Data Labeling Companies: Vendors as Buyers and Partners of Data Holders

Data labeling companies are vendors that add labels, ratings and other human judgments to data for AI developers, and several of them, including Scale AI, Labelbox and SuperAnnotate, also license company records from data holders. This page records 19 such companies, alphabetically, with the legal name on record, the services each lists, its owner and whether it publishes a program to acquire company data (checked 2026-11-09). Six do: Centific, Labelbox (Prism), Scale AI, SuperAnnotate, Toloka and Turing, each on different stated terms.

Labeling vendors sit in two places within AI training data licensing: they supply human data to labs as a service, and a few of them are also counterparties that a data holder can license records to. Their customers are concentrated, as Innodata's Q3 2026 filing shows: its two largest customers made up about 43 % and 31 % of revenue. No vendor pays for its place on the page.

What data labeling and data annotation mean in AI training

Data labeling, also called data annotation, is the work of adding labels, rankings or written answers to raw data so that an AI model can learn from it or be measured against it. The two terms are used interchangeably here. In machine learning generally, a label is the target value a model is trained to predict; for AI training data, the term has widened to cover any human judgment attached to a record.

Typical examples of the work:

• Classification labels: a category assigned to an image, a document or a support ticket.
• Bounding boxes and segmentation: outlines drawn around objects in images and video.
• Transcription: speech converted to text, with speaker and timing marks.
• Rankings: model answers ordered from better to worse.
• Written demonstrations: answers or reasoning written by a person for a model to imitate.
• Ratings against a rubric: scores for accuracy, safety or tone given to model output.

The method behind today's vendors is documented in the research record. The InstructGPT paper of March 2022 describes "about 40 contractors on Upwork and through ScaleAI" who ranked 4 to 9 model responses per prompt. The Llama 2 paper of 2023-07-19 reports 27,540 supervised fine-tuning annotations collected. Rankings of this kind feed reinforcement learning from human feedback, in which a model learns from stated human preferences between its outputs.

Rules for the work exist in at least one jurisdiction. China's Interim Measures for Generative AI Services (Order No. 15 of 2023, issued by the Cyberspace Administration of China and six other agencies, in force since 2023-08-15) require in Article 8 that providers use data labeling with clear rules, assess quality, spot check results and train labelers.

Nineteen data labeling companies on record

The table lists 19 data labeling companies, alphabetically, with the legal name on record, the labeling and data services each lists, whether it acquires company data, and who owns it. It is as of 2026-11-09, and every count in it is the vendor's own claim; sources are the vendor's own page or the filing named.

Reading key:

• Yes: a program page that acquires data from companies, checked 2026-11-09.
• No program found or No program published: none on the pages checked, which is not proof that none exists. No record means no program is on record; No: means the vendor's own page says it does not buy or license third-party datasets, or that none appears on its home page.
• Counts: annotator and customer numbers are the vendor's own claims, never confirmed headcounts.
• Contributor brand: the name under which individuals sign up for tasks; contributor terms are compared on the platforms page.

The legal names of Labelbox and SuperAnnotate read "not stated" because the program pages do not give them. TaskUs has no profile page. Expert marketplaces such as Mercor, micro1 and Handshake are not labeling vendors and sit with the company programs instead.

Which labeling vendors also acquire company data, and on what stated terms

Six of the 19 vendors publish a program that acquires data from companies: Centific, Labelbox (Prism), Scale AI, SuperAnnotate, Toloka and Turing, and each states different terms. They are six of the many company data partnership programs on record, and the only ones run by vendors whose main business is labeling. Every amount below is the vendor's own claim, checked 2026-11-09, never a payment made.

The table compares the six programs on what each acquires, the payment each states, the license terms it states and what it leaves unpublished, as of 2026-11-09.

Labelbox's own Prism pages disagree with each other, and both versions are kept:

• Prism's home page promises a written quote within 48 hours, while its FAQ promises a preliminary valuation within one business day.
• "From first call to first wire in under 21 days" appears on the Prism home page only, not on the FAQ page.

Turing's Project Lazarus page, by contrast, states one payment structure: revenue share with nothing upfront (checked 2026-11-09).

The other vendors in the table state no program to acquire company data. CloudFactory's home page shows no program to buy or license third-party datasets; Innodata, Invisible Technologies, Snorkel AI, Surge AI and TELUS Digital publish none on the pages checked; for Appen, Deccan AI, iMerit, Nexdata, Sama, TaskUs and Unidata no program is on record. Defined.ai, a marketplace rather than a labeling vendor, buys datasets from suppliers for licensing on its marketplace, with no exclusivity and no minimums stated. Payouts and terms of all programs sit side by side in program terms compared, dated.

How a labeling vendor works with a data holder: licensing, preparation and commissioned capture

A data holder meets a labeling vendor in one of four roles: the vendor licenses existing records, prepares them for licensing, commissions new recordings, or sells labeling services for the holder's own data. The roles overlap, and one vendor can take several.

• Licensee and relicensor: the vendor takes a license to the holder's records and offers them to AI developers. Labelbox's Prism and SuperAnnotate describe this role; SuperAnnotate states that the seller approves each buyer.
• Preparer of the data: the vendor carries the work between raw records and a saleable package. Turing states that it covers de-identification, cleanup and packaging, with quality assurance within five business days. Centific states a read-only connection, background extraction and sign-off on an anonymized sample pack before any payment.
• Commissioner of new capture: the vendor pays companies and capture teams to record new material. Toloka's program asks for a partner agreement before recording and blurs faces, palms, fingerprints and tattoos and removes voices.
• Seller of labeling services: the vendor labels the holder's own data for its own use, and nothing is licensed onward. TaskUs lists feedback and evaluation services among its AI offerings, and Appen lists AI training data; neither has a company data program on record.

Preparation is where the vendors differ most. Scale AI states that it anonymizes data inside the company's environment before extraction and uses per-package approval gates. Prism states that medical data needs institutional review board approval or qualified de-identification, and that likeness and voice need explicit timestamped consent. The methods behind such steps, and what they achieve legally, are covered under PII removal and de-identification before licensing.

Who pays labeling vendors: lab customers, revenue concentration and new owners

Labeling vendors are paid mostly by a few AI developers, and their filings show how concentrated that is: Innodata's two largest customers made up about 43 % and 31 % of its Q3 2026 revenue of USD 95.6 million, up 53 % (10-Q, period ended 2026-09-30). The filing does not name the two customers in this record, and Innodata states that five of the "Magnificent Seven" technology companies engage it.

Concentration shows in other filings too. TELUS Digital's 10-K for fiscal 2024 reports that Google was 14.3 % of revenue and its parent TELUS Corporation 24.7 %. Appen's FY25 investor presentation (2026-02-25) states group revenue of USD 230.8 million and adjusts the FY24 comparator for the termination of its Google contract: adjusted 220.9 against statutory 234.3, which gives plus 4.5 % adjusted and minus 1.5 % statutory. Invisible Technologies states 2024 revenue of USD 134 million in a post of 2025-09-16, a claim without a filing behind it.

Ownership also decides who ends up holding licensed data. Four changes are on record:

• iMerit: EXL announced the acquisition on 2026-06-24 and completed it on 2026-07-31, for USD 170 million upfront, up to USD 125 million contingent and USD 15 million deferred (up to USD 310 million), according to EXL's filings.
• TELUS Digital: TELUS Corporation's take-private at USD 4.50 per share, about USD 539 million, was approved on 2025-10-27 and closed on 2025-10-31.
• Scale AI: Meta holds "a minority of Scale's outstanding equity" at a valuation over USD 29 billion (Scale's announcement of 2025-06-12), and Scale does not state the investment amount or the stake.
• Toloka: Toloka states that Nebius no longer has controlling voting power.

Funding rounds, such as Labelbox's Series E announced on 2026-10-14, are tracked under funding and valuations of AI data companies. For a data holder the practical point is narrower: Prism's kill clause includes a seller-defined change of control, so a sale of the vendor is a term to read, not only news.

What a labeling vendor's license terms mean for the records you hand over

The terms that decide what happens to records licensed through a labeling vendor are exclusivity, who else may receive them, deletion, and what happens if the vendor changes hands. Only some of these are published, and the vendors state them differently (checked 2026-11-09).

Prism states that its licenses are exclusive by default, that the seller keeps ownership and continued use, and that buyers pay for stricter exclusivity. A kill clause forces deletion within 30 days after a material breach, a regulatory change or a seller-defined change of control, but models already trained on the data cannot be unlearned. Turing states the opposite starting point: non-exclusive, with Turing as sole seller for the first 90 days and ownership unchanged. How each choice moves price and lock-in is covered under exclusive vs non-exclusive data licenses.

SuperAnnotate states that the company licenses rights for a set period, approves each buyer, can block competitors permanently, and has the data deleted after the agreed period or on request. Scale AI and Centific do not publish exclusivity, term or deletion terms, and Scale does not publish retraining terms. A holder that wants these terms must ask for them in writing before any sample is shared.

Benefits, risks, downsides and alternatives of licensing data through a labeling vendor

Licensing records through a labeling vendor can pay a data holder and hand the preparation work to a specialist, but the license usually reaches labs the holder never contracts with, and the terms are the vendor's to set. The lists below separate what the vendors state from what the record shows.

Benefits

• Payment claims exist: Prism, Scale AI and SuperAnnotate state ranges and Turing states a revenue share, none of them a confirmed payment.
• The vendor carries preparation costs: Turing covers de-identification and packaging, and Centific states no setup fees or deductions.
• Some programs give the seller a say: SuperAnnotate states that the seller approves each buyer.

Risks

• Onward licensing: Prism's license is exclusive by default and trained models cannot be unlearned, so a license cannot be undone in a model already trained.
• Concentration and new owners: a vendor dependent on two customers, or sold to a new parent, can change its priorities within the term.
• Contributor-side complaints: on 2026-10-27 noyb published five GDPR complaints against Scale AI's Outlier platform over trackers, biometric identity checks and access requests.
• Wage litigation: McKinney v. Scale AI (Superior Court of California, County of San Francisco, Case No. CGC-24-620481) settled for a gross USD 12,500,000, with final approval granted on 2026-10-30 (as of 2026-11-09).

Downsides

• Terms are unpublished: Scale AI and Centific do not publish exclusivity, term or deletion terms.
• Payment depends on a result: Centific pays after acceptance and Turing pays each time a record sells.
• Time to first money can be long: Centific states typically six to ten weeks to the first accepted contribution.

Alternatives

• License directly to an AI developer, which removes the vendor's margin and adds the holder's own negotiation.
• Use a marketplace, where the supplier terms are published; Defined.ai states no exclusivity and no minimums.
• Buy labeling services only, keeping the data in-house and licensing nothing onward.
• Do not license at all, and see alternatives to selling data to AI labs, including not selling at all.

A vendor's program is one route among several, and the table above shows that most labeling vendors do not offer it. The choice between them depends on the records, the rights the holder has and the terms it will accept.

How this page is compiled: alphabetical order, check dates, no fees

Vendors appear alphabetically and whether or not any of them pays referral fees; no row is ordered, filtered or described by payout. The claims on this page were checked on the vendors' own pages and in filings on 2026-11-09 unless a sentence gives another date, and the page explains how we source, date and check every figure.

Vendor-written "best" lists, forums and competitor sites are never used as sources, and labeling tools are not reviewed. The page is information, not legal advice.

Where labeling vendors meet buyers, platforms and the seller process

This page records labeling vendors as counterparties of data holders; the pages below cover the other buyers, the contributor side, the seller's choice of counterparty and how the site is funded.

• Human data for AI: the companies, the parent page for companies that supply human judgment.
• Who buys AI training data, every other buyer and program.
• Choosing a buyer for your data, labs, vendors, marketplaces and brokers weighed against each other.
• AI training data marketplaces, including Defined.ai's supplier program.
• Synthetic data companies, companies that generate rather than label data.
• Pay for AI training work, what annotators are paid by platform and field.
• How DataPacts is funded, no vendor pays for a place in the table.
• Dataset vendors and data labs, vendors that sell finished datasets.
• Contributor lawsuits against human data platforms, contributor cases against vendors.

Frequently asked questions

Is data labeling in demand?

Vendors' filings show growth, but no market-wide figure is stated here. Innodata reports Q3 2026 revenue up 53 % (10-Q, period ended 2026-09-30), and TELUS Digital states more than 2 billion labels a year (a claim, checked 2026-11-09). Market estimates, each with its publisher and date, are on the AI training data market page.

Who owns DataAnnotation?

Surge Labs, Inc. operates DataAnnotation as its contributor brand and also sells data services as Surge AI (checked 2026-11-09). Surge's home page lists products, off-the-shelf data, a training lab and an expert workforce, and no company data program. Pay, screening and court records are on the DataAnnotation (Surge Labs) profile.

Is Scale AI a data labeling company?

Scale AI, Inc. is a human-data and data-labeling company. It runs the Outlier contributor platform with its affiliate Smart Ecosystem, Inc., and it publishes a Data Partnerships program that acquires company records (checked 2026-11-09). Meta holds a minority of its equity (announcement of 2025-06-12). The table above carries its full row.

Do data labeling companies hire remote annotators, and are those jobs legit?

Most vendors recruit remote contractors through contributor brands such as Outlier, DataAnnotation, Alignerr, Mindrift and CrowdGen (checked 2026-11-09). Pay claims, screening, payment terms and court records for those brands are compared on the page for AI training jobs and platforms compared.

What is the difference between a data labeling company and a labeling tool?

A labeling tool is software a team uses to label its own data, while a labeling company supplies people and processes. This page records companies as counterparties of data holders. It does not review or compare labeling software, and it names no tools.

Can a data labeling company buy data from my company?

Six of the 19 vendors publish a program that acquires company data: Centific, Labelbox, Scale AI, SuperAnnotate, Toloka and Turing. The other 13 sell labeling services and state no such program on the pages checked. The table above shows which vendor offers what and on which stated terms, and a program is a license, not a sale of the company.