Skip to content

AI data market

Labelbox data platform business analysis: three meters, one workflow

By SourceX Editorial · Reviewed by Noah Loul ·

9 sections

Short answer

A Labelbox data platform business analysis should split three lines Labelbox describes itself: a labeling platform metered in Labelbox Units, managed labeling through its Alignerr workforce with billing modes by task or hour, and Prism, which licenses companies' operational data for an upfront fee plus a revenue share. Its SDK changelog puts the workflow at the center: of 76 items added since July 2024, 4 concern managed labeling and none data licensing.

License terms

Labelbox data platform

This page covers labelbox data platform business analysis: three meters, one workflow.Illustrative

Key takeaways

  • Compare Labelbox line by line: platform consumption in Labelbox Units, managed labor by task or hour, and Prism data licensing by datasets and downstream use.
  • Labelbox publishes each line's meter but no revenue, margins, customer counts or line mix, so those fields stay blank.
  • Of 76 items added to Labelbox's Python SDK between 29 July 2024 and 22 October 2026, 62 are workspace features, 10 maintenance, 4 managed labeling and 0 data licensing.
  • Labor attaches to a project: the SDK starts a labeling service as a workforce record on the project, which makes the workflow the natural comparison anchor.
  • Bundled contracts could make the split an allocation choice, so ask for revenue by meter before allocation and for the share of platform customers that buy labor.

What should a Labelbox data platform business analysis compare?

It should compare three businesses that Labelbox itself describes, each on its own meter: a labeling platform consumed in Labelbox Units, managed labeling performed by a workforce Labelbox manages through its Alignerr package, and Prism, a program that licenses companies' operational data. Blending them into one growth story, or comparing the whole company with a pure software vendor or a pure data vendor, mixes units that do not add.

The evidence for a useful answer is narrow on purpose. Labelbox is private, so this page uses only what Labelbox publishes: its documentation, terms, blog, Prism site and the open-source Python SDK with its dated changelog, all read on 28 October 2026. Revenue, recurring revenue, margins, customer counts and the mix between lines are not in any of those sources, so they stay blank rather than estimated.

The title's claim is a hypothesis to test: the workflow software is the part every other line runs through, so it is the right anchor for comparison, while labor and data have to be priced by their own yardsticks.

Labelbox's own evidence, dated, with the blanks left blank

Labelbox publishes enough to tell its lines apart and to name each meter, and nothing that sizes any of them. Its billing guide describes the Labelbox Unit as a normalized unit of data that is consumed as customers use the products, with Catalog charged monthly and Annotate and Model charged once per data row. Its enterprise terms, updated 11 December 2024, say the customer keeps its rights in the data it submits, while letting Labelbox derive aggregated anonymous data from it and use that for its business purposes, including improving its products. A Labelbox blog post introduces a request button for labeling services that takes a project through Requested, Calibration, Production and Complete stages. The Prism site describes paying companies for real-world operational data from systems such as Salesforce, Jira, Slack and SAP, with an upfront licensing fee at signing, a revenue share on downstream use, and exclusive licenses by default.

The table puts each line beside its meter, the yardstick that fits it, and the fields that stay undisclosed:

Labelbox's own evidence, dated, with the blanks left blank
LineWhat Labelbox publishes (read 2026-10-28)Meter it impliesFitting yardstickNot disclosed
Labeling platformLabelbox Units consumed by product; Catalog monthly, Annotate and Model once per data row (billing guide)Units consumed per data rowUsage-based software: consumption and its retentionRevenue, units consumed, prices realized under enterprise contracts
Managed labelingRequest flow from a project, Requested to Complete (blog); SDK workforce billing modes by task, by hour and by task per turnTasks or hours of human workManaged workforce services: hours, utilization, margin per hourHours delivered, services revenue, workforce cost
Prism data licensingUpfront licensing fee at signing plus revenue share on downstream use; exclusive by default (Prism site, undated)Licensed datasets and their downstream useData licensing: rights, reuse, payment obligationsDatasets licensed, amounts paid, licensing revenue
Customer data termsCustomer keeps rights in its data; Labelbox may derive aggregated anonymous data for its business purposes (terms updated 2024-12-11)None: a limit on reusing customer dataShows the platform does not take customers' data rightsWhether aggregated data is used beyond product improvement
Company levelSDK changelog, 76 added items from 2024-07-29 to 2026-10-22; funding posts not usedProduct effort, not revenueNone until audited figures existRevenue, recurring revenue, margins, revenue by line, customers, headcount

Where Labelbox's product effort goes: 76 changelog additions

Most of what Labelbox ships to programmatic users is workflow software: of 76 items its Python SDK changelog lists as added between version 3.76.0 (29 July 2024) and version 7.17.0 (22 October 2026), 62 are workspace product features, 10 are SDK maintenance, 4 concern the managed labeling service and none concern data licensing. Eighteen of the 62 name a workflow or quality-control term such as a queue, review, rework, consensus, benchmark, performance, label version, issue or assignment. Recent examples include label version history recorded at every hand-off (7.16.0, 20 October 2026) and clearing every reservation in a task queue (7.17.0, 22 October 2026).

The 4 labor items are the plumbing that connects the two businesses. Version 3.76.0 added methods to get, request and check the status of a project's labeling service; 3.77.0 added a request call that validates the project before asking for labelers; 3.78.0 added labeling service dashboards. In the SDK source, starting a labeling service is described as equivalent to the interface action Request Specialized Labelers, and it creates a project boost workforce record, the same object the SDK's separate Alignerr package manages with statuses from set-up to complete and billing modes by task, by hour and by task per turn, plus a separate customer billing mode.

A changelog is a narrow window. It covers what developers can call, not what sales teams sell, and Prism does not appear in it at all. What it does show, dated and reproducibly, is that labor and data hang off a project object whose development continues week by week.

A unit chart of 76 squares for the items added to Labelbox's Python SDK from July 2024 to October 2026: 18 features naming a workflow term, 44 other features, 10 maintenance, 4 managed labeling and 0 data licensing, above three meter cards: Labelbox Units for the platform, per task or per hour for labeling services, and upfront fee plus revenue share for Prism.
Labelbox Python SDK changelog at commit 60f1c41, classified by the article's stated rules; meters from Labelbox's own pages, read 2026-10-28.

Why the workflow is the comparison anchor

Because both other lines attach to it. A labeling service is requested for a project and validated against that project's setup, and the workflow definitions in the SDK name the steps any labeler passes through: initial labeling, review, rework, logic branches, automated quality checks and done. Performance reports added in October 2026 can be filtered by member organization, which the changelog describes as covering other organizations such as workforce providers. The software is built to manage human work whoever supplies it.

That makes usage-based software the fair comparison for the platform line, measured by consumption and how it holds up over time, not by headcount or dataset sales. It also explains why the line split matters for valuation: a customer can run its own labelers, a vendor's labelers or Labelbox's Alignerr workforce through the same workflow, and only the last produces services revenue.

The strongest counter-reading is that the platform is a funnel for services. Labelbox's own materials frame its offering around producing training data, and its July 2024 product post is titled as building a data factory for generative AI and frontier models. If labor is where the money is, the workflow is cost of delivery rather than the product. Labelbox does not publish the split that would settle it.

Labor and licensed data need their own yardsticks

Managed labeling should be judged like a workforce business. The SDK's billing modes by task and by hour, and a separate country rate multiplier setting, point to the measures that decide its economics: hours or tasks delivered, utilization, the spread between what customers pay and what the workforce costs, and how much of that work repeats for the same customer. None of those figures is published.

Prism should be judged like a data licensor. An upfront fee plus a revenue share ties Labelbox's payments to companies to the downstream sales it makes, and exclusivity by default limits what those companies can do with the same data elsewhere. The questions are about rights and reuse: how many datasets, how often each is licensed again, and how revenue share obligations are recorded.

SourceX sits in the same category as Prism, with a different structure on its live pages. Its investors page describes taking licenses to business data straight from the companies that hold it and relicensing that data to AI developers under SourceX's own name, with SourceX handling sourcing, rights review, contracting, preparation, delivery and payment. Its how-it-works page says it usually pays a one-time fee for a dataset with no royalties or usage fees, and that licenses are exclusive for AI training, with the term set in each deal's Dataset Deal Confirmation. Neither structure is better in general; they create different obligations, and an investor should read each against the costs it creates.

The strongest objection: one customer, one contract, three lines

The objection is that the three lines may not be separable in practice. A frontier lab can buy platform access, labeling services and data in one relationship, and Labelbox's enterprise terms already cover labeling services alongside the platform. If prices are bundled, any revenue split is an allocation choice, and comparing the platform alone with software peers would flatter or punish it depending on that choice.

The objection holds where bundles dominate, and it changes the question rather than ending it. Ask for revenue by meter, consumption units, hours or tasks and licensed datasets, before any allocation, and for the share of platform customers that also buy labor. Those numbers, not the label on the business, decide whether the workflow anchors the value or only carries it.

What could make this reading of Labelbox wrong

Several gaps could reverse the reading, and each is visible in the sources themselves:

  • The changelog measures programmatic features, not revenue; managed labeling and Prism are sold through the interface and sales, so their weight can be far larger than 4 of 76 items suggests.
  • Prism's structure was read only through a search excerpt of an undated site; its terms could differ from that summary or change without notice.
  • Bundled enterprise contracts may make the three-line split an accounting choice rather than an economic fact.
  • The enterprise terms let Labelbox derive aggregated anonymous data and use it for business purposes; if that data were sold or licensed, the platform would hold a data asset this page treats as absent.
  • Labelbox is private: no audited statements were found, and funding announcements on its blog are financing events, not revenue, so none of their figures is used.
  • Pricing pages and documentation change; every row of the evidence table carries its read date for that reason.

Questions to ask Labelbox, any data platform and SourceX

The same questions work for any company that mixes software, labor and data, SourceX included. For readers weighing SourceX itself: the company is seeking capital only from accredited investors, under Rule 506(c), and gives investors its company and commercial metrics directly. Its investors page takes requests for materials, and a question from this list is a reasonable place to start.

Each question below asks for a figure tied to one meter:

Checklist · 6 items

  1. Revenue by meter for the last eight quarters: consumption units, labor hours or tasks, and licensed datasets, before any allocation.
  2. The share of platform customers that also buy labeling services, and the reverse.
  3. Gross margin per labor hour or task, and how country rate multipliers change it.
  4. For data licensing: datasets licensed, the number licensed more than once, and how revenue share or one-time payments to suppliers are recorded.
  5. Whether aggregated anonymous data derived from customers is used beyond product improvement, and under which contract clause.
  6. Concentration: the share of revenue from the largest customers in each line.

0 of 6 ticked. Tick each item as you go.

Methods note: sources, dates and the changelog count

Sources were read on 28 October 2026. Labelbox's billing guide, enterprise terms, labeling services blog post, July 2024 product post, Prism site and Series E post were verified through search excerpts restricted to Labelbox's domains, because direct fetches are blocked in this environment; a reviewer must open each. The Python SDK was read from a clone of Labelbox/labelbox-python at commit 60f1c41 (26 October 2026). SourceX's investors page was fetched live; its how-it-works page was read from the repository route that renders it after the live fetch failed.

Changelog count: versions dated 29 July 2024 to 22 October 2026 inclusive (42 versions, 33 with added items). An item is a top-level bullet in an Added section, a bullet directly under a version heading before any section heading, or text following Added in a section heading, in which case the lines beneath that heading are not counted separately. Item text is normalized by removing link targets, turning backticks, underscores and periods into spaces and splitting camel case. Categories are assigned by the first match of word-start patterns: managed labeling service (labeling service, boost, alignerr, specialized labelers), data licensing or acquisition (prism, licens, data acquisition, data partner), SDK maintenance (python followed by a digit, pydantic, mypy, numpy, deprecat, datadog, graphql client, automation test, ssl, self-signed, dependenc), and workspace product feature for everything else. The 18 workflow mentions are product items whose text contains workflow, task queue, queue, reservation, review, rework, consensus, benchmark, performance, label version, issue or issues, assign, auto qa or autoqa, overview, or sync external project; a mention is a match at the start of a word, so assignable counts as assign, not a judgment that the item is a workflow feature. The script and its item-by-item output are bin/labelbox_changelog_mix.py and bin/labelbox_out.txt.

Not measured: revenue or volume of any line, Prism's contract terms beyond the site summary, customer counts, and anything from press coverage. No company is ranked.

Frequently asked questions

Is Labelbox a software company or a services company?

By its own materials it is both, plus a data licensor: a labeling platform metered in Labelbox Units, managed labeling through its own workforce, and the Prism data licensing program. It does not publish revenue by line, so the mix is unknown and each line should be compared on its own meter.

What is a Labelbox Unit?

Labelbox's billing guide describes it as a normalized unit of data consumed as customers use its products, with Catalog charged monthly and Annotate and Model charged once per data row. It is a usage meter, which is why the platform compares best with usage-based software.

What is Labelbox Prism?

Prism is Labelbox's program for licensing companies' real-world operational data for frontier AI. Its site describes an upfront licensing fee at signing, a revenue share on downstream use and exclusive licenses by default, read here through a search excerpt of an undated page.

Does Labelbox publish its revenue or growth?

Not in any source read for this page. Labelbox is private, and its funding announcements are financing events rather than revenue, so this analysis leaves revenue, recurring revenue, margins and growth blank.

How does SourceX's model differ from Prism's?

Both license business data, with different structures on their own pages. SourceX's live pages describe licensing data straight from the companies that hold it and relicensing it under its own name, usually paying a one-time fee with no royalties, with exclusivity for AI training and the term set in each Dataset Deal Confirmation; Prism describes an upfront fee plus a revenue share.

Sources

  • Labelbox's billing guide describes the Labelbox Unit as a normalized unit consumed as customers use its products, with Catalog charged monthly and Annotate and Model charged once per data row. Source
  • Labelbox's enterprise terms, updated December 11, 2024, say the customer retains rights in data it submits, cover labeling services, and let Labelbox use customer data internally to generate aggregated anonymous data that it may use for business purposes, including improving its products. Source
  • A Labelbox blog post introduces a request button for labeling services with Requested, Calibration, Production and Complete stages, the same workflow for managed contracts and self-serve. Source
  • Labelbox's Prism site describes paying companies for real-world operational data, with an upfront licensing fee at signing, a revenue share on downstream use and exclusive licenses by default. Source
  • Labelbox's July 2024 product post is titled Building the data factory for GenAI and frontier models. Source
  • Labelbox has published financing announcements on its blog, including a Series E post; this article uses none of their figures. Source
  • Labelbox's Python SDK changelog lists dated versions; 76 added items between 3.76.0 (2024-07-29) and 7.17.0 (2026-10-22), including labeling service methods in 3.76.0 to 3.78.0 and member-organization filters for performance reports in 7.14.0. Source
  • The SDK's LabelingService starts a service with a mutation that upserts a project boost workforce and describes this as equivalent to the interface action Request Specialized Labelers; request() validates the project before requesting labelers. Source
  • The SDK's Alignerr package defines project boost workforce statuses from SET_UP to COMPLETE, billing modes BY_TASK, BY_HOUR and BY_TASK_PER_TURN, a customer billing mode and a country rate multiplier setting. Source
  • The SDK's workflow definitions name initial labeling, initial rework, review, rework, logic, done, custom rework and AutoQA nodes. Source
  • SourceX's investors page says it licenses business data directly from companies and licenses it onward to AI developers in its own name, managing sourcing, rights review, contracting, preparation, delivery and payment, and shares company and commercial metrics with investors directly. Source
  • SourceX's how-it-works page says it usually pays a one-time fee for a dataset with no royalties or usage fees, and that licenses are exclusive for AI training with the term set in each Dataset Deal Confirmation. 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