Data Labeling Companies: A Buyer’s Guide to 26 Vendors (2026)

Last updated October 2026. Vendor facts, certifications, pricing and review figures were checked on 6 August 2026, and every ownership, funding and pricing fact was re-checked on 15 October 2026. This category changes fast, and we re-check every figure quarterly.

Scale AI’s largest customer moved to end the relationship within days of Meta taking a reported 49% stake. Appen’s group operating revenue fell 14.2% the year after Google terminated its contract. Mercor went from a $2 billion valuation to $10 billion in eight months. Three data points, one market that reorganised itself inside eighteen months, and almost none of the guides ranking for data labeling companies reflect any of it.

This is written for the person choosing and defending an annotation vendor: the ML lead, the head of data, the AI product owner who has to put three names in front of a procurement committee and explain the shortlist. It is not a guide to annotation work. If you are looking for a job labeling data, the rest of this page will not help you.

The numbering is not a ranking and nobody paid for placement. What you walk out with: 26 vendors with sourced core facts and public review figures printed with their sample sizes, the three questions that shorten a shortlist faster than any feature comparison, per-modality cost benchmarks, and a pilot protocol plus an SLA clause skeleton you can lift straight into a procurement pack.

The 26 data labeling companies compared

Twenty-six entries is more than anyone holds in their head. One scan view first, on the axes that actually shorten a shortlist. The sections below carry the detail.

#	Category	Company	Delivery model	Workforce model (geography disclosed?)	Ownership / independence	Pricing posture	Public customer reviews	Best for
1	Enterprise / frontier-lab	Scale AI	Hybrid	Open gig crowd (No)	Meta non-voting minority stake (reported 49%), June 2025	Quote-only	None usable (G2 n=1)	Large frontier-scale programmes, ownership conflict permitting
2	Enterprise / frontier-lab	Appen	Hybrid	Open vetted crowd + secure facilities (Partial)	Public, ASX: APX, no controlling shareholder	Quote-only	G2 4.2 (n=35); Capterra 4.1 (n=40)	Multilingual breadth with a checkable public owner
3	Enterprise / frontier-lab	iMerit	Managed + platform	Employed in-house (Yes)	Acquired by EXL (Nasdaq: EXLS), Aug 2026	Quote-only	G2 4.7 (n=13), small sample	Regulated work needing employed annotators
4	Enterprise / frontier-lab	Sama	Managed + platform	Employed in-house, impact sourcing (Yes)	Private, last round 2021	Quote-only	G2 4.6 (n=11), small sample	Impact-sourced delivery with published office addresses
5	Enterprise / frontier-lab	Innodata	Managed + platform	Employed specialists, 8 countries (Yes)	Public, Nasdaq: INOD; top two clients ~71% of Q2 2026 revenue	Quote-only	None usable (Trustpilot n=2)	Employed expert teams from a listed vendor
6	Expert-data marketplace	Surge AI	Platform + managed expert	Credential-gated experts (No)	Bootstrapped to 2025, then first raise	Quote-only	None (G2 n=0)	Credential-gated RLHF and evaluation work
7	Expert-data marketplace	Mercor	Marketplace	Vetted experts (No)	VC-backed, $10B, October 2025	Quote-only	None usable; Trustpilot 4.4 (n=640) is contributor-side and solicited	Hourly domain-expert capacity at speed
8	Expert-data marketplace	Turing	Managed + platform	Vetted experts (No)	VC-backed, ~$2.2B	Quote-only	None usable; Trustpilot 2.8 (n=194) contributor-side	Coding, reasoning and STEM training data
9	Expert-data marketplace	Handshake AI	Fellowship network	Credential-gated, US only (Yes)	VC-backed, $3.3B (2022)	Quote-only	None. Circulating figures are the recruiting product	US-only academic expert data
10	Expert-data marketplace	Micro1	Platform + expert network	Vetted experts (No)	VC-backed; $500M Sept 2025, reported $4B Sept 2026	Quote-only	None; Trustpilot n≈9,333 is candidate-side	RL environments and agent evaluation
11	BPO and crowd-scale	TELUS Digital	Hybrid at scale	BPO + open crowd (Yes)	Subsidiary of TELUS Corp; privatization closed 31 Oct 2025, TIXT off the NYSE from 11 Nov 2025	Quote-only	None (G2 n=0)	500+ language coverage at volume
12	BPO and crowd-scale	TaskUs	Managed BPO	BPO delivery centres, 31 sites (Yes)	Public, Nasdaq: TASK; take-private voted down 8 Oct 2025	Quote-only	G2 4.0 (n=11), small sample	Certification-heavy, audit-heavy programmes
13	BPO and crowd-scale	LXT (incl. clickworker)	Managed + open crowd	Managed + 10M crowd (Yes)	Private; clickworker acquired by LXT, Dec 2024	Quote-only	None; Trustpilot 2.6 (n=6) contributor-side, and clickworker 2.6 worker-side	Named secure facilities and on-site deployment
14	BPO and crowd-scale	Toloka	Platform + marketplace	Crowd to vetted experts (Partial)	Nebius Group; Bezos Expeditions round, May 2025	Structure published, rates not	None usable; G2 profile carries no rating	Crowd breadth moving toward vetted experts
15	Managed mid-market	CloudFactory	Managed + platform	Employed in-house (Yes)	Private, VC/PE, last round 2019	Quote-only, published as such	G2 4.5 (n=11), small sample	A dedicated employed team on a mid-market budget
16	Managed mid-market	Cogito Tech	Managed service	BPO, three named Noida centres (Yes)	Private, bootstrapped	Quote-only, no pricing page	None printed; G2 sample size unresolved	Named India delivery centres, if you verify the certs yourself
17	Managed mid-market	Label Your Data	Hybrid	Employed annotator teams (Yes)	Stated subsidiary of SupportYourApp	Published	G2 4.9 (n=15), small sample	Published per-object rates and no minimums
18	Annotation platform	Labelbox	Platform + expert crowd	Vetted expert crowd (Partial)	VC-backed, Series E Oct 2026	Quote-only since the pricing page was removed	G2 4.5 (n=48)	Platform-first teams wanting an expert network attached
19	Annotation platform	SuperAnnotate	Platform + expert network	Vetted experts (No)	VC-backed, Series B extension 2025	Quote-only	G2 4.9 (n=284), the largest sample here	Teams that want the deepest review evidence in the category
20	Annotation platform	Encord	Platform-first	Partner/expert network (No)	VC-backed, $60M Series C, Feb 2026	Quote-only	G2 4.8 (n=61)	Physical AI, robotics and AV data workflows
21	Annotation platform	V7 (V7 Darwin)	Platform-first	Contracted specialists (No)	Independent, VC-backed, last round 2022	Quote-only	G2 4.8 (n=54)	Medical and document workflows with a full security page
22	Annotation platform	Kili Technology	Platform-first hybrid	Not disclosed (No)	VC-backed, Balderton-led Series A	Partially published	G2 4.7 (n=53)	European enterprises wanting on-prem or air-gapped
23	Open source / self-serve	Label Studio (HumanSignal)	OSS + SaaS	Self-hosted or vendor-managed (Partial)	VC-backed, $30M reported	Published ($99/mo Starter base)	None usable (Capterra n=2)	Apache 2.0 self-hosting with a paid cloud tier
24	Open source / self-serve	CVAT.ai	OSS + managed cloud	Self-hosted or first-party teams (No)	Independent, no announced round	Self-serve tiers published	G2 4.5 (n=31), read live	Fully published pricing and MIT-licensed core
25	Open source / self-serve	Roboflow	Platform	Undisclosed for labeling add-on (No)	VC-backed, $40M Series B, Nov 2024	Fully published	G2 4.7 (n=108)	Computer vision teams wanting self-serve labeling
26	Cloud-native	AWS SageMaker Ground Truth	Cloud-native, closed to new customers 30 Jul 2026	Marketplace vendors or your own; MTurk reported closed Sept 2026 (No)	AWS / Amazon (Nasdaq: AMZN)	Fully published	G2 4.1 (n=19)	Bringing your own workforce onto rented tooling. Existing AWS customers only since 30 July 2026

Two findings from our own same-day sweep carry into any evaluation. First, the longer the certification list on a marketing page, the less likely any of it is evidenced. Second, a collection artifact quietly poisons side-by-side rating comparisons: Mercor pays for a Trustpilot subscription and actively solicits reviews, with 637 of its 640 reviews arriving inside the last twelve months, producing a 4.4. Turing sits at 2.8 (n=194) and LXT at 2.6 (n=6), and neither solicits at all. A material part of that gap is a collection-method difference, not a quality difference, and every roundup we read prints those three numbers together without saying so.

Attribution: Forage AI research team, certification and review sweep of 28 providers, 6 August 2026

How we evaluated these companies

Nine criteria, published up front, applied identically to all 26 entries. Where a vendor publishes nothing against a criterion, this guide prints the absence rather than leaving the cell blank.

Criterion	What we checked
Delivery model	Platform, managed service, or workforce supply
Workforce model	Employed, contractor, crowd, or expert network
Geography disclosure	Whether delivery-centre locations are disclosed at all
Ownership and independence	Who owns the vendor, and whether that creates a conflict with your lab
Modality and domain fit	Which data types and verticals they actually staff for
Published quality method	Whether a gold-set, consensus or IAA method is documented
Security and compliance	Framework, Type, and scope: read together, not as a badge
Pricing posture	Published, on request, or enterprise-only
Public review evidence	Printed with its sample size, or recorded as unchecked

One finding is worth stating before the list, because it changes how the compliance rows should be read. Across 28 vendor security pages read on the same day, breadth of certification claims ran inverse to strength of evidence. The vendors claiming the most frameworks were the least likely to publish a verifying artefact or state a scope. Certification is a floor, and the floor is lower than most procurement checklists assume.

Forage AI published this article and does not sell data labeling or annotation in any form, which is why it is not one of the 26 entries. An earlier version carried it at #5 with a not-for-you verdict; a roster slot spent on a company you cannot hire for the job is a slot wasted, so it went to Innodata instead.

What the review evidence is worth

Customer-side review volume in this category is close to zero: Scale AI n=1 on G2, Surge AI n=0, TELUS Digital n=0, Innodata n=0. Every rating below is printed with its sample size and access date, and rows that were not checked say so rather than claiming an absence nobody verified. Trustpilot here is a worker channel, not a customer one: Appen carries 4.1 from buyers and 1.4 from workers (n=366).

The 26 data labeling companies, by category

Those are the market conditions the following 26 entries were assessed under.

The numbering is not a ranking. Categories are ordered by buyer-decision logic, and within each category entries are ordered by market prominence rather than preference. Nobody paid for placement.

Every entry carries the same 14-row core-facts table and the same 8-row review table, so the roster reads as one comparison rather than a pile of unrelated specs. Filter on delivery model first and workforce geography second, because geography is a hard disqualifier the moment you carry a data-residency obligation. Modality and domain fit come third; review scores come last.

Absence of published pricing, geography or certifications is printed as a finding rather than left blank. A rating without its sample size is not a claim.

A. Enterprise and frontier-lab providers

The five vendors that take the largest programmes, and the tier where the 2025 and 2026 ownership changes landed first.

1. Scale AI

Attribute	Detail
Delivery model	Hybrid: platform plus managed service
Workforce model	Open gig crowd, recruited through wholly-owned Remotasks and Outlier
Workforce geography	Not disclosed on scale.com. Third-party reporting (Washington Post, 2023) places the crowd in Kenya, the Philippines and Venezuela, at roughly 240,000 workers
Ownership / independence	Meta holds a non-voting minority stake, closed 13 June 2025; press reported it at 49% for ~$14.3B, valuing Scale at ~$29B, and neither company discloses the percentage. Founder Alexandr Wang left to lead Meta Superintelligence Labs and remains on Scale’s board. Francis deSouza became CEO on 10 August 2026, succeeding interim CEO Jason Droege
Best for	Frontier-scale programmes where volume and tooling maturity outweigh the ownership question
Not for	Any buyer who competes with Meta, or whose confidentiality posture cannot absorb a strategic investor with a reported 49% economic interest
Modalities	Image, video, 3D/LiDAR, text, audio, document, multimodal
Specialist domains	Defence and government, autonomous vehicles, frontier LLM training and evaluation
Core features	Data Engine, Nucleus, Scale GenAI Platform, Remotasks and Outlier contributor networks
Quality method	Multi-pass review and model-assisted pre-labeling; no published IAA methodology
Security & compliance	SOC 2 Type 2, ISO 27001:2022, DoD IL4 Provisional Authorization, FedRAMP High Authorized (scale.com/legal/security)
Pricing posture	Quote-only. Scale’s own documentation publishes no figures
Typical engagement / minimums	Not disclosed
Watch-out	The ownership conflict above, and a June 2025 Business Insider report that at least 85 Google Docs containing confidential client project material were publicly accessible by link; the files were locked down and an investigation opened after notification

Review attribute	Detail
G2	Seller page shows 5.0 from n=1. Report as no public reviews / fewer than 5 reviews on G2 as of 6 August 2026
Employee sentiment (Glassdoor)	3.7 (n≈424), accessed 6 August 2026. Employee sentiment, not customer sentiment
Evidence quality	The largest and most-discussed vendor in the category has essentially no customer-review footprint. That absence is itself a finding
What users like	Growth and exposure, on the employee side (Glassdoor)
What users flag	Work-life balance, on the employee side (Glassdoor)
Checked, nothing found	Capterra, TrustRadius, Gartner Peer Insights, Trustpilot. No usable figure as of 6 August 2026.

What it is. A hybrid platform and managed service running on an open gig crowd, historically the default enterprise choice for high-volume computer vision and, more recently, frontier-lab training and evaluation data.

Best for. Programmes at frontier scale with mature tooling requirements. Not for buyers who compete with Meta, or for anyone whose confidentiality review will ask who else has an economic interest in the vendor handling their data.

What customers say. Almost nothing, publicly. The largest vendor in the category has no usable customer-review footprint, and the only sizeable figure available is employee sentiment, which says nothing about delivery quality.

2. Appen

Attribute	Detail
Delivery model	Hybrid: ADAP platform plus managed services
Workforce model	Open vetted crowd, over 1 million contributors across 170+ countries and 235+ languages, plus secure in-house facilities for restricted work
Workforce geography	Disclosed in part. Named secure facilities in the UK, Cavite (Philippines) and Wuxi (China)
Ownership / independence	Public company, ASX: APX, no controlling shareholder. The checkable-ownership example in this roster
Best for	Multilingual and speech breadth at volume, from a vendor whose ownership you can verify on an exchange
Not for	Buyers who need a single dedicated employed cohort, or who cannot tolerate crowd-level variance on ambiguous tasks
Modalities	Text, audio and speech, image, video, document, multimodal
Specialist domains	Search relevance, speech and language, generative AI evaluation
Core features	ADAP platform, crowd management, secure facility delivery, LLM evaluation services
Quality method	Contributor scoring and multi-pass review; no published IAA threshold
Security & compliance	SOC 2 Type II, ISO 27001:2013 (TÜV Rheinland NA), US health-data privacy, GDPR, ISO 9001 (UK), Cyber Essentials
Pricing posture	Quote-only. appen.com/pricing returns 404
Typical engagement / minimums	Not disclosed
Watch-out	Client concentration. The Google contract was US$82.8M of FY23 revenue of US$273.0M, and its termination, announced 23 January 2024, removed all of it. FY2025 results (25 February 2026) showed US$230.8M of revenue, up 4.5% on an adjusted basis excluding the FY24 Google impact and down 1.5% statutory, underlying EBITDA before FX up from US$3.5M to US$12.2M, and FY2026 guidance of US$270 to $300M, so the concentration risk is documented and the recovery is too

Review attribute	Detail
G2	4.2/5 (n=35), accessed 6 August 2026. Customer-side
Capterra	4.1/5 (n=40), accessed 6 August 2026. Customer-side
Trustpilot	1.4/5 (n=366), re-read 6 August 2026. Contributor-side, not customer reviews. Earlier reads in this research pass recorded 1.8 to 2.2; the 1.4 is the figure from the live page and is the one printed
Employee sentiment (Glassdoor)	3.5 global (n≈2,200) and 2.9 US (n≈573), accessed 6 August 2026. Employee sentiment
Evidence quality	The clearest illustration in this article of why populations must be labelled: 4.1 to 4.2 from buyers (n=35 and n=40), 1.4 from workers (n=366), same brand
What users like	Scale of language coverage and project management, on G2 and Capterra
What users flag	Platform usability and turnaround consistency, on G2 and Capterra
Checked, nothing found	TrustRadius, Gartner Peer Insights. No usable figure as of 6 August 2026.

What it is. A 1996-founded Sydney company running a hybrid of the ADAP platform and managed services, on an open vetted crowd, with named secure facilities for restricted work.

Best for. Multilingual and speech programmes at volume where verifiable ownership matters. Not for teams needing one persistent employed cohort, or highly ambiguous taxonomies where crowd variance is expensive.

What customers say. The buyer-side figures are the strongest in the enterprise tier. Trustpilot’s much lower figure is contributor-side and measures the labour relationship, not the delivery. Everest Group named Appen a Leader in its 2024 Data Annotation and Labeling PEAK Matrix assessment.

3. iMerit

Attribute	Detail
Delivery model	Managed service plus the Ango Hub platform layer
Workforce model	Employed in-house. iMerit states its people work full-time for iMerit at its secure facilities and are not remote or outsourced workers
Workforce geography	The fullest geography disclosure in the enterprise tier. Delivery centres in Silicon Valley, New Orleans, Kolkata, Bengaluru and around eight further Indian cities, plus Thimphu, Bhutan (120+ people). Over 5,000 full-time staff, more than 50% women
Ownership / independence	Acquired by EXL (Nasdaq: EXLS) for up to $310M ($170M upfront plus up to $140M in earnouts), signed 24 June 2026 and completed 3 August 2026. Founder Radha Basu stays as EVP and Head of iMerit. Before the sale it was private and impact-investor backed (~$24.3M raised)
Best for	Regulated and high-consequence work where you need to know who the annotators are and that they are employed
Not for	Teams needing a self-serve platform, or surge capacity in dozens of languages overnight
Modalities	Image, video, 3D/LiDAR, text, document, medical imaging
Specialist domains	Medical and clinical, autonomous vehicles, geospatial, agriculture, commerce
Core features	Ango Hub, dedicated managed teams, domain-expert annotation pods
Quality method	Dedicated QA layer with expert review. Their “output accuracy above 98%” claim is vendor self-reported and is not printed here as a benchmark
Security & compliance	SOC 2 Type 2, ISO 27001, US health-data privacy, GDPR, ISO 9001:2015, TISAX
Pricing posture	Quote-only
Typical engagement / minimums	Not disclosed
Watch-out	Ownership changed in August 2026. iMerit is now a unit of EXL, a listed analytics and BPO company with its own client base, so ask the who-else-do-you-serve question again and check whether your contract survives a change of control

Review attribute	Detail
G2	4.7/5 (n=13) on the seller page, accessed 6 August 2026. Small sample; treat as directional
Employee sentiment (Glassdoor)	3.3 (n≈454), accessed 6 August 2026. Employee sentiment
Evidence quality	n=13 is below the level at which a rating discriminates between vendors. Use it as a signal of presence, not of superiority
What users like	Domain expertise and responsiveness of dedicated teams (G2)
What users flag	Onboarding time on complex taxonomies (G2)
Checked, nothing found	Capterra, TrustRadius, Gartner Peer Insights, Trustpilot. No usable figure as of 6 August 2026.

What it is. A 2012-founded managed service with the strongest workforce-employment guarantee in this roster, plus Ango Hub as a platform layer for teams that want visibility into the pipeline.

Best for. Regulated, high-consequence programmes where annotator identity and continuity are contractual concerns. Not for self-serve buyers or overnight multilingual surge.

What customers say. The G2 sample is too small to discriminate between vendors and should be read as a signal of presence. Everest Group listed iMerit as a Major Contributor in its 2024 assessment.

4. Sama

Attribute	Detail
Delivery model	Managed service through in-house impact and BPO delivery centres, plus the Sama Platform
Workforce model	Employed in-house, impact sourcing, B Corp certified
Workforce geography	Best geography disclosure in the enterprise tier. Offices published in San Francisco, Montréal, San José (Curridabat) Costa Rica, Nairobi, Gulu, Kampala and The Hague
Ownership / independence	Private. Last disclosed round $70M led by CDPQ, 4 November 2021. No round found after 2021 and no valuation disclosed
Best for	Buyers with an ESG or impact-sourcing mandate who also need published delivery addresses
Not for	Anyone whose procurement checklist requires a completed SOC 2 today
Modalities	Image, video, 3D/LiDAR, text, multimodal
Specialist domains	Autonomous vehicles, retail and e-commerce, agriculture, content moderation
Core features	Sama Platform, managed delivery pods, impact-sourcing programme with published impact reporting
Quality method	Dedicated QA layer with published impact and quality reporting cadence
Security & compliance	Achieved: ISO 27001, ISO 9001, TISAX, GDPR, CCPA. SOC 2 is listed on Sama’s own page as certification in progress, not achieved. ISO 42001 also in progress
Pricing posture	Quote-only
Typical engagement / minimums	Not disclosed
Watch-out	The SOC 2 status above. Sama is the only vendor in the enterprise tier without a completed SOC 2, and competing roundups routinely list it as certified

Review attribute	Detail
G2	4.6/5 (n=11), accessed 6 August 2026. Small sample
Employee sentiment (Glassdoor)	3.5 (n≈472), accessed 6 August 2026. Employee sentiment
Evidence quality	n=11 does not support a comparative claim. The published office addresses are stronger evidence than the rating
What users like	Delivery consistency and account management (G2)
What users flag	Pricing flexibility on smaller volumes (G2)
Checked, nothing found	Capterra, TrustRadius, Gartner Peer Insights, Trustpilot. No usable figure as of 6 August 2026.

What it is. A 2008-founded managed service delivering through owned impact and BPO centres, one of the first AI companies certified as a B Corp, with more than 13,000 people connected to digital work.

Best for. Programmes with an explicit ESG mandate that also need named delivery locations. Not for procurement processes that require a completed SOC 2 report at signature.

What customers say. The G2 sample is too small to rank on. Sama’s client disclosure is the strongest first-party list in the enterprise tier, naming Getty Images, Continental, Verizon, SWIFT, Siemens, Qualcomm, Sony, Microsoft, Walmart, eBay and NASA.

5. Innodata

Attribute	Detail
Delivery model	Managed service, plus its own Annotation Platform sold as SaaS or on-premise
Workforce model	Employed specialists, not a crowd. Innodata describes its annotators as specialists rather than crowd labour; 10,107 employees at 31 December 2025, 10,020 of them full-time (10-K)
Workforce geography	Disclosed by country in the 10-K: the Philippines, India, Sri Lanka, Canada, the UK, Israel, the US and Germany. The website claims 20+ delivery locations
Ownership / independence	Public company, Nasdaq: INOD. Founded 1988, Ridgefield Park, New Jersey. Rahul Singhal became President and CEO on 30 September 2026; founder Jack Abuhoff moved to Executive Chairman
Best for	Model builders that want employed, country-disclosed expert teams for fine-tuning, RLHF, red-teaming and evaluation data, from a vendor whose finances are filed with the SEC
Not for	Buyers who would sit behind the vendor’s two largest customers for capacity, or who want a self-serve price
Modalities	Image, video, LiDAR and sensor data, text, documents, code, speech and audio in 120+ languages
Specialist domains	Generative AI and frontier models, healthcare and life sciences, financial services, automotive, retail, media, US public sector (Innodata Federal)
Core features	Supervised fine-tuning data, RLHF and preference data, safety and red-teaming, benchmarks, RL environments, Annotation Platform, Innodata Federal
Quality method	Published process: written guidelines, multi-stage QA, sampling, reviewer feedback and arbitration; the platform supports single-pass, double-pass and blind double-pass work with inter-annotator agreement. No accuracy threshold is published
Security & compliance	The 10-K states ISO 27001:2022. The federal page also lists ISO 9001, SSAE 16 Type II (the older name for a SOC 1 financial-controls report, not SOC 2), NIST 800-171, a health-data privacy commitment and CMMC readiness (readiness, not certification). No SOC 2 claim found, and innodata.com/security returned 404 on 15 October 2026
Pricing posture	Quote-only. Platform tiers list seats but no prices
Typical engagement / minimums	Not disclosed for services. The platform’s Basic tier starts at 2 seats
Watch-out	Client concentration, disclosed because it has to be: one customer was 58% of FY2025 revenue, and in Q2 2026 the two largest were 37% and 34%, about 71% together (10-Q). Q2 2026 revenue was $92.1M, up 58%, on guidance of 40%+ growth for 2026. The securities class action that followed a February 2024 short report was dismissed with prejudice on 10 June 2026

Review attribute	Detail
G2	Not re-verified on 15 October 2026 (the page blocked automated reads). The August sweep recorded no public reviews (n=0)
Capterra	Not verified as of 15 October 2026
Trustpilot	2.9/5 (n=2), accessed 15 October 2026. Below any usable threshold
Employee sentiment (Glassdoor)	Not printed. A live read was blocked and cached figures are undated
Evidence quality	No usable customer-side rating. The SEC filings are the stronger evidence: revenue, headcount by country and client concentration, all stated under disclosure obligations
Checked, nothing found	TrustRadius, Gartner Peer Insights. No usable figure as of 6 August 2026.
What users like	No public review corpus to draw from.
What users flag	No public review corpus to draw from.

What it is. A 1988-founded New Jersey data-services company that has turned into one of the larger managed AI training-data vendors, selling fine-tuning, preference, red-teaming and evaluation data mostly to big model builders, with employed teams in eight disclosed countries.

Best for. Programmes that want a listed vendor with employed, country-disclosed specialists and a published QA process. Not for buyers who cannot accept that most of the vendor’s capacity goes to two customers.

What customers say. Nothing usable on the review platforms. As with TaskUs, the filings tell you more than any rating would: who the vendor depends on, how fast it is growing, and that a new CEO took over on 30 September 2026.

B. Expert-data marketplaces: RLHF, evaluations and domain experts

This is the tier most “2026” rosters do not contain at all, and it is where the money moved after June 2025. It is also the tier with the thinnest published certification evidence in the roster, which is an uncomfortable combination.

6. Surge AI

Attribute	Detail
Delivery model	Platform plus managed expert delivery
Workforce model	Credential-gated expert network with published recruiting bars
Workforce geography	Not disclosed. The network is quantified purely by credential (200,000+ PhDs, 500+ disciplines, 80+ languages, 30,000+ publications) and never by country
Ownership / independence	Bootstrapped with roughly $300k of personal savings and no venture capital from its 2021 launch until July 2025, when it began its first raise at a valuation Bloomberg reported at at least $25 billion. Founder Edwin Chen holds a majority. Around 110 full-time employees
Best for	Frontier RLHF, preference data and evaluation work where the credential of the annotator is the product
Not for	Regulated buyers who need a certification artefact, or anyone with a data-residency obligation
Modalities	Text, code, multimodal, preference and rubric data
Specialist domains	Law, medicine, finance, journalism, software engineering, consulting
Core features	Expert recruiting with published bars (legal roles require Supreme Court clerk, appellate advocate or senior counsel; journalism requires a Pulitzer, National Magazine Award or equivalent), managed expert delivery, RLHF and evaluation programmes
Quality method	Credential gating at recruitment. No published QA sampling or IAA methodology
Security & compliance	None published on surgehq.ai. The security URL returns 404. A third-party aggregator asserts SOC 2, GDPR, PCI, US health-data privacy, ISO 27001, FedRAMP and CSA STAR; none of it is first-party and it is not printed here as fact
Pricing posture	Quote-only. The only published rate is contractor pay, up to $200 to $400 per hour for a journalist role, which is supply-side, not customer price
Typical engagement / minimums	Not disclosed
Watch-out	Zero published certifications, zero disclosed geography, and no named clients on their own site. Press reports Anthropic, OpenAI, Google, Microsoft, Meta and the US Army as customers, which is press, not disclosure

Review attribute	Detail
Evidence quality	Zero public buyer evidence of any kind. For a company at this reported valuation, that is a striking finding rather than a neutral blank
Checked, nothing found	G2, Capterra, TrustRadius, Gartner Peer Insights, Trustpilot, Glassdoor. No usable figure as of 6 August 2026.
What users like	No public review corpus to draw from.
What users flag	No public review corpus to draw from.

What it is. A 2020-founded expert-data company with the most explicitly credential-gated recruiting model in the roster, selling RLHF, preference and evaluation data to frontier labs.

Best for. Work where the annotator’s professional credential is the deliverable. Not for regulated buyers, because there is no certification artefact to hand your security team.

What customers say. Nothing, anywhere. In a category where review volume mostly tracks customer-marketing motion, silence from a bootstrapped company selling to a handful of labs is expected, but your diligence has to come from references rather than from platforms.

7. Mercor

Attribute	Detail
Delivery model	Marketplace
Workforce model	Vetted expert network, 30,000+ experts, matched to lab projects on hourly contracts
Workforce geography	Not disclosed
Ownership / independence	$350M Series C at a $10B valuation, 27 October 2025, led by Felicis with Benchmark, General Catalyst and Robinhood Ventures. Five times the $2B Series B eight months earlier
Best for	Standing up domain-expert capacity quickly, on hourly terms, with an evidenced security posture
Not for	Buyers who need a fixed per-unit price, or a workforce whose location they can name
Modalities	Text, code, multimodal, preference and rubric data, expert evaluations
Specialist domains	Accounting, finance, computer science research, medicine, law
Core features	Live expert marketplace with posted roles and rates, hourly contracting, lab-project matching
Quality method	Expert vetting at intake. No published QA sampling or IAA methodology
Security & compliance	SOC 2 Type 2, and only that, but genuinely evidenced: a downloadable SOC 2 Type II 2026 report, a 2026 penetration test, a Mandiant report and a BC/DR plan on trust.mercor.com. Narrow claims with real artefacts, the direct foil to Cogito Tech
Pricing posture	Quote-only for buyers. The rates posted publicly (for example “Accounting Expert $80 to $120/hr”, “Computer Science PhD Researchers $70 to $90/hr”) are supply-side
Typical engagement / minimums	Hourly contracts; disburses roughly $1.5M per day to contractors at $85 to $95 per hour average
Watch-out	A security incident in late March 2026. Mercor was affected by a supply-chain attack involving LiteLLM, with unauthorised access between 24 and 30 March 2026; Mercor confirmed it publicly on 31 March. The Lapsus$ group claimed roughly 4TB of data; reporting describes Slack data, internal ticketing, source code, database records and candidate personal information among the material involved. Meta paused its contracts indefinitely, while OpenAI investigated its exposure and did not pause. A class action filed 1 April 2026 alleges more than 40,000 people were affected; Mercor says only “a very limited subset” of its nearly five million experts had sensitive information affected and has published no number. Five suits were on file within the first week and at least seven in federal courts in California and Texas by late April 2026. Mercor notified affected individuals on 25 and 26 June 2026 and offered identity protection; the class actions were undecided as of late September 2026. It held SOC 2 Type 2 throughout, which is the point: the certificate did not prevent this (Fortune, 2 April 2026; TechCrunch, 9 April 2026; Mercor statement, 25 June 2026)

Review attribute	Detail
G2	4.9/5 (n=6), accessed 6 August 2026. Below a usable sample size, and the page is CAPTCHA-walled, with the count recorded as 6 in some views and 7 in others. Treat as unverified
Trustpilot	4.4/5 (n=640) on mercor.io, accessed 6 August 2026. Contributor-side, not customer reviews. The profile is claimed and on a paid subscription, reviews are actively solicited, and 637 of the 640 arrived inside the last twelve months. A secondary domain profile shows 2.5/5 (n=5), unsolicited
Employee sentiment (Glassdoor)	4.2 (n≈162), accessed 6 August 2026. Employee sentiment
Evidence quality	Solicitation asymmetry applies here. Mercor’s 4.4 sits beside Turing’s unsolicited 2.8 (n=194) and LXT’s unsolicited 2.6 (n=6). A material part of that gap is a collection-method difference, not a quality difference
What users like	Timely payment and clear task flow, from contributors on Trustpilot
What users flag	Inconsistent project availability, from contributors on Trustpilot
Checked, nothing found	Capterra, TrustRadius, Gartner Peer Insights. No usable figure as of 6 August 2026.

What it is. A 2023-founded San Francisco marketplace matching credentialed domain experts to AI-lab projects on hourly contracts, and the fastest-repricing asset in the category.

Best for. Buyers who need expert hours quickly and value a downloadable audit report over a long certification list. Not for fixed per-unit budgeting, or for anyone whose obligation requires naming the workforce’s jurisdiction.

What customers say. Nothing usable on the buyer side. The visible 4.4 comes from 640 contributors on a paid, actively solicited profile, which is a different population and a different collection method from the unsolicited profiles it gets compared against.

8. Turing

Attribute	Detail
Delivery model	Managed plus platform
Workforce model	Vetted expert network across coding, reasoning and STEM
Workforce geography	Not disclosed
Ownership / independence	Private, VC-backed. Roughly $247M raised at a ~$2.2B valuation; $111M Series E in March 2025 led by Khazanah Nasional. ARR reported around $300M, up from $167M
Best for	Coding, reasoning and agentic training data for model developers
Not for	Buyers with a security questionnaire, because there is nothing published to answer it with
Modalities	Text, code, reasoning traces, RL environments
Specialist domains	Software engineering, STEM, enterprise knowledge work
Core features	Domain Software Engineering, Enterprise Knowledge Work, Frontier STEM Capability, off-the-shelf datasets, RL environments
Quality method	Expert vetting at intake. No published QA methodology
Security & compliance	None found. turing.com/security and /trust both return 404 and trust.turing.com does not resolve. No SOC 2, ISO 27001, US health-data privacy, GDPR or PCI DSS claim anywhere on turing.com
Pricing posture	Quote-only. Published rates are expert-side
Typical engagement / minimums	Not disclosed
Watch-out	Beyond the missing security pages, turing.com/about, /about-us, /company/about and /company/news all returned 404 on 6 August 2026, so the About link in their own navigation is broken

Review attribute	Detail
G2	4.2/5 (n=18), accessed 6 August 2026. The page is CAPTCHA-walled to automated reads. The figure held consistent across G2’s own product and comparison views on re-check, but it has not been confirmed on a rendered live page
TrustRadius / Gartner Peer Insights	TrustRadius 3.95/5 (n=4), accessed 6 August 2026. Below the n=5 threshold, so formally no public reviews. Gartner not systematically checked
Trustpilot	2.8/5 (n=194), accessed 6 August 2026. Contributor-side. Bimodal: roughly 70% five-star and 22% one-star, with the company replying to about 17% of negatives. Unsolicited profile
Employee sentiment (Glassdoor)	3.5 (n≈764), accessed 6 August 2026. Employee sentiment
Evidence quality	Solicitation asymmetry applies. Turing’s 2.8 is unsolicited and Mercor’s 4.4 is paid and solicited, so the two are not directly comparable
What users like	Pay rates and project interest, from contributors on Trustpilot
What users flag	Screening process length and payment disputes, from contributors on Trustpilot
Checked, nothing found	Capterra. No usable figure as of 6 August 2026.

What it is. A 2018-founded company that pivoted from remote-developer staffing into frontier-AI human data, now selling coding, reasoning, STEM and agentic training data to labs.

Best for. Model developers who need software and reasoning data at pace. Not for any buyer whose security review requires a published framework, because Turing publishes none.

What customers say. Effectively nothing verifiable. All three visible figures are blocked, below threshold, or contributor-side, so treat them as unusable for vendor comparison.

9. Handshake AI

Attribute	Detail
Delivery model	Fellowship-run expert network
Workforce model	Credential-gated fellows recruited from an 18M+ student and alumni graph; 100k+ fellows, $100M+ paid out
Workforce geography	The sharpest constraint in the roster, and it is published: United States only. “You must be based in the U.S. and have a valid work authorization … STEM OPT is not supported”
Ownership / independence	Private, VC-backed. Last reported valuation $3.3B (2022); acqui-hired the nine-person Cleanlab team in January 2026, terms undisclosed. Private-markets research firm Sacra estimates roughly $1.1B annualised gross revenue by April 2026
Best for	US-data-residency and US-persons requirements, and academic-credentialed expert data
Not for	Any programme needing non-US coverage or non-English language work
Modalities	Text, code, multimodal, expert evaluations
Specialist domains	Academic disciplines across the graph, plus medicine, law and finance at the top rate band
Core features	Handshake AI Fellowship, verified academic identity graph built on the parent network’s 18M+ students and alumni, credential gating by degree level
Quality method	Credential verification through the academic graph. No published QA sampling methodology for the data business
Security & compliance	Trust portal lists SOC 2 Type II, PCI DSS v4.0.1, GDPR, CCPA, TX-RAMP, UK Cyber Essentials and EU-U.S. DPF. The scope covers the recruiting products, not the AI data unit. A real certificate that does not cover what you would be buying
Pricing posture	Quote-only for buyers. Published rates are fellow-side: Handshake’s own fellowship postings run roughly $40 to $125 per hour, with higher bands for medicine, law and finance that Handshake does not publish as a consolidated rate card
Typical engagement / minimums	Not disclosed
Watch-out	The certification scope gap above. Also, the “trusted by” logos on the trust portal are employers on the career network, not AI-data customers, and should not be read as client references

Review attribute	Detail
G2	No review-platform coverage of Handshake AI’s data business as of 6 August 2026. Any “Handshake” G2 figure belongs to the college-recruiting product and is not printed here
Capterra	Same. No AI-data profile exists; the Capterra listings under this name are the recruiting product and an unrelated wholesale-commerce vendor
TrustRadius / Gartner Peer Insights	Same. Gartner’s vendor page states Handshake has exactly one product in one market, and it is not this one
Trustpilot	A single handshake.ai-specific review (n=1), which is below any threshold. Formally, no public reviews
Employee sentiment (Glassdoor)	The 3.2 (n≈294, accessed 6 August 2026) figure belongs to the parent career-network company, not the AI division, and is not a rating of this business
Evidence quality	The clearest name collision in the category. Every publicly circulating “Handshake” rating measures a different product
What users like	Not applicable for the AI data business
What users flag	Not applicable for the AI data business

What it is. The AI data division of Handshake, the Gen-Z career network, launched in 2025 and run as a fellowship that converts a twelve-year academic identity graph into a credentialed expert workforce.

Best for. Buyers with US-persons or US-residency requirements who need academic credentials verified rather than asserted. Not for multilingual or non-US work, where the published work-authorisation constraint is disqualifying by design.

What customers say. There is no review-platform coverage of this business anywhere. Note also that handshake.ai is a separate, unrelated company doing audit and assurance for AI agents, which is a third entity in the same name space.

10. Micro1

Attribute	Detail
Delivery model	Platform plus expert network, positioned as a “data lab”
Workforce model	Top-percentile vetted experts, with anti-cheat integrity scoring
Workforce geography	Not disclosed
Ownership / independence	$35M Series A at a $500M valuation, 12 September 2025, led by 01 Advisors. Roughly $41.6M total across four rounds. Forbes reported in September 2026 that Micro1 raised $100M+ at a $4B valuation; Micro1 declined to comment and has not announced it, so treat it as reported
Best for	RL environments, frontier evaluations and expert-demonstrated robotics data
Not for	Government or regulated buyers who need a certification artefact
Modalities	Text, code, multimodal, robotics demonstration data
Specialist domains	Robotics, agent evaluation, academic and clinical expertise
Core features	Realm (RL environments and frontier evaluations), Cortex (contextual evaluation for production agents), Robotics (expert-demonstrated real-world data)
Quality method	Anti-cheat integrity scoring on the expert side. No published QA sampling or IAA methodology
Security & compliance	None. The Legal Center contains only website privacy, cookie, candidate and referral terms, and trust.micro1.ai does not resolve. Notable given they market a Government partnerships track
Pricing posture	Quote-only
Typical engagement / minimums	Not disclosed
Watch-out	No published certifications alongside a Government track, and no named clients; a third-party analyst estimate names Microsoft, which is unverified

Review attribute	Detail
Trustpilot	Approximately 4/5 (n≈9,333), accessed 6 August 2026. Candidate-side reviews of the AI interviewer product, not customer reviews of a labeling service. Pagination was inconsistent; treat n as approximate
Evidence quality	The large visible sample is the single most misleading number in this category, because it measures job candidates rather than buyers. On the buyer side, there is nothing
What users like	Interview experience, from candidates on Trustpilot
What users flag	Interview outcome opacity, from candidates on Trustpilot
Checked, nothing found	G2, Capterra, TrustRadius, Gartner Peer Insights, Glassdoor. No usable figure as of 6 August 2026.

What it is. An April 2022-founded expert network and platform selling RL environments, agent evaluation and robotics demonstration data, recruiting professors, MDs and PhDs with integrity scoring on top.

Best for. Frontier evaluation and robotics work where expert demonstration is the deliverable. Not for regulated or public-sector buyers, given the absence of any published certification.

What customers say. Nothing on the buyer side. The large visible Trustpilot volume is candidate sentiment about an AI interviewer, and reading it as customer satisfaction would be a category error.

C. High-volume BPO and crowd-scale workforce

Where multilingual breadth and surge capacity actually live, and where the ownership corrections in this article cluster.

11. TELUS Digital (AI Data Solutions)

Attribute	Detail
Delivery model	Hybrid at the largest scale in the roster: BPO delivery centres, a proprietary training platform, and an open crowd
Workforce model	BPO delivery centres plus an open “AI Community” crowd of 1M+ annotators
Workforce geography	Broadly disclosed: 82,000+ team members across 35+ countries including the Philippines, Canada, US, Guatemala, El Salvador, Ireland, Bulgaria, Romania, China, Morocco and South Africa. 500+ annotation languages and dialects; 2B+ labels annually
Ownership / independence	Wholly owned subsidiary of TELUS Corporation (TSX: T / NYSE: TU). The privatization closed on 31 October 2025 at US$4.50 per share, roughly US$539M for the shares TELUS did not own; NYSE removal took effect 11 November 2025 and TIXT no longer trades. The legal entity is still TELUS International (Cda) Inc. and the brand is still TELUS Digital, and there was no rebrand at privatization. Any guide describing TELUS Digital as NYSE-listed is stale. In its Q2 2026 results (31 July 2026) TELUS booked a C$2.1B goodwill and intangible impairment on the TELUS Digital unit, citing faster declines in legacy services as customers automate with AI, alongside a 55% dividend cut and a moratorium on acquisitions
Best for	Programmes needing genuine 500-language coverage and large surge capacity
Not for	Buyers needing a certification scope that covers the whole entity rather than the labeling facilities
Modalities	Text, audio and speech, image, video, document, multimodal
Specialist domains	Search relevance, content moderation, localisation, speech
Core features	AI Data Solutions built by acquisition (Lionbridge AI, completed 31 December 2020 at roughly C$1.2B, and Playment in 2021), proprietary annotation platform, AI Community crowd
Quality method	Platform-managed multi-pass review. No published IAA threshold
Security & compliance	Note the hedging, verbatim: “We are SOC 2 compliant and TISAX certified, and our labeling facilities are ISO 27001 certified.” No Type I or II designation is given, and ISO 27001 is scoped to labeling facilities rather than the entity
Pricing posture	Quote-only
Typical engagement / minimums	Not disclosed
Watch-out	The delisting correction above, the C$2.1B impairment the parent took on this unit in July 2026 (ask how the AI data business is being funded), the unspecified SOC 2 Type, and the facility-scoped ISO 27001. No named clients on their own site; they lead with Everest Group placement instead

Review attribute	Detail
Employee sentiment (Glassdoor)	Not printed; no consolidated figure isolated for the AI Data Solutions unit
Evidence quality	One of the largest annotation operations on earth carries no customer-review footprint. Absence is the finding
Checked, nothing found	G2, Capterra, TrustRadius, Gartner Peer Insights, Trustpilot. No usable figure as of 6 August 2026.
What users like	No public review corpus to draw from.
What users flag	No public review corpus to draw from.

What it is. Founded in 2005 as TELUS International and renamed TELUS Digital in 2024, with an AI Data Solutions arm assembled by acquiring Lionbridge AI and Playment, running the largest published label volume in this roster.

Best for. Language breadth and volume that few operations can match. Not for security reviews that will not accept an ISO 27001 scoped only to the labeling facilities.

What customers say. Nothing publicly. Everest Group named TELUS International a Leader in its 2024 Data Annotation and Labeling PEAK Matrix assessment, which is analyst placement rather than customer evidence, and the two are not substitutes.

12. TaskUs

Attribute	Detail
Delivery model	Managed BPO plus a Work@Home model
Workforce model	BPO delivery centres, 64,000+ employees, 30+ languages
Workforce geography	31 named sites across Colombia, Croatia, Egypt, Greece, India, Ireland, Japan, Malaysia, Mexico, the Philippines (11 sites), Serbia, Taiwan and the US
Ownership / independence	Publicly listed, Nasdaq: TASK. The $16.50 per share, $1.62B Blackstone and co-founder take-private announced 9 May 2025 was voted down by stockholders on 8 October 2025, with no termination fee. TaskUs continues to trade; Q2 2026 revenue was $308.9M, up 5%, with about 63,200 staff (5 August 2026)
Best for	Certification-heavy, audit-heavy programmes needing site-level accountability
Not for	Small-volume or short-duration engagements
Modalities	Text, image, video, audio, document, multimodal
Specialist domains	Content moderation, trust and safety, customer experience, fintech and crypto operations
Core features	Managed delivery pods, Work@Home model, 31 named sites, audit-ready reporting
Quality method	Dedicated QA layer with site-level reporting. No published IAA threshold
Security & compliance	The fullest stack in the roster: SOC 2 Type II, ISO 27001, ISO 27701, US health-data privacy, GDPR, PCI DSS, HITRUST. Corroborated in the 10-K for SOC 2 Type 2, ISO 27001, HITRUST and PCI DSS
Pricing posture	Quote-only
Typical engagement / minimums	Enterprise contracts; not published
Watch-out	Named in litigation over the 2024-25 Coinbase data breach, where an amended complaint alleges agents in India took bribes to exfiltrate customer data. Pending litigation, allegations rather than findings. Separately, client concentration is disclosed by SEC obligation: Meta Platforms was the largest client at 19% of FY2023 revenue (22% in FY2022), with top-10 clients at 55% and top-20 at 68% of FY2023 revenue

Review attribute	Detail
G2	4.0/5 (n=11) on the seller page, accessed 6 August 2026. Small sample
Employee sentiment (Glassdoor)	4.0 (n≈18,094), accessed 6 August 2026. Employee sentiment, and the largest workforce sample in this article
Evidence quality	The two 4.0 figures are coincidental and measure different things. n=11 from buyers, n≈18,094 from employees
What users like	Delivery discipline and reporting (G2); culture and facilities (Glassdoor)
What users flag	Cost at lower volumes (G2); shift patterns (Glassdoor)
Checked, nothing found	Capterra, TrustRadius, Gartner Peer Insights, Trustpilot. No usable figure as of 6 August 2026.

What it is. A 2008-founded BPO headquartered in New Braunfels, Texas, with delivery originating in Bacoor, Cavite, now running annotation, moderation and trust-and-safety work at 31 named sites.

Best for. Programmes where the certification stack and site-level auditability are procurement requirements. Not for small pilots, where a 64,000-person operation is the wrong shape.

What customers say. The G2 sample is too small to rank on. The far more useful evidence is the SEC-disclosed client concentration, a fact a private vendor never has to give you.

13. LXT (including clickworker)

Attribute	Detail
Delivery model	Managed service plus open crowd hybrid
Workforce model	Managed annotation and collection delivery, plus clickworker’s crowd of over 10 million contributors
Workforce geography	The fullest geography disclosure in the roster. Offices in Canada, the US, UK, Egypt, India, Romania, Turkey, Germany and Australia; 150+ countries and 1,000+ language locales; named secure facilities: two in Toronto, one in Mississauga, one in Montreal, one in Cairo. On-site deployment into the client’s own facility is offered
Ownership / independence	Private, headquartered in Toronto. clickworker was acquired by LXT: definitive agreement announced 17 December 2024, technology integration completed 31 July 2025, so roundups listing them separately are listing one company twice
Best for	Confidentiality-sensitive multilingual work needing named facilities or on-site delivery
Not for	Buyers requiring SOC 2, which LXT does not claim
Modalities	Audio and speech, text, image, video, document, multimodal
Specialist domains	Speech and language data collection, localisation, search relevance
Core features	Managed delivery, supervised annotation inside secure facilities, on-site deployment, clickworker crowd access
Quality method	Supervised delivery inside named facilities. No published IAA threshold
Security & compliance	LXT: ISO 27001 (technology and physical facilities), GDPR, US health-data privacy, OWASP Top Ten, PCI DSS compliant facilities. No SOC 2, no ISO 27701. clickworker: ISO 27001, GDPR, and a Crowdsourcing Code of Conduct certification, the only labour-fairness credential in this roster
Pricing posture	Quote-only for LXT. clickworker’s pricing URL resolves to a page with no rates on it, which is a fair illustration of the category norm
Typical engagement / minimums	Not disclosed
Watch-out	LXT publishes no named clients; all 20 case studies are anonymised. clickworker’s crowd figure has grown from roughly 500k in 2013 to 10 million now, so always date-stamp it. The Trustpilot location field reads “Anguilla”, which is a platform auto-placeholder and not LXT’s location

Review attribute	Detail
G2	No profile located as of 6 August 2026. Searches for “LXT” return LTX (Lightricks), a different company; do not use that profile
Capterra	No profile located as of 6 August 2026
TrustRadius / Gartner Peer Insights	No profile located; Gartner not systematically checked
Trustpilot	LXT 2.6/5 (n=6), accessed 6 August 2026, unclaimed profile, all six reviewers are contributors rather than buyers. clickworker 2.6/5, re-read 6 August 2026, worker-side; the review count did not resolve consistently and is not printed
Employee sentiment (Glassdoor)	3.4 (n≈134), accessed 6 August 2026. Employee sentiment
Evidence quality	No customer-side evidence exists for either brand. LXT’s 2.6 is unsolicited and n=6, which is below threshold; the solicitation-asymmetry caveat applies when it is placed beside Mercor’s solicited 4.4
What users like	Timely payment on some clickworker task types, from workers
What users flag	Unpaid or rejected tasks, account suspensions handled by automated support only, and withdrawal fees, from workers on clickworker’s profile

What it is. A 2010-founded Toronto company selling managed annotation and data collection, now holding clickworker’s ten-million-contributor crowd, with a confidentiality posture that includes supervised annotation inside named facilities and deployment into the client’s own site.

Best for. Multilingual work with a hard confidentiality constraint, where named facilities beat an anonymous crowd. Not for procurement checklists that require SOC 2.

What customers say. Nothing. Neither brand carries a customer-review profile. Both worker-side profiles sit at 2.6, and clickworker’s complaint pattern about unpaid tasks and automated-only support is consistent with that rating rather than in tension with it.

14. Toloka

Attribute	Detail
Delivery model	Platform plus marketplace
Workforce model	Open crowd evolving into a vetted expert network; contributor-facing brand is Mindrift.ai
Workforce geography	200,000+ experts across 100+ countries, 40+ languages, 50+ verified knowledge domains
Ownership / independence	Founded 2014 inside Yandex, now Amsterdam-based (Toloka AI BV) and spun into Nebius Group NV (Nasdaq: NBIS). A May 2025 round led by Bezos Expeditions saw Nebius relinquish controlling voting power while remaining a shareholder; Shopify CTO Mikhail Parakhin is Board Chairman. Amount not stated by Toloka; press reports roughly $72M
Best for	Crowd breadth with a real certification stack, moving toward expert and safety evaluation work
Not for	Healthcare or card-data workloads, since neither a US health-data privacy statement nor PCI DSS is claimed
Modalities	Text, image, audio, video, multimodal, preference data
Specialist domains	Red-teaming, safety evaluation, search relevance, domain-expert evaluation
Core features	Crowd and expert marketplace, Mindrift.ai contributor brand, red-teaming and evaluation programmes
Quality method	Platform consensus and overlap controls. No published IAA threshold
Security & compliance	SOC 2 Type II, ISO 27001, ISO 27701, GDPR. One of only two vendors in the roster claiming ISO 27701 (the other is TaskUs). No US health-data privacy statement, no PCI DSS
Pricing posture	Structurally transparent, numerically opaque. No hidden fees, no minimums and no contracts; the buyer sets the per-task price and the platform adds a platform fee and an LLM-QA cost as an unpublished percentage of it. The price-structure document returned 403 on 6 August 2026
Typical engagement / minimums	No minimums, per their own statement
Watch-out	The unpublished percentage above means you cannot model total cost from the published structure alone. Ask for the platform fee in writing before the pilot

Review attribute	Detail
G2	A profile exists but carries no rating and no sample size as of 6 August 2026; G2’s own “not enough reviews to provide buying insight” notice sits on the page
Trustpilot	Present but sample size not visible, accessed 6 August 2026. Worker-side. Complaints centre on low effective hourly pay, unpaid training and tests, and rejected submissions
Employee sentiment (Glassdoor)	n≈82, rating not visible as of 6 August 2026. Employee sentiment
Evidence quality	No usable customer-side rating with a visible sample size. The strongest evidence Toloka offers is its named client list, not a score
What users like	Task availability, from workers
What users flag	Effective hourly pay and rejected submissions, from workers
Checked, nothing found	Capterra, TrustRadius, Gartner Peer Insights. No usable figure as of 6 August 2026.

What it is. A crowd platform with a decade of history that has repositioned around vetted domain experts, red-teaming and safety evaluation, now operating outside Yandex control under Nebius Group.

Best for. Buyers who want crowd breadth plus a certification stack that includes ISO 27701. Not for regulated healthcare or payment-data work.

What customers say. No usable public rating exists. Toloka’s client disclosure is vendor-published rather than a logo wall: Anthropic, Amazon, Microsoft, poolside, Recraft and Shopify.

D. Managed mid-market and domain specialists

Sub-enterprise budgets that still need a dedicated team, and the tier where published rate cards start to appear.

15. CloudFactory

Attribute	Detail
Delivery model	Managed service with a platform layer
Workforce model	Employed in-house delivery teams, not an open crowd
Workforce geography	Delivery in Nepal, Kenya, the Philippines and Colombia; core team across the UK, US, Nepal and Kenya; 100% remote operation. No HQ address published on their own site; /locations returns 404
Ownership / independence	Private, VC and PE backed. Latest announced round $65M on 20 November 2019, including FTV Capital. Trackers disagree on total raised ($78.1M against up to $153M) and CloudFactory publishes no funding summary
Best for	A dedicated employed team on a mid-market budget, with a QA mechanism you can inspect
Not for	Buyers who need the vendor’s legal HQ and full funding history documented first-party
Modalities	Image, video, text, document, audio
Specialist domains	Insurtech, retail, agriculture, transportation, medical imaging support
Core features	Five-person “toli” teams with weekly face-to-face accuracy reviews, managed delivery pods, platform layer
Quality method	The most citable QA mechanism in the roster: workers organised into teams of five with weekly face-to-face accuracy reviews. A mechanism rather than an adjective
Security & compliance	The strongest disclosure page in the managed tier: ISO 27001:2022, SOC 2 validated against Security, Availability and Confidentiality (Type not stated), US health-data privacy as a compliant business associate, GDPR “for over seven years”, ISO 9001:2015. Not claimed: ISO 27701, CCPA, PCI DSS
Pricing posture	Quote-only, but unusually they publish a page saying so: a consumption-based model based on annual spend commitments and customised rates, piece-rate and consumption-based, no free trial but a free analysis
Typical engagement / minimums	Annual commitments of spend; specific minimums not published
Watch-out	The SOC 2 Type is not stated, and the missing HQ address and funding disclosure make financial diligence a third-party exercise

Review attribute	Detail
G2	4.5/5 (n=11) on the seller page, accessed 6 August 2026. Small sample
Employee sentiment (Glassdoor)	n≈547, rating not visible as of 6 August 2026. Employee sentiment
Evidence quality	n=11 is directional only. The published QA mechanism is stronger evidence than the score
What users like	Team consistency and communication (G2)
What users flag	Ramp time on new task types (G2)
Checked, nothing found	Capterra, TrustRadius, Gartner Peer Insights, Trustpilot. No usable figure as of 6 August 2026.

What it is. A 2010-founded managed service started in Nepal on the premise that talent is distributed more evenly than opportunity, running employed delivery teams rather than a crowd.

Best for. Mid-market buyers who want a persistent employed cohort and a QA mechanism they can audit. Not for anyone whose diligence requires the vendor’s own site to document HQ and funding.

What customers say. The G2 sample is directional only. Named clients include Compound Eye, Driver Technologies, Deepcell, Matterport, Expensify, Ibotta, Sartorius and Activ Surgical, against a claimed 700+ client base.

16. Cogito Tech

Attribute	Detail
Delivery model	Managed service
Workforce model	BPO delivery centres with managed teams of trained specialists by industry, platform-agnostic
Workforce geography	Best delivery-centre disclosure in the roster: three named Noida addresses (A-83 Sector-2; C-01 Sector-59; C-40 Sector-59). Headcount not disclosed
Ownership / independence	Privately held and bootstrapped. No disclosed funding, no institutional investors, no parent, not public. Founded 2011 by Rohan Agrawal, Levittown, New York
Best for	Buyers who value named delivery addresses and are willing to verify certifications themselves
Not for	Regulated buyers who need certification artefacts at signature
Modalities	Image, video, text, document, audio, multimodal
Specialist domains	Healthcare, autonomous vehicles, retail, content moderation, RLHF and red teaming
Core features	Managed expert teams, HITL workflows, content moderation, RLHF, fine-tuning, prompt engineering, red teaming
Quality method	Managed teams with industry-trained specialists. No published QA sampling or IAA methodology
Security & compliance	Claims six: “EU GDPR Certified”, “SOC 2 TYPE II Certified”, a “Certified” badge for US health-data privacy, “ISO 27001:2013 accredited secure facilities worldwide”, “ISO 9001 certified” and CCPA compliance. No verifying artefact is published for any of them. Two factual notes: the US health-data privacy rule has no certifying body, so a “certified” badge for it is not a real designation; and the ISO claim cites the superseded 2013 revision where peers cite 27001:2022
Pricing posture	Quote-only. There is no pricing page on the site at all
Typical engagement / minimums	Not disclosed
Watch-out	The certification position above is the “most claims, least evidence” end of the spectrum in this roster. Separately, the logo wall (OpenAI, Medtronic, AWS, NBC Universal, Siemens, Samsara, Stryker) carries no backing case studies and should be treated as claimed rather than verified. Name collision: Cogito Tech is not Cogito Corporation, the contact-centre AI company acquired by Verint in 2024, and not Epic Cogito

Review attribute	Detail
G2	A “COGITO Tech Data labeling” profile exists showing about 4.7, but the sample size resolved as 12 in one view and 14 in another as of 6 August 2026, so no figure is printed here. Check you are on that profile and not one of the several unrelated “Cogito” products
Employee sentiment (Glassdoor)	n≈45, rating not visible as of 6 August 2026. Employee sentiment
Evidence quality	No usable public rating. Better evidence sits in the named testimonial sources than in the logo wall
Checked, nothing found	Capterra, TrustRadius, Gartner Peer Insights, Trustpilot. No usable figure as of 6 August 2026.
What users like	No public review corpus to draw from.
What users flag	No public review corpus to draw from.

What it is. A 2011-founded managed annotation provider running BPO delivery from three named Noida centres, covering annotation, moderation, RLHF and red teaming for enterprise buyers.

Best for. Buyers who weight physical delivery-address disclosure heavily and will do their own certification verification. Not for regulated programmes that need the report in hand.

What customers say. No usable public rating exists. The named testimonial sources (Machine AI, VISUA Labs, Tilesview.ai, Axiom Bio, Smart Eye, Verdure Imaging) are better evidence than the tier-1 logo wall, because they are attributable.

17. Label Your Data

Attribute	Detail
Delivery model	Hybrid service plus platform
Workforce model	Managed employed annotator teams, not an open crowd. The site runs a named annotator roster and the FAQ calls annotators “our employees”
Workforce geography	1,000+ teammates across 22 countries and 3 offices; named annotator countries include Poland, Ukraine, Bulgaria, Slovenia, Lithuania, Moldova, Morocco and Kenya. Their own pages disagree: 22 against 25 countries, and roughly 200 to 500 against 1,000+ annotators
Ownership / independence	Founded 2020. Stated subsidiary of SupportYourApp, sourced only to the vendor’s own article; no VC round found. US entity in Wilmington DE, EU entity in Nicosia, Cyprus; no single stated HQ
Best for	Mid-market teams that want published rates and no annual commitment
Not for	Buyers who require a SOC 2 report or a dedicated security page
Modalities	Image, video, text, audio, document
Specialist domains	Retail, agriculture, security, medical imaging support
Core features	Managed employed annotator teams, free pilot, self-serve entry path, published rate card
Quality method	Managed team review. No published IAA threshold
Security & compliance	Claims PCI DSS Level 1, ISO 27001, GDPR and US health-data privacy, plus CCPA adherence. No SOC 2 claim, and no security page: both /security and /data-security return 404
Pricing posture	Published, one of only six in the roster: bounding box from $0.02 per object; keypoint $0.015 per object; NLP $0.02 per entity; $6 per annotator hour
Typical engagement / minimums	No minimum monthly volumes or annual contracts. Free pilot; self-serve requires a $100 balance top-up after the trial
Watch-out	The internal inconsistencies in their own headcount and country figures, and the parent-company relationship sourced only to their own content

Review attribute	Detail
G2	4.9/5 (n=15), accessed 6 August 2026. A high score on a small sample
Evidence quality	This is exactly the comparison this article warns about. A 4.9 from 15 and a 4.9 from 284 are not the same claim, and only one of them discriminates between vendors
What users like	Turnaround speed and price transparency (G2)
What users flag	Capacity limits on large surges (G2)
Checked, nothing found	Capterra, TrustRadius, Gartner Peer Insights, Trustpilot, Glassdoor. No usable figure as of 6 August 2026.

What it is. A 2020-founded hybrid service running employed annotator teams across 22 countries, and one of the few vendors in this market that publishes an actual rate card.

Best for. Mid-market programmes that need real numbers before a procurement cycle starts and cannot commit annually. Not for buyers whose security review requires SOC 2 or a security page that resolves.

What customers say. A 4.9 from 15 buyers is not the same claim as a 4.9 from 284, and only one of them discriminates. Named clients include ABB, UiPath, Yale, Bizerba, George Washington University, Toptal, Hypatos, Ouster, Zendar, Respeecher, Geberit and NODAR.

E. Annotation platforms and tooling-first hybrids

You supply the workforce; they supply the software. Several of these attach an expert network, which moves them along the spectrum without changing which risk you retain by default.

18. Labelbox

Attribute	Detail
Delivery model	Hybrid: SaaS platform plus a vetted expert crowd branded Alignerr, plus an add-on team of professional labelers
Workforce model	Vetted expert crowd, 2.6M+ contributors claimed
Workforce geography	Breadth disclosed, delivery footprint not. 40+ countries and 75+ languages, named markets including the US, India, UK, Germany, Canada, France, Japan, Brazil, South Korea and Australia; credential mix claimed at 50k+ PhDs, 200k+ master’s and 85k+ licensed professionals
Ownership / independence	Private, VC-backed. Series E of more than $100M (equity plus a credit facility) led by Infinitum Partners, 14 October 2026, valuation not disclosed; before that, $189M across rounds to a $110M Series D on 6 January 2022 led by SoftBank Vision Fund 2. Acquired Upcraft in February 2026 to grow Alignerr
Best for	Platform-first teams that want an expert network attached without changing tools
Not for	Anyone budgeting from published rates, because there are none any more
Modalities	Image, video, text, audio, document, multimodal, LLM and RL data
Specialist domains	Frontier model training and evaluation, computer vision, content understanding
Core features	Catalog, Annotate, Model, plus Horizon, Terra and Recursion; Alignerr expert network
Quality method	Consensus and review workflows inside the platform. No published IAA threshold
Security & compliance	Security page states SOC 2 Type II and a US health-data privacy programme. ISO 27001 appears there only as a standard that informs the programme, not as a held certification, while a company blog post claims ISO 27001:2022; the two have not been reconciled and the trust portal is gated
Pricing posture	Quote-only, following a correction. Labelbox removed its public pricing page: both the pricing URL and the calculator return 404, verified 6 August 2026, and “Pricing” no longer appears in navigation or footer. The per-LBU and per-hour figures circulating in other roundups cite a dead page and are not reproduced here
Typical engagement / minimums	Unit definitions survive without dollar values: 500 free LBU credits per month; 1 LBU per 60 data rows for Catalog storage; 1 LBU per data row for image, text and audio annotation; 20 LBU per data row for live multimodal and LLM work; 1 LBU per 5 data rows for Model
Watch-out	The removed pricing page, and the unreconciled ISO 27001 position between the security page and the blog

Review attribute	Detail
G2	4.5/5 (n=48), accessed 6 August 2026. The n=47 that circulates is an older crawl of the same page
Evidence quality	n=48 is one of the six usable customer samples in this roster, alongside SuperAnnotate, Encord, V7, Kili and Roboflow
What users like	Catalog search and model-assisted labeling workflows (G2)
What users flag	Cost predictability and the credit model (G2)
Checked, nothing found	Capterra, TrustRadius, Gartner Peer Insights, Trustpilot, Glassdoor. No usable figure as of 6 August 2026.

What it is. A 2018-founded San Francisco platform company now branding itself around reinforcement-learning data, with an expert crowd (Alignerr) attached to the tooling.

Best for. Teams that want one platform plus optional labour without switching vendors. Not for buyers who need to model cost before a sales conversation.

What customers say. A genuinely usable sample by this category’s standards. Named clients include Meta, through a first-party case study covering 820 expert-authored problems for a Meta benchmark, plus Google, Ideogram, Speak, ElevenLabs and Suno.

19. SuperAnnotate

Attribute	Detail
Delivery model	Hybrid: platform plus a global network of vetted experts
Workforce model	Vetted expert network
Workforce geography	Not disclosed beyond “global network.” No country list, no delivery-centre map
Ownership / independence	Private, VC-backed, roughly $67.5M total. Latest is a ~$13.5M Series B extension in June 2025 led by Dell Technologies Capital, bringing Series B to $50M; the initial $36M Series B (late 2024) was led by Socium Ventures with NVIDIA, Databricks Ventures and Play Time Ventures
Best for	Teams that want the deepest customer-review evidence in the category alongside a real trust centre
Not for	Buyers with a data-residency obligation, since annotator geography is undisclosed
Modalities	Image, video, text, audio, document, multimodal, LLM data
Specialist domains	Computer vision, LLM fine-tuning and evaluation, healthcare imaging support
Core features	Annotation platform, orchestration and QA tooling, expert network, model evaluation workflows
Quality method	Multi-stage review inside the platform with configurable QA steps. No published IAA threshold
Security & compliance	The strongest certification evidence in the platform tier: SOC 2 Type II, ISO/IEC 27001:2022, GDPR, CCPA, with the SOC 2 Type II and ISO 27001:2022 reports downloadable on request through a clean trust centre. Important caveat: a sentence on their marketing security page listing US health-data privacy, PCI DSS, ISO 27017, 27018 and 9001 describes their authentication vendor’s certifications, not SuperAnnotate’s, and those are not attributed to SuperAnnotate here
Pricing posture	Quote-only. Three tiers with no dollar figures; the only quantified differentiator is compute hours (1k / 2.5k / 10k)
Typical engagement / minimums	Not disclosed
Watch-out	Client disclosure is thin relative to the review profile: Databricks and ServiceNow are named, against a G2 sample larger than any peer’s

Review attribute	Detail
G2	4.9/5 (n=284), accessed 6 August 2026. The review-volume outlier of the entire category, against Labelbox at n=48 and Encord at n=61
Capterra	5.0/5 (n=2), accessed 6 August 2026. Below threshold; not a rating
Evidence quality	The strongest customer-side sample in the roster. It also demonstrates that review volume is not evenly earned across this market: a 4.9 from 284 and a 4.9 from 15 are not the same claim
What users like	Annotation throughput tooling and responsive support (G2)
What users flag	Learning curve on advanced orchestration features (G2)
Checked, nothing found	TrustRadius, Gartner Peer Insights, Trustpilot, Glassdoor. No usable figure as of 6 August 2026.

What it is. A 2018-founded platform and expert network founded by brothers Vahan and Tigran Petrosyan, headquartered in San Francisco with substantial engineering in Yerevan.

Best for. Buyers who weight third-party review evidence and want reports available on request. Not for programmes with a hard data-residency requirement.

What customers say. The only large customer sample in the platform tier. Note the asymmetry: the review footprint is far larger than the named-client footprint, which is unusual and worth asking about in references.

20. Encord

Attribute	Detail
Delivery model	Platform-first with managed services attached
Workforce model	Partner and expert network
Workforce geography	Not disclosed. No annotator location, delivery-centre or country information anywhere on encord.com, despite selling managed annotation services
Ownership / independence	Private, VC-backed, $110M total. Latest is a $60M Series C led by Wellington Management in February 2026; prior $30M Series B in August 2024 led by Next47. Y Combinator-backed, London with a San Francisco office
Best for	Physical AI work: robotics, autonomous vehicles and drone data
Not for	Buyers who need the SOC 2 Type stated before signature
Modalities	Image, video, 3D/LiDAR, DICOM and NIfTI, geospatial, ECG, text, multimodal
Specialist domains	Physical AI, robotics, autonomous vehicles, medical imaging
Core features	Annotation, Index, Active, Data Agents, plus a services arm covering annotation, data collection and physical AI data
Quality method	Active provides quality analytics over labelled data. No published IAA threshold
Security & compliance	The “narrow claims, real evidence” end of the spectrum: AICPA SOC 2 (stated as having successfully completed its SOC 2 examination), plus US health-data privacy and GDPR self-attested through Vanta automation rather than third-party audit certificates. Type I against Type II is not stated on the security page, so a Type II claim cannot be supported from that page alone. Explicitly not claimed: ISO 27001, ISO 27701, CCPA, PCI DSS. Published specifics: AES-256 at rest, TLS 1.2/1.3 in transit, US and EU deployment regions, VPC and on-prem options
Pricing posture	Quote-only. Three tiers, a long feature matrix, and zero dollar figures; no currency symbol appears on the rendered page. Quantified tiering is data volume only (up to 500k / 100m / 1bn+)
Typical engagement / minimums	Not disclosed. Many items are marked “add on” and unpriced, including on-prem, DICOM and NIfTI, geospatial, ECG, 3D/LiDAR, LLM evaluations and a solutions architect
Watch-out	Selling managed annotation while disclosing no annotator geography is the single biggest gap in an otherwise strong evidence posture. Encord’s own About page also publishes no founding date and no address

Review attribute	Detail
G2	4.8/5 (n=61), accessed 6 August 2026
Evidence quality	n=61 is a usable sample and one of the better ones in this article
What users like	Video and medical imaging workflows, and support responsiveness (G2)
What users flag	Pricing opacity and add-on structure (G2)
Checked, nothing found	Capterra, TrustRadius, Gartner Peer Insights, Trustpilot, Glassdoor. No usable figure as of 6 August 2026.

What it is. A 2020-founded London platform company, Y Combinator-backed, pivoting hard toward physical AI data in 2026 while keeping a managed services arm.

Best for. Robotics, AV and drone programmes that need specialised modality support. Not for buyers who cannot proceed without a stated SOC 2 Type.

What customers say. A usable sample by this category’s standards. UiPath appears as a first-party case study, alongside 18 customer stories of which most are anonymised by industry, against a claim of 300+ top AI teams.

21. V7 (V7 Darwin)

Attribute	Detail
Delivery model	Platform-first with a vetted partner and expert network
Workforce model	Contracted specialists including qualified radiologists and pathologists. V7 does not employ the labelers; it sources, manages and contracts them
Workforce geography	Not disclosed. “Various geographies,” none named
Ownership / independence	Independent, VC-backed. Latest is a $33M Series A on 28 November 2022 co-led by Radical Ventures and Temasek, with Air Street, Amadeus and Partech; roughly $36M total. No Series B as of 6 August 2026
Best for	Medical and document workflows that need clinician-level annotators and a full security page
Not for	Buyers who need the annotators to be employees of the vendor
Modalities	Image, video, medical imaging, document, text
Specialist domains	Medical and clinical, life sciences, insurance and document processing, industrial inspection
Core features	Darwin platform, auto-annotation, workflow orchestration, contracted specialist network
Quality method	Review stages within Darwin workflows. V7’s published cost-reduction percentages are vendor self-reported and are not printed here as benchmarks
Security & compliance	SOC 2 Type II (audited), ISO 27001, US health-data privacy, GDPR, on a properly published security page. No ISO 27701, CCPA or PCI DSS. The services page adds “certified” annotators for health data and “FDA-compliant annotation workflows”
Pricing posture	Quote-only. Custom package based on platform, users and data volume
Typical engagement / minimums	Not disclosed
Watch-out	The annotators are contracted rather than employed, which matters for continuity and for confidentiality controls. Several client names circulating in third-party databases were not confirmed first-party and are not printed here

Review attribute	Detail
G2	4.8/5 (n=54), accessed 6 August 2026
Evidence quality	n=54 is usable. Combined with a full security page, V7 has one of the better evidence positions in the platform tier
What users like	Medical imaging workflows and annotation speed (G2)
What users flag	Pricing structure clarity for smaller teams (G2)
Checked, nothing found	Capterra, TrustRadius, Gartner Peer Insights, Trustpilot, Glassdoor. No usable figure as of 6 August 2026.

What it is. A London platform company founded in 2018 (with an earlier San Francisco predecessor), selling Darwin plus a contracted network of specialists that includes clinicians.

Best for. Regulated medical and document workflows where the annotator’s professional qualification matters. Not for buyers whose contract requires employed annotators.

What customers say. A usable sample, and combined with a full security page it gives V7 one of the better evidence positions in the tier. Named clients include Genmab, Miovision, Abyss Solutions, CattleEye, Imidex, InformAI, Intelligent Ultrasound, Raptor Maps, Safari AI, Digital Divide Data, Velmeni, Yembo and Franklin.ai.

22. Kili Technology

Attribute	Detail
Delivery model	Platform-first hybrid; labeling services are an optional paid add-on
Workforce model	Not disclosed
Workforce geography	Not disclosed. Deployment regions (cloud, on-prem, air-gapped) are infrastructure, not labour
Ownership / independence	Private, VC-backed. Seed €5.7M in January 2021 (Serena Capital and Headline), Series A $25M led by Balderton Capital. No valuation disclosed. Paris-registered SAS
Best for	European enterprises needing on-prem or air-gapped deployment with a live trust centre
Not for	Buyers whose evaluation requires knowing where the labeling workforce sits
Modalities	Image, video, text, document, audio, satellite and geospatial
Specialist domains	Financial services and KYC, insurance, defence and geospatial, gaming
Core features	Annotation platform, quality workflows, on-prem and air-gapped deployment, optional dedicated labeling services
Quality method	Platform review and consensus workflows. No published IAA threshold
Security & compliance	SOC 2 Type II, ISO 27001:2022, US health-data privacy, with annual audits, plus a live trust centre. No ISO 27701 or PCI DSS
Pricing posture	Partially published: free trial at $0 per month (1 seat; 100 assets for text, document and image, 5 for video and satellite); Grow is a custom subscription (up to 20 seats, up to 50,000 assets); Enterprise is a custom contract. Net effect: tier structure published, no dollar rates
Typical engagement / minimums	Grow caps at 20 seats and 50,000 assets; Enterprise is negotiated
Watch-out	Workforce is undisclosed while labeling services are sold as an add-on, which is the same gap Encord has

Review attribute	Detail
G2	4.7/5 (n=53), accessed 6 August 2026. The n=49 that circulates is an older crawl of the same page
Evidence quality	n=53 is usable, and mid-range for the platform tier
What users like	Deployment flexibility and quality workflows (G2)
What users flag	Onboarding complexity for smaller teams (G2)
Checked, nothing found	Capterra, TrustRadius, Gartner Peer Insights, Trustpilot, Glassdoor. No usable figure as of 6 August 2026.

What it is. A 2018-founded Paris platform company selling annotation software with on-prem and air-gapped options, plus dedicated labeling services as a paid add-on.

Best for. European enterprises with deployment constraints and a real security review. Not for anyone who needs the workforce’s location documented.

What customers say. A usable sample, though the exact figure needs a live read. Written case studies are stronger evidence than a logo wall here, covering LCL Bank (KYC document extraction), Covéa, Eidos-Montréal, Jellysmack and Enabled Intelligence. One neutral note for the record: Kili is the single publisher in this SERP whose own roundup discloses that it ranks itself. That is a fact about the SERP rather than an endorsement, and it belongs here because it is the honest counterpart to the disclosure at the top of this article.

F. Open-source, self-serve and cloud-native

The options the vendor-authored roundups have the least reason to mention, and the only tier where pricing is mostly published.

23. Label Studio (HumanSignal)

Attribute	Detail
Delivery model	Open source, plus SaaS, plus managed data services
Workforce model	Self-hosted (you) or vendor-managed; “native speakers in 91 countries” for the managed services arm
Workforce geography	Partially disclosed. The 91-country claim covers language coverage; specific delivery-centre or physical data lab locations are not disclosed. HumanSignal publishes no company address on its own site
Ownership / independence	Private, VC-backed, $30M total from Redpoint Ventures, Unusual Ventures, Bow Capital and Swift Ventures. No round-by-round breakdown, no dates and no valuation published, so whether $30M is current cannot be confirmed first-party
Best for	Teams that want a permissively licensed self-hosted core with an optional paid cloud tier
Not for	Buyers with a formal certification checklist, because the disclosure is the weakest in the platform tier
Modalities	Image, video, text, audio, time series, document, multimodal
Specialist domains	General-purpose; strongest in NLP and audio community usage
Core features	Label Studio Community Edition (self-hosted, free), Label Studio Enterprise (SaaS or on-prem), managed data services covering from-scratch data creation, annotation and evaluation by domain experts
Quality method	Configurable review and agreement tooling in Enterprise, including LLM-as-a-judge. No published IAA threshold
Security & compliance	Weakest formal disclosure in the platform tier. Enterprise is described as SOC 2 and US health-data privacy compliant cloud or on-prem, with SOC 2 and health-privacy reporting as an Enterprise-only feature and no SOC 2 Type stated. The security page describes controls (TLS enforced, encryption at rest, SAML 2.0, separated data and control planes) but names no certification, with details behind a gated brief. Not claimed anywhere: ISO 27001, ISO 27701, GDPR, CCPA, PCI DSS, and the absence of any GDPR statement is notable given every EU-facing peer carries one
Pricing posture	Published, one of only six: Community Edition free and self-hosted; Starter Cloud $99 per month base plus $49 per month for each additional user (up to 12 users, free trial); Enterprise custom. Managed data services are quote-only. HumanSignal acquired Erud AI in November 2025 and launched the data services division on the back of it
Typical engagement / minimums	Starter Cloud caps at 12 users. Enterprise is purchasable through AWS Marketplace against committed AWS spend
Watch-out	Licence: Apache 2.0, verified from the repository. Permissive, no copyleft, no source-available commercial restriction, which materially differentiates it from source-available competitors. Also: three mutually inconsistent user counts appear across their own site, so none is cited here

Review attribute	Detail
G2	Figures not visible as of 6 August 2026; live read required
Capterra	4.0/5 (n=2), accessed 6 August 2026. Below threshold; not a rating
Evidence quality	A structural finding worth stating once: open-source tools with developer user bases barely register on commercial B2B review sites. The absence reflects buyer behaviour, not product quality
What users like	Flexibility of the labeling interface configuration, from the small Capterra sample
What users flag	Setup effort for self-hosting, from the small Capterra sample
Checked, nothing found	TrustRadius, Gartner Peer Insights, Trustpilot, Glassdoor. No usable figure as of 6 August 2026.

What it is. A 2019-founded company (formerly Heartex) publishing the most widely used open-source labeling interface, with a paid cloud tier and a managed data services arm on top.

Best for. Engineering teams that want to self-host under a permissive licence and buy support later. Not for procurement processes that require a named certification with a stated Type.

What customers say. Almost nothing on the review platforms, which is normal for open-source tooling and should not be read as a quality signal in either direction. Named clients include Bombora, Geberit, Outreach, Trivago and Zendesk.

24. CVAT.ai

Attribute	Detail
Delivery model	Open source, plus managed cloud, plus enterprise self-host, plus first-party annotation services
Workforce model	Self-hosted (you) or first-party Computer Vision and Audio Annotation Services teams
Workforce geography	Not disclosed for the labeling-services workforce. Registered agent in Wilmington DE; EU entity CVAT.ai Europe Ltd in Paphos, Cyprus; cvat.ai publishes no HQ city
Ownership / independence	Independent. No funding round announced on cvat.ai, and no first-party or credible-press priced round located. Project started inside Intel in 2017, open-sourced 2018, spun out as CVAT.ai Corporation in 2022 by ex-Intel engineers Nikita Manovich and Boris Sekachev
Best for	Computer vision teams that want fully published pricing and a permissively licensed core
Not for	Regulated buyers, because no security certification is claimed at all
Modalities	Image, video, 3D point cloud, audio
Specialist domains	General computer vision; audio annotation as a first-party service
Core features	CVAT Community (self-hosted OSS), CVAT Online (managed cloud), CVAT Enterprise (self-hosted commercial), first-party annotation services teams
Quality method	Built-in quality control and honeypot-style validation in the platform. No published IAA threshold
Security & compliance	The sharpest contrast in the open-source tier: GDPR, CCPA and EU AI Act compliance only. No SOC 2, no ISO 27001, no ISO 27701, no US health-data privacy statement and no PCI DSS anywhere on the pricing or enterprise pages
Pricing posture	Fully published: Free $0; Solo $33 per month monthly or $23 billed yearly; Team $33 per user per month monthly with a 2-seat minimum ($66) or $23 per user per month yearly (2-seat minimum, $46); Enterprise Basic and Premium are now contact-sales on the live pricing page (15 October 2026); the $12,000-per-year Basic price printed in August is no longer shown
Typical engagement / minimums	2-seat minimum on Team; Team caps at 50 members
Watch-out	Licence: MIT for CVAT Community, with a caveat published in the same README. Code under /serverless is MIT but bundles third-party assets under separate licences including non-commercial restrictions, and FFmpeg components are LGPL or GPL. A free and open-source option with non-commercial components inside it is exactly the detail a legal team finds late

Review attribute	Detail
G2	4.5/5 (n=31), accessed 6 August 2026, read live. A conflicting 4.6 to 4.8 appears elsewhere; the platform page figure is the one used here
Evidence quality	One of the few figures in this article read live rather than from a search index. n=31 is usable but modest
What users like	Annotation speed for video and the value of the free tier (G2)
What users flag	Self-hosting maintenance burden (G2)
Checked, nothing found	Capterra, TrustRadius, Gartner Peer Insights, Trustpilot, Glassdoor. No usable figure as of 6 August 2026.

What it is. The Computer Vision Annotation Tool, started inside Intel in 2017 and now an independent company selling managed cloud and enterprise self-host alongside the open-source core.

Best for. Computer vision teams that want to model cost precisely before committing. Not for buyers whose security review requires SOC 2 or ISO 27001.

What customers say. One of the few figures in this article confirmed by live read. Client evidence is unusual: CVAT publishes named individuals with job titles rather than corporate logos, and only two of the four are identifiable companies.

25. Roboflow

Attribute	Detail
Delivery model	Platform (self-serve plus enterprise) with an optional managed labeling add-on
Workforce model	Undisclosed for the labeling add-on
Workforce geography	Not disclosed. Roboflow also publishes no street address or HQ city on its own pages
Ownership / independence	Private, VC-backed. $40M Series B on 19 November 2024 led by GV, with Craft Ventures and Y Combinator; valuation not disclosed. Prior ~$2.1M seed (January 2021) and $20M Series A (September 2021, Craft Ventures)
Best for	Computer vision teams that want to start labeling today at a published rate
Not for	Buyers needing ISO 27001 or GDPR statements in the compliance section
Modalities	Image, video
Specialist domains	Industrial inspection, agriculture, sports analytics, logistics
Core features	Annotate, Train, Deploy, Workflows, Universe; optional managed labeling
Quality method	Platform review workflows. No published IAA threshold
Security & compliance	SOC 2 and US health-data privacy only. No ISO 27001, ISO 27701, PCI DSS or GDPR appears in the Compliance section, and the badge reads “SOC 2” without specifying Type publicly, though a report is available on request
Pricing posture	Published, and changed on 18 September 2026: the Core plan now starts at $0 with 10 credits and unlimited seats, with paid credits on top (the live page should be read before budgeting, because tier prices moved twice in two months); data labeling services from $0.10 per bounding box; Enterprise custom
Typical engagement / minimums	Self-serve from free; no annual minimum below Enterprise
Watch-out	The labeling add-on’s workforce is undisclosed, so the vendor’s transparency on pricing is not matched on labour

Review attribute	Detail
G2	4.7/5 (n=108), accessed 6 August 2026. One of only two ratings in the roster built on more than 100 customer reviews, the other being SuperAnnotate’s n=284. A conflicting 4.8 appears elsewhere; the platform page figure is used here
Evidence quality	A proof point for this article’s review-vacuum finding. Two vendors out of 26, Roboflow and SuperAnnotate, carry a customer rating at a sample size that would be unremarkable in any other software category
What users like	Speed from dataset to deployed model, and documentation quality (G2)
What users flag	Credit consumption predictability at higher volumes (G2)
Checked, nothing found	Capterra, TrustRadius, Gartner Peer Insights, Trustpilot, Glassdoor. No usable figure as of 6 August 2026.

What it is. A 2020-founded computer vision platform (Y Combinator S20) covering annotation through deployment, with labeling available as a priced add-on rather than a sales conversation.

Best for. Computer vision teams that want to move today without procurement. Not for buyers whose compliance section has to include ISO 27001 or an explicit GDPR statement.

What customers say. One of only two customer-review positions in this roster built on more than 100 reviews. Trust-centre named clients include John Deere, Cardinal Health, USG, Intel, Rivian, Medtronic and Column, with full case studies for BNSF Railway, USG and Fletcher Sports.

26. AWS SageMaker Ground Truth

Attribute	Detail
Delivery model	Cloud-native service, not a company
Workforce model	Originally three options: Amazon Mechanical Turk (open crowd), third-party vendors through AWS Marketplace (managed service), or your own private workforce. Mechanical Turk was reported closed on 30 September 2026 after AWS told workers on 25 August, and AWS lists it with Ground Truth among the features moved to maintenance, so the open-crowd option should be treated as gone. The private-workforce path is the concrete BYO example the three-purchases framework needs
Workforce geography	Not disclosed as a service attribute. The MTurk crowd is global, and Marketplace vendor workforces vary by listing
Ownership / independence	AWS Inc., Seattle; parent Amazon.com (Nasdaq: AMZN). The largest and most independently verifiable owner in the roster, and also a hyperscaler that may compete with you elsewhere
Best for	Teams already using Ground Truth on AWS who want to keep the data and the workforce inside their own account
Not for	New buyers, since the service closed to new customers on 30 July 2026; and anyone who needs the certification to cover the labeling workflow rather than the cloud underneath it
Modalities	Image, video, text, 3D point cloud
Specialist domains	General-purpose; strongest where the data already sits in S3
Core features	Labeling job orchestration, three workforce options, automated data labeling, 3D point cloud workflows
Quality method	Consensus and annotation consolidation built into labeling jobs. No published IAA threshold
Security & compliance	A useful nuance: Ground Truth inherits the AWS compliance programme (SOC 2, ISO 27001, ISO 27701, US health-data privacy eligibility, PCI DSS, GDPR), but none of it is asserted on the Ground Truth page itself. “Your cloud already does this” also means the certification belongs to the cloud, not to the labeling workflow
Pricing posture	Fully published (US East, N. Virginia): reviewed objects per month under 50,000 at $0.08 each; 50,000 to 1,000,000 at $0.04; above 1,000,000 at $0.02. 3D point clouds: $3.00 per single frame; $3.00 for the first frame of a sequence and $1.50 thereafter. MTurk labour, suggested price per labeler: image classification $0.012, text classification $0.012, NER $0.024, bounding box $0.036, semantic segmentation $0.84. Vendor workforce cost per label is set by the vendor and published per listing in AWS Marketplace. Note for verification: those rate tables rendered empty in a live browser on 6 August 2026; the figures were read from AWS’s own backing pricing feed and are authoritative, but a spot-checker will see blank tables
Typical engagement / minimums	Pay as you go; no minimum
Watch-out	Two dated facts from AWS’s own service availability notice, and they are different products. First: Ground Truth Plus, the AWS-managed labeling workforce, reached end of support on 30 June 2026 and is no longer available. Roundups still recommending it as a live managed option, including for its old cost-saving claim, are recommending something you cannot buy. Second: self-service SageMaker Ground Truth, the service profiled here, moved to maintenance and closed to new customers on 30 July 2026. AWS states that customers already using it can continue to do so. If you are not already on it, this entry is a reference point rather than an option

Review attribute	Detail
G2	4.1/5 (n=19), accessed 6 August 2026
Employee sentiment (Glassdoor)	Not applicable at the service level
Evidence quality	n=19 is modest but usable, and it is the lowest customer rating among the vendors in this roster carrying a usable sample
What users like	Integration with the rest of the AWS stack (G2)
What users flag	Console usability and job configuration complexity (G2)
Checked, nothing found	Capterra, TrustRadius, Gartner Peer Insights, Trustpilot. No usable figure as of 6 August 2026.

What it is. An AWS service launched at re:Invent in November 2018, offering labeling job orchestration on top of three different workforce options rather than a workforce of its own.

Best for. Teams already using Ground Truth on AWS who want to staff labeling with their own people on rented tooling. Not for new buyers, because AWS closed the service to them on 30 July 2026, or for anyone who needs a certification scoped to the labeling workflow itself.

What customers say. The lowest usable customer rating in this roster. Named clients include T-Mobile, the National Football League, AstraZeneca and Tyson Foods.

📊 2 of 26 vendors carry a customer rating built on more than 100 reviews. SuperAnnotate, G2 4.9 (n=284) and Roboflow, G2 4.7 (n=108). Scale AI shows n=1, Surge AI n=0, TELUS Digital n=0. Source: Forage AI review sweep, accessed 6 August 2026.

We printed the review sample size for every vendor in this roster, and for several of the biggest names the honest answer was that there weren’t any. Review scores cannot carry your shortlist in this category. They mostly measure which vendors run a customer-marketing motion, and the vendors with the largest delivery footprints run the smallest ones.

Attribution: Forage AI research team, review sweep of 28 providers, 6 August 2026

Attribution: Forage AI research team, four-sweep vendor audit of 28 providers, 6 August 2026

What do data annotation services actually cost?

The SERP keeps the roster and the cost benchmarks in separate articles, which is convenient for publishers and useless for buyers. Here they are together.

Per-unit benchmarks by modality

Task type	Published 2026 benchmark range
Image classification	$0.02 to $0.15 per label
Object detection (bounding box)	$0.05 to $0.90 per object
Semantic segmentation	$0.50 to $2.00+ per mask
Text sentiment or intent classification	$0.01 to $0.10 per item
Named entity recognition	$0.05 to $0.25 per entity
Audio transcription	$0.50 to $3.00 per minute
Video annotation	$1.00 to $10.00 per minute
Medical or clinical items	$1.00 to $5.00+ per item
Hourly managed rates	$6 to $60+ per hour

The verifiable published rates, read from vendor pages on 6 August 2026, are narrower and more useful because you can check them. AWS SageMaker Ground Truth charges $0.08, $0.04 and $0.02 per reviewed object across its three volume tiers, and $3.00 per 3D point-cloud frame. CVAT.ai runs $23 to $33 per user per month, with Enterprise now quoted on request. Roboflow’s Core plan starts at $0 since 18 September 2026, with labeling from $0.10 per bounding box. Label Studio charges $99 per month for Starter Cloud plus $49 per additional user. Label Your Data publishes $0.02 per bounding-box object, $0.015 per keypoint, $0.02 per NLP entity and $6 per annotator hour.

6 of 28 vendors publish real pricing figures.

The other 22 are quote-only, so budget for a procurement cycle rather than a checkout. Source: Forage AI pricing sweep, 28 vendors, accessed 6 August 2026.

The five pricing models

Model	What you pay for	What it hides	When it fits
Per-unit	Each label, box, entity or minute	Rework, QA overhead, setup, minimums	Stable taxonomy, predictable volume
Hourly	Annotator time	Throughput variance between cohorts	Ambiguous or exploratory work
Per-project	A defined deliverable	Change-order pricing when scope moves	Fixed-scope one-offs
Dedicated FTE	A named cohort’s capacity	Idle time you still pay for	Long programmes with continuity needs
Platform seat or usage	Software access and consumption	Your own labour cost, which is the larger number	You already employ annotators

Common misconception: expert-network published rates are not prices.

They publish what they pay contractors, not what they charge clients. Mercor posts roles at $60 to $120 per hour, Handshake AI publishes $40 to $125 per hour for fellows, Surge AI up to $200 to $400 per hour for a journalist role, and Turing publishes expert-side rates. Every one of those is supply-side. It reads as pricing transparency and it is the opposite: it tells you the vendor’s cost of goods and nothing about your invoice.

The landed-cost model

The per-unit rate is not the number that decides your bill. Landed cost is:

per-unit rate × rework multiplier + QA overhead + setup and minimums + guideline versioning + re-training the annotator cohort

Worked example. Vendor A quotes $0.01 per record at a 20% error rate. Vendor B quotes $0.04 per record at 1%. On 1,000,000 records, Vendor A’s direct cost is $10,000 and Vendor B’s is $40,000, and Vendor A looks four times cheaper. Then add rework: 200,000 records go back through at least once, which is another $2,000 of labeling and, more expensively, a QA cycle to find them, a guideline revision to prevent recurrence, and re-training for the cohort. Your own reviewers absorb that time at a fully loaded cost that dwarfs the labeling line. Vendor B’s 1% error rate is not a quality nicety. It is the reason the cheaper quote costs more.

Two conditional rules fall out of that. Below roughly 50,000 units, a managed vendor’s minimums usually dominate the economics, so a platform plus internal labour wins. And if the taxonomy is still moving, a fixed per-unit contract prices the wrong thing, because you will pay full rate to label a schema you are about to change. This is the same total-cost-of-ownership framing that applies to any managed data operation: the visible line item is rarely the expensive one.

On the labour side, rates stratify into bands that barely overlap. AFP’s October 2025 reporting documented $0.05 to $0.25 per task at the bottom, with some Remotasks work around $0.01 per task, against $30 to $50 per hour for specialised annotation. Entry-level gig work sits around $15 to $20 per hour where it is hourly at all. At the top, Mercor pays contractors $85 to $95 per hour on average, and Handshake AI’s published fellowship postings reach $125 per hour with higher unpublished bands in medicine, law and finance. Expert RLHF and medical review is often priced per example, in the $50 to $100 range. If a quote for expert-domain work looks like commodity pricing, one of the two numbers is wrong.

There is peer-reviewed evidence for why label error is expensive rather than merely annoying. Northcutt, Athalye and Mueller found label errors averaging at least 3.3% across the test sets of ten of the most commonly used computer vision, NLP and audio benchmarks, with at least 6% of the ImageNet validation set mislabelled. The consequence that matters to procurement is that model rankings flip: on corrected ImageNet labels, ResNet-18 outperforms ResNet-50 once the share of originally mislabelled test examples rises by just 6%. Label quality is not a soft contract concern. It changes which model you ship.

Attribution: Curtis Northcutt, Anish Athalye and Jonas Mueller, “Pervasive Label Errors in Test Sets Destabilize Machine Learning Benchmarks”, NeurIPS 2021 Datasets and Benchmarks Track.

What gets annotated: modality coverage

Modality is the fastest first filter on a long vendor list. Computer vision covers bounding box, polygon, semantic segmentation and instance segmentation, keypoint and landmark work, 3D cuboids and LiDAR point clouds, and video object tracking. Text and NLP covers NER, sentiment and intent classification, and increasingly RLHF, SFT, preference and rubric data and RL environments. Audio covers transcription and diarization. Documents cover OCR and key-value extraction. Multimodal sits across all of it.

Two method entities show up in almost every modern engagement and change the economics: model-assisted pre-labeling, where a model proposes and humans correct, and active learning, where the system chooses what to send to a human next. Neither removes the human. Both move where the human’s time goes.

The single most useful number in this market is not the market size. It is the sourcing split. Teams that assume in-house is the conservative default are choosing the minority position, and the reasons for choosing it should be specific rather than reflexive.

Attribution: Grand View Research, Data Labeling Solution and Services Market 2025–2030, accessed 6 August 2026.

How should you run the evaluation?

Label error is measurable before you sign, and measuring it is a procurement step rather than a research one.

No ranking roundup gives you a pilot protocol, an acceptance criterion or a contract clause. They all end at “contact them.” Here is what to do between the shortlist and the signature.

A two-week, three-vendor pilot protocol

Run two or three vendors in parallel, on one or two task types, inside two weeks. NDA and DPA signed before any sample data moves, without exception. A batch of roughly 500 units returned in five to ten business days from NDA is what a mature vendor should manage; the ones that cannot are telling you about their queue.

Score the output against your own benchmark set with real ground truth, not the vendor’s gold set, because a vendor scoring itself on its own gold items is measuring its ability to hit its own targets. Use a quality framework for external data sources to define what “correct” means before the batch goes out. Agreeing acceptance criteria after seeing the output is how pilots turn into arguments.

Measuring quality: gold sets, consensus, and IAA thresholds

Write the accuracy floor as a kappa or alpha threshold against a named gold set, not as a percentage. Percentages hide task difficulty and hide chance agreement; agreement statistics do not.

The standard interpretation bands come from Landis and Koch (1977): 0.00 to 0.20 slight, 0.21 to 0.40 fair, 0.41 to 0.60 moderate, 0.61 to 0.80 substantial, and 0.81 to 1.00 almost perfect.

0.61 to 0.80 is “substantial” inter-annotator agreement.

Write the accuracy floor as a kappa threshold against a named gold set, not as a percentage. Source: Landis & Koch, Biometrics 33(1), 1977, 159–174.

For anything subjective, a result under 0.80 is where you start asking questions rather than where you sign. Specify which statistic you mean, because Cohen’s kappa, Fleiss’ kappa and Krippendorff’s alpha are not interchangeable and a vendor reporting “our IAA is 0.9” without naming the metric has told you nothing.

Guideline authorship and versioning is the largest single driver of label quality and the most common cause of a failed engagement, and it is almost never in the contract. Taxonomies drift, so the acceptance threshold gets re-measured on a cadence rather than set once at kickoff.

The RFP: what to send, and what to ask for back

Of the data labeling RFP templates ranking for this search in October 2026, the useful ones sit behind a demo form. A useful RFP is short. It tells the vendor exactly what you are labelling, and it asks for documents rather than adjectives, so the replies can be compared line by line and the disqualifiers fire before anyone books a call.

RFP section	What you send	What the vendor must return
Scope and spec	Modality, volume by month, the draft taxonomy, a sample of your hardest edge cases, target start date	A throughput and ramp plan for your volume, and which taxonomy items they would push back on
Ownership	The who-owns-you question, and your own sector	Owner and any parent, and whether they serve a competitor of yours
Workforce	Your data-residency and confidentiality constraints	Named facilities or countries, employment status of annotators, every subcontractor
Security	Your NDA and DPA, signed before any sample moves	The certification report itself, its Type, period and systems in scope
Quality	Your acceptance metric and threshold, and who holds the gold set	A QA plan naming the sampling rate, strata and the agreement statistic
Pricing schedule	Volume bands and the rework you expect	Unit price at each band, rework price, setup fees, minimums, guideline-change charges
Pilot	About 500 units and your benchmark set	Turnaround commitment and agreement to be scored on your set, not theirs
Contract and exit	The SLA clause skeleton below	Redlines against it, plus export formats and what happens to the trained cohort at exit

Score the returns in two passes. First pass is pass or fail against the six disqualifiers further down; second pass is the pilot result against your own benchmark set. Weighted scorecards come last, if at all. The twelve questions below are the ones every section above has to answer, in the order that disqualifies fastest.

1. Who owns you, and who else do you serve?
2. Where does the labour physically sit, and can you name the facilities?
3. Are annotators employed, contracted, or crowd-sourced?
4. Do you subcontract, and to whom?
5. SOC 2: Type I or Type II, what period does the report cover, and what systems are in scope?
6. Can I have the report?
7. Who authors and versions the guidelines?
8. What is your QA sampling rate, and which strata does it cover?
9. Do you report inter-annotator agreement, on what metric, at what threshold?
10. What happens to the trained annotator cohort if I leave?
11. Who owns the labels, the guidelines and the ontology?
12. What export formats, and can I get the raw annotations plus the audit trail?

The SLA clause skeleton

Eight clauses. Most annotation contracts contain two of them.

1. Accuracy floor expressed as a Landis-Koch kappa threshold against a named gold-standard set, with the statistic specified.
2. Measurement method: sampling rate, strata covered, who measures, and how often it is re-measured as guidelines change.
3. Gold-standard definition: who builds it, who holds it, and how it is refreshed to prevent gold-set overfit.
4. Rework terms: what triggers rework, what the turnaround is, and who pays.
5. Ownership: labels, guidelines and ontology, named explicitly, because “the deliverable” is not the same as “the ontology we developed together.”
6. Subcontracting disclosure: prior written notice, with the right to refuse a named fourth party.
7. Access controls: named-individual annotator logins, no shared accounts, and an audit trail of who labelled what.
8. Exit and portability: export formats, the audit trail, and what happens to the trained cohort at termination.

The failure modes nobody writes down

Annotator churn mid-project, which resets everything the cohort learned about your edge cases. Silent subcontracting to a fourth party you never approved. QA sampling that only inspects the easy strata, so the measured quality is real and irrelevant. Gold-set overfit, where annotators learn the gold items and the measured quality stops predicting real quality. And the pilot-team bait-and-switch, where the A-team runs your pilot and the B-team runs your production volume. Contract for the cohort, not just for the vendor.

Once you have selected and signed, running the programme is a separate discipline, and the operational playbook for running a training-data program covers the part that starts after procurement ends.

Every kappa threshold in every annotation SLA descends from one 1977 paper, and almost no annotation contract cites it. The gap between an industry that quotes those bands informally and an industry that writes them into contracts is where most annotation disputes live. Named-individual logins and an audit trail are cheap clauses to write and expensive ones to be missing, as the June 2025 document-exposure report showed.

Attribution: J. Richard Landis and Gary G. Koch, “The Measurement of Observer Agreement for Categorical Data”, Biometrics 33(1), 1977, 159–174.

When should you not hire a data labeling company?

All of that assumes you are hiring somebody. Four cases where the answer is not a vendor at all.

1. Keep it in-house. When the domain expert is already on your payroll and cannot be replicated externally, when the volume is small enough that a vendor’s minimum exceeds the work, when the taxonomy is still pre-product-market-fit, or when the data legally cannot leave your jurisdiction. Grand View Research puts in-house at roughly 15% of the 2024 market, so this is the minority case rather than the default.

2. Licence a finished dataset. Licensing a finished dataset somebody has already built is a different purchase with different economics and a much shorter time to first usable data, at the cost of any claim to a proprietary asset.

3. Programmatic or weak supervision. Write label functions and heuristics instead of hiring labelers. Snorkel AI is the reference implementation, named here rather than profiled because it is a different purchase, not a different vendor.

4. You do not have the data yet. Every vendor in this article assumes you already possess the corpus to be labelled, and sourcing it is a separate purchase that gets specialised fast in regulated verticals, which is what our 2026 guide to buying legal data covers for court records, dockets and case law.

Auto-labeling and synthetic data have not removed the human. They moved the human to audit, calibration and edge cases. Peer-reviewed work on LLM-assisted annotation with active learning, including AutoLabel presented at KSEM 2024 and published in the ACM Digital Library, demonstrates both the automated path and its limits, which is a better anchor than the 10x speedup and 75% savings figures circulating in this market, all of which come from companies selling the automation. Whether you need labelled data at all is worth settling before any of this, and that decision often turns on fine-tuning versus retrieval.

The strongest available counter-signal to “automation solved it” is the Meta case above. The company that paid a reported $14.3 billion for 49% of Scale AI reassigned roughly 6,500 of its own engineers to write training data by hand in March 2026, and conceded the restructuring had gone wrong three months later.

Two data points sit in tension and both are true. Grand View Research puts 84.6% of the 2024 market in the outsourced segment, which says the market has settled the build-versus-buy question. Meta’s March 2026 in-housing says the settlement does not hold at the top of the value curve, where the annotation is expert judgement rather than throughput. The practical read: outsource the volume, keep the judgement, and be explicit about which of the two a given task is before it goes into an RFP.

Attribution: Grand View Research, Data Labeling Solution and Services Market 2025–2030, set against reporting on Meta’s Applied AI Engineering unit, March to June 2026.

How to choose: routing your shortlist

You came for the list, and the list is above. Here is the repeatable method for turning 26 names into three.

Start with the three questions. Everything in this article converges on the same three, and they shorten a shortlist faster than any feature comparison because each one is answerable in a week and each one produces a yes or a no rather than a score:

1. Who owns you, and who else do you serve? The criterion this market produced in 2025 and 2026. Meta inside Scale AI, iMerit inside EXL, TIXT delisted, clickworker inside LXT, Shaip inside Ubiquity. Ownership is the one variable you inherit and cannot negotiate.
2. Where does the labour physically sit, and will you name the facilities? Twelve of the 28 vendors we swept publish zero contributor geographies. The moment you carry a data-residency or US-persons obligation, an unanswered version of this question is a disqualification, not a gap.
3. Which certification framework, which Type, what scope, and can I have the report? Four parts, one question. It separates Cogito Tech’s six unevidenced claims from Mercor’s single evidenced one, and it is what catches a genuine certificate scoped to the wrong product, as Handshake’s is.

Answer those and most of the 26 are gone. The rest of this section is the ordered path for what survives.

The routing path, in order.

1. Which purchase? Platform, managed outcome, your workforce on their tooling, or licence a dataset. This eliminates more of the roster than anything else here.
2. Does workforce geography disqualify anyone? Undisclosed geography is a disqualifier the moment you carry a data-residency or US-persons obligation.
3. Does ownership disqualify anyone? Does the vendor’s owner compete with you, and how would you find out if that changed mid-contract?
4. Modality and domain fit. Does the marketing page say “multimodal” while the case studies say bounding boxes?
5. Certification framework, Type and scope. Evidenced, not asserted.
6. Pricing posture and landed cost. Can you model the number, and does the procurement shape fit your cycle?
7. Review evidence, with sample size. Read last, and read sceptically.

The use-case routing table

Use case	Category to start in	2-3 names to look at	The decisive constraint
Medical and clinical	Enterprise, platform-first with clinicians	iMerit, V7, Encord	Annotator qualification and health-data privacy scope
Autonomous vehicles and robotics	Enterprise, platform-first	Scale AI, Encord, Sama	3D and LiDAR tooling maturity
LLM, RLHF and evaluations	Expert-data marketplaces	Surge AI, Mercor, Handshake AI	Credential verification and no certification artefact to lean on
Multilingual speech and language	BPO and crowd-scale	TELUS Digital, Appen, LXT	Language locale count and secure-facility options
Documents and OCR	Platform-first, mid-market managed	Kili Technology, V7, Cogito Tech	Key-value extraction quality and document confidentiality
Content moderation	BPO and crowd-scale	TaskUs, Sama, TELUS Digital	Wellbeing programmes and site-level accountability
Small-volume self-serve	Open source and cloud-native	Roboflow, CVAT.ai, Label Your Data	Published pricing and no minimum commitment

The six disqualifiers

Disqualifying is faster than ranking. Any one of these ends the conversation:

• Workforce geography undisclosed while you carry a data-residency obligation.
• The vendor’s owner competes with you.
• A certification claimed with no artefact and no stated scope.
• A pricing posture incompatible with your procurement cycle.
• No published quality method, and no willingness to describe one in the pilot.
• A product past end of support.

Do not read “#1 in a guide” as market leadership. In at least 12 of roughly 15 ranking guides on this SERP, #1 is the publisher, and the disclosure standard set at the top of this article is the one to hold them to. The same diligence discipline applies upstream, and the enterprise evaluation checklist covers the version of this for data suppliers rather than annotation vendors.

A tie-breaker, because scorecards produce ties. If two vendors land within about 5% of each other on a weighted comparison, do not re-weight until one wins. Ask instead which dimension, if it failed in year two, would cost the most. For most enterprise annotation programmes that is annotator cohort continuity, because a cohort that has learned your edge cases is the asset you are actually buying and cannot be re-procured quickly. Where the programme touches regulated data, ownership and workforce geography displace it, because those two fail in ways that are not recoverable with money.

Frequently asked questions

How much does data labeling cost?

Published 2026 benchmarks run from roughly $0.02 per image classification label to $0.50 to $2.00+ for semantic segmentation, $0.50 to $3.00 per audio minute and $1.00 to $5.00+ for medical items, with hourly managed rates between $6 and $60+. Only six of the 28 vendors we checked on 6 August 2026 publish real figures, so most of this market prices by quote. The correction most buyers need: the lowest per-unit rate is frequently the most expensive vendor, because $0.01 per record at a 20% error rate costs more than $0.04 at 1% once rework, QA cycles and cohort re-training are counted.

What is the difference between a data labeling platform and a data labeling service?

They are opposite risk transfers. A platform sells you software and leaves the labour, throughput and quality risk with you; a service sells you an outcome and takes process visibility away in exchange. A third path that almost no roundup names is renting the vendor’s tooling and staffing it with your own workforce, of which AWS SageMaker Ground Truth’s private-workforce option is the clearest illustration, though AWS closed that service to new customers on 30 July 2026. Most vendors sit on a spectrum rather than in a box, so the useful question is which risk you retain after signing, not which label the vendor uses.

Should we outsource data labeling or do it in-house?

Outsourcing is the norm. Grand View Research put 84.6% of the 2024 data labeling solution and services market in the outsourced segment, which leaves in-house at roughly 15%. In-house wins in four specific situations: the domain expert is already on your payroll, the volume is small enough that vendor minimums dominate, the taxonomy is still moving, or the data cannot legally leave your jurisdiction. Outside those, the question is not whether to outsource but which of five workforce models to outsource to, because they behave differently on price, variance, ramp and confidentiality.

How do I evaluate annotation quality before signing a contract?

Run a two-week pilot with two or three vendors on a batch of roughly 500 units, with the NDA and DPA signed first, and score the output against your own benchmark set rather than the vendor’s gold set. Then write the accuracy floor into the contract as an agreement threshold rather than a percentage: on the Landis-Koch scale, published in Biometrics in 1977, 0.61 to 0.80 is substantial agreement and 0.81 to 1.00 is almost perfect. For subjective tasks, anything under 0.80 is a question rather than a signature. Name the statistic too, because Cohen’s kappa, Fleiss’ kappa and Krippendorff’s alpha answer different questions.

What certifications should a data annotation vendor have, and is my data safe?

Ask three questions rather than one: which framework, which Type, and what is in scope. Only three of the 28 vendors we checked claim ISO 27701 and only four claim PCI DSS, so a checklist that assumes either is standard will disqualify most of the market before you have evaluated anything. A certificate can be entirely genuine while covering a different product: Handshake’s SOC 2 Type II covers its recruiting products rather than its AI data unit. Ask for the report, the scope, and the report date.

The Ownership Question

Eighteen months ago, “who owns this vendor?” was a curiosity you might get to in the third meeting. Then Meta took a non-voting stake in Scale AI, reported at 49% for $14.3 billion, and Google, Microsoft, xAI and OpenAI were all reported reducing or ending their engagements inside a week. TELUS International’s privatization closed on 31 October 2025 and the shares came off the NYSE days later. TaskUs stayed public because its shareholders voted a take-private down on 8 October 2025. clickworker turned out to be inside LXT. Shaip turned out to be inside Ubiquity Global Services. In August 2026, iMerit completed its sale to EXL. Centific turned out to be Pactera EDGE with a new name.

None of those are scandals. They are ordinary corporate events, and every one of them changes the answer to a question your RFP probably does not ask: who owns this vendor, and who else do they serve?

That is the criterion this market produced in 2025 and 2026, and it belongs at the top of an evaluation rather than the bottom. It is checkable in an afternoon for a public company and genuinely hard for a private one, which is itself the finding. Certification you can request. Pricing you can negotiate. Ownership you inherit, and it can change without you.

Put it in the RFP as question one. The guides ranking for this search still do not ask it.

Before you label anything, you need the data

Every company in this guide assumes you already have the corpus. Labeling adds structure to data you possess. It does not produce the data.

If the harder half of your problem is acquiring it, at scale, from the open web and from documents, and keeping it current as sources change, that is what Forage AI does. Forage AI is an end-to-end managed data pipeline: it automates discovery, extraction, structuring, quality checks and delivery, built to your own fields and systems, so you receive the data rather than a tool to maintain. Every delivery passes a 3x QA team, and from sign-off to first dataset takes 1-2 weeks. Clients own the data, and we do not resell it. For a broader view of how this category has developed, modern data extraction services in 2026 covers the landscape.

That is a different problem from annotation, and if annotation is what you need, the 26 entries above are where to look.

See what Forage AI builds for AI teams →

Talk to our expert →

About the author

Forage AI Research Team. We build and run managed data pipelines for web and document data. That work puts us in the same procurement conversations buyers have here, which is why we write these guides.

This edition was built from a same-day sweep of 28 vendor websites, security pages, pricing pages and public review profiles on 6 August 2026. Ownership, funding and pricing were re-checked on 15 October 2026. Forage AI does not sell data labeling or annotation and is not one of the 26 entries.

Related Articles

• Top AI Training Data Providers: Buyer’s Guide to the Four Categories: The four categories of the AI training data market, and which one your problem belongs to.
• Best AI Dataset Marketplaces: 8 Platforms Compared: Licensing a finished dataset instead of building one.
• AI Training Data: The Operational Playbook: Running a training-data programme after the vendor is selected.
• The Best Data as a Service (DaaS) Companies in 2026: The delivered-data end of the same supply chain.
• What Is Data for AI (And How You Can Use It): Labeled versus unlabeled data, and what each one is for.