Skip to content

AI data market

Handshake AI business model analysis: a talent network buying archives

By SourceX Editorial · Reviewed by Noah Loul ·

10 sections

Short answer

Handshake AI's business model, read from Handshake's own pages, has two supply lines. A career network of verified graduate-level experts sells hours to AI labs under contracts that assign the work to Handshake, and a data partnerships program acquires operational records from companies. The lines face different limits, hours versus history, and Handshake publishes no revenue, fellow hours or license terms for either, so investors should value them separately.

System export

Handshake AI business model

This page covers handshake AI business model analysis: a talent network buying archives.Illustrative

Key takeaways

  • Handshake runs two AI supply lines: a fellowship that sells verified experts' hours and a data partnerships program that acquires companies' operational records.
  • Its network figures (18 million, 20 million, 30 million, 500,000 PhDs) describe different populations on different pages and cannot be added or turned into a growth rate.
  • The contractor agreement assigns fellows' work product to Handshake and bans AI tools without written permission; the network's visible asset is verification and scheduling, while any reuse of the output depends on unpublished lab agreements.
  • The records program's license terms, exclusivity and reuse across labs are not published, which leaves its economics undetermined.
  • The test for investors is the split of AI revenue between expert hours, licensed records and enterprise work, none of which Handshake discloses.

What does Handshake AI's business model say about expert supply?

Handshake's own pages describe two kinds of supply sold into AI development. The first is people: graduate-level experts recruited from Handshake's career network of students, alumni and employers, verified through university partnerships and engaged as independent contractors whose work product is assigned to Handshake. The second is records: operational data that Handshake acquires from companies through a program it calls data partnerships. Expert supply, in Handshake's case, is a recruiting asset first and a data asset second.

That split is the decision an investor has to make before any comparison. A talent network grows when it can recruit, verify and schedule more qualified hours; a records program grows when more companies with usable history agree to license it. The two lines face different limits, carry different rights and would deserve different multiples, so a reader who treats Handshake AI as one business will misread whichever line is growing.

Useful evidence would be revenue by line, hours delivered by fellows, the number of companies that have licensed records and whether any acquired dataset reaches more than one lab. Handshake publishes none of these on the pages read for this article on 2026-10-28. The table below sets out what it does publish and keeps every missing field visible; the rest of the page tests what the published facts support.

Five Handshake AI activities and the fields Handshake leaves empty

Five fields cover what Handshake says about its AI activities on its own domain. Every entry is the company's statement, and the last column names what an investor would need and cannot find there.

Two patterns stand out. Most figures describe the size of a network, not the volume of work or money moving through it. And the one figure that touches the records program, a partner count shown on its licensing page, is self-reported and carries no date in the excerpt verified here:

Five Handshake AI activities and the fields Handshake leaves empty
ActivityWhat Handshake saysWhere (all read 2026-10-28)Kind of figureNot disclosed
Expert fellowshipGraduate-level experts provide human validation of AI models for labs; fellows are 1099 contractorsIntroducing Handshake AI post; fellowship FAQ for university partnersProgram descriptionFellow count, hours delivered, revenue, lab names
Data partnershipsAcquires operational records from companies; data is scoped, de-identified and confidential, and the company decides what is sharedData partnerships page; employer data program postProgram descriptionLicense terms, exclusivity, revenue, reuse across labs
Partner eligibility and countSeeks companies with 20+ full-time employees and 3+ years in operation; shows 1K+ partner companies, undatedData partnerships licensing pageEligibility rule; self-reported countDate of the count, active versus signed partners
Enterprise AIPitches model ownership and agent deployment, citing 30M+ professionals and 200+ researchersEnterprise AI pageNetwork and staff claimsCustomers, revenue, contract terms
AcquisitionsBought Cleanlab, whose leaders join to run research, and Taro, a software engineering communityAcquisition posts on Handshake's blogCorporate eventsPrices, dates in the excerpts, integration results

Five network counts that cannot be added together

Handshake's network figures describe different populations on different pages, so they cannot be summed, compared over time or read as active supply.

The launch post introducing Handshake AI cites more than 500,000 PhDs, over 50,000 in STEM. The About page counts 18 million knowledge workers, 1,600+ educational institutions and 1 million employers. The Taro acquisition post refers to more than 20 million students, alumni and professionals, and the Enterprise AI page cites 30M+ professionals.

Each count answers a different question: registered members, graduate holders, institutions, employers or professionals. None counts the people doing AI work in a given week, and none carries a date in the excerpts. The rise from 18 million to 30 million may reflect growth, a broader definition or both; the pages do not say, so no growth rate can be computed from them.

The figure that does carry economic meaning is the one Handshake leaves out: active fellows and the hours they deliver. The fellowship FAQ for university partners describes typical time commitments of 5 to 20 hours a week and projects that run from two weeks to more than three months, which is the shape of part-time, project-based labor. Supply in that model is the number of qualified hours a network can schedule, not the size of its membership.

Who owns what the expert network produces?

Handshake's contract terms send the network's output to Handshake and its partner labs, not to the experts who produce it. Whether Handshake may reuse any of that output depends on its agreements with labs, which are not published; on the contributor side, the experts keep no rights in it.

The Handshake AI Contractor Agreement says the contractor irrevocably assigns all right, title and interest in intellectual property created under the agreement, including deliverables, and that the contractor keeps no right to use the work product. Where a right cannot be assigned, the contractor grants Handshake an exclusive, perpetual, royalty-free license with the right to sublicense. The same agreement prohibits large language models and other AI tools in the work unless Handshake authorizes them in writing.

The fellowship FAQ written for university partners frames the same point from the fellow's side: no previous or future intellectual property moves from fellows to Handshake or its partner labs, while materials created within the program are owned by Handshake and its partner labs. The FAQ also says Handshake cannot name the labs because of confidentiality agreements.

Two investor conclusions follow. First, the network's defensible asset is verification and scheduling: the university partnerships that the launch post says supply verified credentials instead of self-reported profiles. Second, the ban on AI tools in contributor work shows what buyers are paying for, human judgment they can trust was not generated by a model, and it puts a quality-control cost on every hour the network sells.

The second supply line: records companies already hold

Handshake's data partnerships buy what no expert produces on request: records of work that companies have already done. The program is a second supply line with its own constraints, and Handshake's pages describe its intent more fully than its terms.

The data partnerships page says Handshake acquires operational records from businesses across industries, that all data is scoped, de-identified and confidential, and that the company determines exactly what is shared. The employer data program post adds that terms are scoped upfront and protected under a nondisclosure agreement, and that Handshake uses such data to build real-world environments where AI models learn to carry out tasks. The licensing page asks for companies with 20+ full-time employees and 3+ years in operation and names Slack, Google Drive and NetSuite among the sources it works with.

What the pages do not publish decides how an investor should value the line: whether a license is exclusive, who owns the prepared dataset after delivery, whether the same records can support environments for more than one lab, and how much revenue the line earns. An archive program whose datasets can serve several buyers behaves differently from one where each acquisition feeds a single commissioned project.

Supply here is bounded by history, not hours. A company with three years of records cannot produce a fourth year on request, so the line grows by adding companies, and each new company brings a fixed stock of past work along with its own rights questions about customer, employee and vendor content.

Network supply against archive supply, criterion by criterion

Network supply and archive supply differ on every criterion that shapes value: what is sold, when it is created, who owns it, what limits it and what a buyer can check before paying. The comparison below uses only what each company publishes, Handshake's pages read on 2026-10-28 and SourceX's live pages, and ranks no one.

Read across a row, not down a column. The fellowship and the data partnerships program sit inside one company yet land on opposite sides of most rows, which is why Handshake AI should be modeled as two lines. SourceX appears as a labeled application of the same test, described only in the words of its live pages:

Network supply against archive supply, criterion by criterion
CriterionHandshake fellowshipHandshake data partnershipsSourceX, from its live pages
What is suppliedExpert hours spent validating model outputsOperational records from companiesRecords a company already keeps, licensed by that company
When it is createdDuring the project, to a lab's specificationBefore the deal, in the normal course of businessBefore the deal; the company confirms the data was not created or altered with AI for the sale
Who owns the resultAssigned to Handshake; program materials owned by Handshake and partner labsNot publishedThe company keeps its original data; SourceX owns the de-identified dataset under the signed agreement, subject to the company's IP
What limits supplyQualified experts' available hours and project matchingCompanies with 20+ full-time employees and 3+ years in operationCompanies with 50 or more full-time employees at peak in the U.S., Canada and the EU, plus listed English-language markets
Published exclusivityNot applicable: work made to orderNot publishedExclusive for AI training; the term is set in each Dataset Deal Confirmation
What a buyer can verify firstCredentials checked through university partnershipsScope set by the company; de-identified; under NDARights confirmed by the company, and data not previously licensed for AI training
Hub diagram with Handshake AI at the center: a talent network supplying expert hours, with work assigned to Handshake and supply limited by qualified hours, and data partnerships supplying company records, with license terms not published and supply limited by years of history; revenue by line is marked not disclosed.
Handshake AI's two supply lines as described on Handshake's own pages, read 2026-10-28.

Could the archives simply feed Handshake's environment work?

Another reading is that Handshake's two supply lines are one business: records bought from companies become raw material for environments that experts then build and grade for labs, all sold as a service. The evidence supports that reading in part.

The employer data program post says acquired data is used to build real-world environments where models learn to execute tasks. The Enterprise AI page lists RL environments among the topics its researchers work on, and the Cleanlab acquisition post says Cleanlab's chief executive, chief scientist and chief technology officer join to lead research, with Handshake describing the deal as deepening its frontier AI research. Taken together, the archives may be an input to commissioned environment work rather than a separate licensing line.

If that reading holds, the right comparison for the data partnerships is an internal cost of goods, not a licensing business with resale economics, and the question becomes how much of each environment's value comes from the acquired records versus the expert hours around them. An investor can test it with two numbers Handshake does not publish: the share of acquired datasets used for more than one lab, and the share of AI revenue earned without paying for new expert hours.

Six ways the two-line reading of Handshake could fail

Six risks could make this reading wrong or make the expert-supply model less durable than its network figures suggest.

At a glance

Six ways the two-line reading of Handshake could fail

  • Undated, self-reported figures: network counts and the partner count appear without dates in the excerpts verified here, so neither growth nor scale can be computed from them.
  • Press figures excluded: revenue and run-rate figures attributed to Handshake AI in press coverage were not read on Handshake's own pages and are left out; an investor should ask for audited numbers.
  • Buyer concentration: the fellowship FAQ says Handshake cannot name its labs because of confidentiality agreements, so the share of revenue from the largest lab is unknown.
  • Quality control: the contractor agreement's ban on AI tools shows that model-generated work is a live risk in human data, and enforcing it is a cost that grows with every hour sold.
  • Rights in acquired records: business archives hold content created by customers, employees and vendors, and a program description cannot show that each partner had the right to license it.
  • Automation of grading: if labs replace more expert grading with automated evaluation, demand for network hours could fall while demand for recorded work context holds, shifting value between Handshake's two lines.

Six numbers to request from Handshake, and the same from SourceX

Ask for numbers that separate hours from records, with dates and definitions. The same questions apply to SourceX.

SourceX sits wholly on the records side of the comparison. Its investor page describes a company that takes licenses to business data straight from the companies holding it and grants licenses to AI developers under its own name, running sourcing, rights checks, contracts, preparation, delivery and payment along the way. Nothing moves to SourceX until the supplying company has accepted the scope and the terms and signed, and that signature is the source of SourceX's right to use, sublicense and distribute the de-identified dataset. Revenue, deal counts and supplier counts are not on its website; company and commercial metrics go to investors directly. SourceX is currently raising capital under Rule 506(c), which limits purchasers to accredited investors, and requests for materials go through its investor page.

  • What share of AI revenue came from fellow hours, from records licensed by companies and from enterprise work in the last fiscal year?
  • How many fellows delivered paid work in the last quarter, and how many hours?
  • How many companies have licensed records, and how many datasets reached more than one lab?
  • Who owns a prepared dataset built from a company's records, and is the license exclusive?
  • What share of revenue comes from the largest lab and the five largest?
  • How is AI-generated content detected in contributor work, and how often is work rejected for it?

Methods note: the Handshake evidence set and what it leaves out

The evidence table and the network-versus-archive comparison are editorial frameworks built for this article on 2026-10-28. The evidence set is Handshake's own domain: the post introducing Handshake AI, the About page, the fellowship FAQ for university partners, the Handshake AI Contractor Agreement, the data partnerships page and its licensing page, the employer data program post, the Enterprise AI page and the Cleanlab and Taro acquisition posts. SourceX's column uses its live how-it-works and investor pages.

Every Handshake source was verified through search excerpts on Handshake's domain, not by opening the page, and a reviewer should open each before release. Publication dates did not appear in the excerpts, so the table dates each entry by the day it was read. Excluded: press estimates of revenue, valuation or growth, investor statements not published by Handshake, job postings, pay rates and the payment figures on the licensing page.

Not measured: Handshake's revenue by line, margins, fellow counts, hours delivered, number of data partners with signed licenses, reuse of acquired datasets and customer concentration. No company is ranked, and SourceX metrics are not part of the comparison because SourceX publishes none.

Frequently asked questions

What does Handshake AI do?

Handshake's pages describe a service that connects graduate-level experts from its career network with AI labs, where fellows provide expert human validation of AI models as independent contractors. Handshake also runs a data partnerships program that acquires operational records from companies and an enterprise AI offering.

Does Handshake publish revenue for Handshake AI?

Not on the Handshake pages read for this article on 2026-10-28. Press coverage has attributed figures to the business, but those are not Handshake's own disclosures and are excluded here; an investor should ask for audited revenue by line.

Who owns the work Handshake AI fellows produce?

Handshake's contractor agreement says contractors irrevocably assign intellectual property in their deliverables to Handshake and keep no right to use the work product. The fellowship FAQ for university partners says materials created in the program are owned by Handshake and its partner labs, while fellows' prior and future work stays theirs.

How is a talent network different from a company-archive licensing business?

A talent network sells new work made to a client's specification, so supply depends on qualified hours. An archive business licenses records a company already created, so supply depends on how many companies hold usable history and can grant the rights. The two lines face different limits and deserve separate valuation.

What should an investor ask Handshake about its data partnerships?

Ask who owns a prepared dataset, whether licenses are exclusive, how many companies have signed and whether any dataset has served more than one lab. Those answers decide whether the program behaves like licensing or like an input to commissioned environment work.

Sources

  • Handshake describes Handshake AI as a service providing expert human model validation for frontier AI development, citing more than 500,000 PhDs, over 50,000 in STEM, and credentials verified through university partnerships. Source
  • Handshake's About page says it connects 18 million knowledge workers, 1,600+ educational institutions and 1 million employers. Source
  • Handshake's fellowship FAQ for university partners says fellows are 1099 contractors working typically 5 to 20 hours a week on projects of two weeks to over three months, that program materials are owned by Handshake and partner labs, and that labs cannot be named because of confidentiality agreements. Source
  • The Handshake AI Contractor Agreement assigns contractors' intellectual property in deliverables to Handshake, leaves contractors no right to use work product, grants a fallback exclusive license, and prohibits AI tools without written authorization. Source
  • Handshake's data partnerships page says it acquires operational records from businesses, and that all data is scoped, de-identified and confidential, with the company deciding what is shared. Source
  • Handshake's data partnerships licensing page seeks companies with 20+ full-time employees and 3+ years in operation, names sources such as Slack, Google Drive and NetSuite, and shows an undated count of 1K+ partner companies. Source
  • Handshake's employer data program post says terms are scoped upfront under NDA, data is de-identified, and Handshake uses it to build real-world environments where AI models learn to execute tasks. Source
  • Handshake's Enterprise AI page pitches model ownership and agent deployment and cites 30M+ professionals, 200+ researchers and work on RL environments. Source
  • Handshake's blog says it acquired Cleanlab, whose CEO, chief scientist and CTO join to lead research, deepening its frontier AI research. Source
  • Handshake's Taro acquisition post says it acquired Taro, a software engineering community, and refers to a network of more than 20 million students, alumni and professionals. Source

Request SourceX investor materials

SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.

See if your company qualifies