Skip to content

Deal economics

Annotation versus data licensing economics: labor resets, rights repeat

By SourceX Editorial · Reviewed by Noah Loul ·

9 sections

Short answer

Annotation revenue is paid labor: each new unit needs new hours, the buyer usually owns the output and revenue follows the staffing schedule. Licensing revenue is paid rights: the licensor keeps title, a refresh or second license can reuse the asset, and timing follows the license terms. Before applying one multiple, investors should split the two lines and run each through the same shocks.

Value drivers

Annotation

This page covers annotation versus data licensing economics: labor resets, rights repeat.Illustrative

Key takeaways

  • Output ownership is the dividing line: Scale AI's services agreement assigns the output to the customer, so the vendor's reusable asset is its process, not the data.
  • License timing depends on license type: IFRS 15 recognizes a right to access over time and a right to use at a point in time.
  • TaskUs's 2025 annual report warns that helping clients automate may lift revenue now and remove the work later.
  • Licensing revenue steps with partnerships, as IAC's 2025 report shows for People Inc.; annotation revenue moves with ordered volume.
  • Normalize to scope: revenue per labor hour or accepted unit for annotation, per licensed record-year and per dataset version for licensing.

Why do annotation revenue and licensing revenue behave differently?

Annotation revenue is paid labor and licensing revenue is paid rights, and that difference drives how each responds to growth, to a lost buyer and to automation. In annotation, people label, rank, write or check data to a buyer's specification; each new unit of revenue needs new hours, and the buyer usually owns what is produced. In licensing, a rights holder grants permission to use a dataset that already exists; the next license can reuse the same asset, and the licensor keeps title.

Four forces separate the lines: the labor each new sale requires, who holds the intellectual property in the result, when revenue is recognized, and what happens when the buyer's needs change. An investor who blends the two into one AI data revenue figure, or values both at one multiple, loses the information that predicts how the business will behave next year.

This page compares the two on those forces using contracts and filings, then gives a shock test and a set of metrics normalized to the scope each line sells, so a blended company can be split before it is valued.

Who owns the deliverable decides whether revenue can repeat

In a typical annotation contract the buyer owns the output, so the vendor cannot sell the same deliverable twice. Scale AI's Main Services Agreement, as shown on its legal page and checked by search on 2026-10-27, states that "Customer will own the Output and Customer Application." The same agreement bars the customer from selling, licensing or distributing the output to an unaffiliated third party, and lets Scale use customer materials and output to analyze and improve its own products and services.

That clause structure is the economic core of annotation. The vendor's reusable asset is its workforce, tooling, quality process and any improvement rights the contract grants, not the labeled data it delivered. Revenue repeats only when the buyer orders more work.

A license runs the other way. The licensor grants a right to use and keeps ownership, so the same dataset version can be licensed again, refreshed with new records, or extended to a new use, each without rebuilding it. Whether that repeat is available in practice depends on exclusivity and field-of-use terms, which is why the share of licensed scope under exclusive grants belongs in diligence.

Two parallel tracks: on the annotation track each order needs new labor and the output goes to the buyer; on the licensing track one asset is built once and then licensed, refreshed and licensed again.
How each line earns its next sale. Editorial diagram based on the contract and filing evidence in this article.

Labor clocks and license clocks: when each kind of revenue lands

Labor-priced work earns revenue as units or hours are delivered, so annotation revenue follows the staffing schedule; license revenue follows the license terms, and accounting standards make the shape of those terms matter. Under IFRS 15, a license is a right to access the entity's intellectual property, recognized over time, only when the contract requires or the customer reasonably expects the entity to undertake activities that significantly affect that IP and the related conditions in paragraph B58 hold; otherwise it is a right to use the IP as it exists, recognized at a point in time under paragraph B61.

Filings show how that plays out for content. Meredith Corporation's annual report for fiscal 2020 described revenue from functional licenses as recognized at the point in time when access to completed content is granted, and noted that because licensed content is continuously added, the pattern resembles over-time recognition (paraphrased). A refresh schedule, not the calendar, sets the cadence.

Partnership timing shows up in reported results as well. IAC's annual report for 2025 attributes the rise in People Inc.'s Licensing and Other revenue mainly to Apple News+ and syndication partners and to a full year of revenue from an OpenAI partnership that began in May 2024 (paraphrased). Licensing lines tend to step up when a partnership starts and to step down when one ends; annotation lines move with the volume of work ordered.

For diligence, ask a licensing business for its license revenue split between point-in-time grants and over-time access, and for the refresh schedule that drives the point-in-time deliveries. Ask an annotation business for ordered work not yet delivered and how quickly it has historically been completed or cancelled. Each answer shows how much of next year is already decided, in the form that matters for that line.

Annotation versus data licensing economics under five shocks

The quickest way to see the difference is to run both lines through the same events. The table is an editorial framework built from the contract and filing evidence above; it describes usual arrangements, and a particular company's contracts can move any cell.

One row has direct filing support. TaskUs warns in its 2025 annual report on Form 10-K that "In some cases, TaskUs is supporting these initiatives, which may lead to revenue growth in the near term but may ultimately result in the automation of some services that TaskUs currently provides for these clients." Labor that trains a model can be the labor the model replaces.

Annotation versus data licensing economics under five shocks
EventAnnotation services lineData licensing lineMetric to watch
Largest buyer stops orderingRevenue falls as projects wind downContracted fees run to the end of the term; refreshes stopShare of next 12 months' revenue under non-cancellable commitment
Buyer automates the taskDemand for that labor shrinksDemand for source records can persistShare of labor revenue in tasks the buyer is automating
Second buyer wants the same dataNew labor needed; first buyer owns the outputSame version can be licensed again unless exclusiveLicenses sold per dataset version
Labor costs riseMargin compresses unless prices followMostly a one-time preparation cost per versionGross margin per labor hour; preparation cost per version
Buyer asks for exclusivityAlready implied by output ownershipRepeat sales in that field end and must be priced inShare of licensed scope under exclusive grants

Scope-normalized metrics that put both lines on one page

Comparable metrics divide each line's results by the scope it actually sells: labor hours and accepted units for annotation, licensed record-years and dataset versions for licensing. Raw revenue totals reward size, and a large annotation book and a large license catalog can share a revenue figure while behaving in opposite ways.

Ask for each line separately. A company that reports only a blended AI data revenue number, or that cannot say which contracts transfer ownership of the output, has not yet shown which business it is.

Scope-normalized metrics that put both lines on one page
MetricAnnotation definitionLicensing definitionWhat it reveals
Revenue per unit of scopeRevenue per accepted labeled unit or labor hourRevenue per licensed record-yearPricing power per unit of input
RepeatabilityShare of revenue from repeat orders by the same buyerLicenses and refreshes sold per dataset versionWhether growth needs new inputs
CommitmentOrdered work not yet deliveredContracted license fees not yet recognizedVisibility, and what can be cancelled
ConcentrationTop buyer's share of annotation revenueTop licensee's share of license revenueExposure to one decision
OwnershipShare of contracts assigning output to the buyerShare of scope licensed exclusivelyWhether the asset can be reused

Does the label platform hide a labor business?

Words like platform and marketplace can sit on top of economics that are mostly labor. A company may describe a data platform while most of its revenue comes from managed teams producing labels the buyer owns; another may call itself a services firm while earning most of its margin from relicensing a dataset it controls.

Four questions sort the revenue regardless of the label: who owns the rights in what is delivered, who performs the labor, who supplies the tooling, and who captures the recurring value when the buyer comes back. If the answers are buyer, vendor staff, vendor and nobody until the next order, the line is services. If they are the licensor, a one-time preparation team, the licensor and the licensor again on refresh, it is licensing. Revenue that mixes the two should be split before any multiple is applied.

Illustrative: a company sells evaluation datasets to model developers and calls the product a data platform. If each dataset is written to one buyer's specification, delivered once and owned by that buyer, the four answers are buyer, vendor staff, vendor and nobody until the next order, so the line is services however it is marketed. If the same company holds a benchmark set it controls, refreshes it quarterly and licenses the current version to several developers, that slice is licensing. Most such companies hold some of each, and the split, not the product name, sets the valuation question.

Where the two-line distinction blurs or fails

The distinction is useful because it predicts behavior, and it fails where contracts mix the two forms. Six cases deserve attention.

  • Vendors with reuse rights. Scale's agreement lets it use customer materials and output to improve its own products, so an annotation vendor can accumulate an asset even when the deliverable belongs to the buyer.
  • Licenses that need heavy custom preparation for each buyer behave like services: the second sale is cheap only if the asset is reusable as delivered.
  • Automation. The TaskUs risk factor describes near-term growth from supporting client automation that may later remove the work; annotation growth can be self-limiting.
  • Partnership dependence. The IAC filing shows licensing revenue moving with the start of a single partnership; a licensing line with few licensees carries renewal risk of a different kind.
  • Accounting form. A right-to-use license recognized at a point in time can make a single quarter look large; read the license type before reading the trend.
  • Substitutes for both. Synthetic data and openly licensed corpora can reduce demand for some labeling and some licensed content at once, so neither line is immune.

Questions for any data company, and where SourceX's model sits

SourceX sits on the licensing side of this framework, with a preparation step inside each deal; it is described here only from its published pages. It helps companies turn selected business records into licensed datasets for AI developers and runs sourcing and qualification, rights review, contracting, data preparation, delivery and payment. The company approves the scope and terms and signs before any data is delivered, after which SourceX may use, sublicense and distribute the de-identified dataset as the agreement allows. De-identification is performed by SourceX directly or through its processing partner, which is the labor component the shock test would examine.

SourceX is raising from accredited investors under Rule 506(c) and shares company and commercial metrics with investors directly through its investors page; none appear here. The questions below are the ones to put to any AI data company.

  • Which contracts assign ownership of the output to the buyer, and what share of revenue do they carry?
  • For each dataset version, how many licenses and refreshes have been sold, and at what preparation cost per version?
  • How much of next year's revenue is under non-cancellable commitment, by line?
  • Which revenue depends on tasks your buyers are training models to automate?
  • What share of licensed scope is exclusive, and how was the lost repeat value priced?

How this comparison was assembled, and what it leaves out

This is a framework built from contract language, accounting text and filings, not a measurement of any company's revenue split. Sources: Scale AI's Main Services Agreement page; TaskUs's 2025 annual report on Form 10-K and 2026 proxy statement; IFRS 15 paragraphs B58 and B61; Meredith Corporation's fiscal 2020 annual report; and IAC's 2025 annual report. All were checked through search excerpts on their own domains on 2026-10-27, and the filing statements are paraphrased except where quoted.

The TaskUs proxy statement adds one figure: it describes its AI Services business growing nearly 59 percent in 2025 (paraphrased from the excerpt). AI-related service work can grow quickly, which is a reason to measure its durability rather than assume it.

Not measured: margins, revenue mix or contract terms at any named company beyond what the cited text states, prices, and the size of either market. The shock table and the metric definitions are editorial and should be tested against a company's actual contracts.

Frequently asked questions

Is data annotation revenue recurring?

It recurs only when the buyer orders more work. Because the buyer usually owns the labeled output, the vendor cannot resell it, so each new period of revenue needs new labor.

Why can data licensing revenue repeat without new work?

The licensor keeps ownership and grants a right to use, so the same dataset version can be licensed to another buyer or refreshed with new records. Exclusivity and field-of-use terms can limit that repeat, so check them before assuming it.

When is license revenue recognized?

Under IFRS 15, a license that is a right to access the entity's intellectual property is recognized over time when the conditions in paragraph B58 are met; otherwise it is a right to use, recognized at a point in time. Content licenses with continuous additions can look like over-time revenue even when each delivery is recognized at a point in time.

Should investors use one multiple for an AI data company with both lines?

Not until the lines are split. Annotation revenue behaves like labor services and licensing revenue like rights income, so they respond differently to buyer loss, automation and labor costs.

What single metric best separates the two?

The share of revenue from contracts that assign ownership of the output to the buyer. Paired with licenses sold per dataset version, it shows how much revenue can repeat without new labor.

Sources

  • Scale AI's Main Services Agreement gives the customer ownership of the output, bars the customer from reselling it to unaffiliated third parties and lets Scale use customer materials and output to improve its products. Source
  • TaskUs's 2025 annual report warns that supporting client automation initiatives may bring near-term revenue growth but may ultimately automate services it provides. Source
  • TaskUs's 2026 proxy statement describes growth of nearly 59 percent in its AI Services business for 2025. Source
  • IFRS 15 treats a license as a right to access recognized over time only when paragraph B58's conditions are met; otherwise it is a right to use recognized at a point in time (B61). Source
  • Meredith's fiscal 2020 annual report recognized functional license revenue when access to completed content was granted and noted that continuous additions make the pattern resemble over-time recognition. Source
  • IAC's 2025 annual report attributes People Inc.'s higher Licensing and Other revenue mainly to Apple News+, syndication partners and a full year of OpenAI revenue from a partnership that began in May 2024. 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