential interest before deciding whether to proceed with a license.

DataPayouts assessment

Start with a data assessment by describing your business and the types of records it maintains. Keep customer documents and other sensitive records out of the initial inquiry.

This guide provides general information with a U.S. emphasis. Obtain legal advice for the proposed use and applicable jurisdiction. Program details reflect public information available on October 7, 2026 and may change.