RESEARCHYou Can't Buy Taste. We Tested It.

Appen alternatives: where to go when the crowd model stops fitting

Appen has been selling human data since 1996 and still has the broadest linguistic reach in the category, with a stated network of more than a million contributors across over 170 countries and more than 500 locales. Those numbers are the reason to use it. They are also, indirectly, the reason buyers look elsewhere: a distributed crowd managed statistically is the right instrument for well-specified work at volume and the wrong one when a single wrong label is expensive or the task requires credentials. This guide sorts the alternatives by which of those you are running into.

Key takeaways

  • The crowd model's failure mode is quiet. Agreement metrics stay healthy while a task is systematically misunderstood, and the problem surfaces in model evaluation rather than in the labeling pipeline.
  • For expert judgment instead of generalist throughput, the options are Surge AI, Mercor, Handshake AI, and iMerit.
  • For employed rather than contracted workforces, look at iMerit, Sama, and Innodata.
  • Appen's ISO 27001 is facility-scoped to named sites rather than company-wide, which matters if your project runs elsewhere.
  • Nobody here matches Appen's 500-plus locales. If language coverage is the requirement, the shortlist is short.
  • If the data does not exist yet, this is the wrong category entirely.

Why buyers look for alternatives

The task needs credentials. Crowd quality is managed through redundancy, gold-standard tasks, and agreement thresholds. That works when the judgment required is general. It does not work when the label requires a physician, an attorney, or a structural engineer, because agreement among people who are all wrong in the same way looks exactly like agreement.

Consistency across a long program. A flexible pool turns over. For a program running quarters rather than weeks, where the same edge cases recur and the taxonomy evolves, an employed workforce tends to outperform because the same people accumulate context.

Compliance scope. Appen holds SOC 2 Type II and ISO 27001, but the ISO certification is scoped to named facilities in the UK, the Philippines, and China rather than the company as a whole, and its HIPAA claim covers UK secure facilities only. If your work runs somewhere else, those certifications may not cover it.

Delivery evidence. Appen presents a logo wall including Oracle, Amazon, Microsoft, Nvidia, Adobe, Bloomberg, Airbus, and Boeing, with a self-reported claim that 80 percent of the world's leading large language model builders are customers, stated without methodology. Its case studies page returned no retrievable content when checked. Its G2 rating, 4.2 from 35 reviews, mixes crowdworker feedback with buyer feedback, which dilutes the buyer signal.

Concentration risk, demonstrated. First half FY2025 revenue fell 18.4 percent year over year after Google terminated its contract, with global services down 36.5 percent. The recovery is real, with first half FY2026 up 17.5 percent to 119.9 million US dollars, but it was led by an 80 percent increase in Appen China, which is a concentration of a different kind and worth raising if your data has residency constraints.

When to stay with Appen

Language and locale coverage. More than 500 locales, more than 170 countries, more than 100 languages, and a pre-built dataset catalog spanning over 80 languages, with a dedicated speech and audio practice covering text-to-speech and multi-speaker transcription. On this dimension Appen is not merely competitive, it is categorically ahead, and no alternative below comes close.

Volume at low unit cost. For well-specified, high-volume work where per-unit cost dominates, the crowd model is the economically correct instrument.

Workforce disclosure. Appen publishes its network size, geographic reach, and named secure facility locations. Several larger competitors publish none of that.

Financial transparency. Appen is listed on the Australian Securities Exchange, so you can read its financials before signing. Almost nothing else in this category offers that.

Alternatives by what you need

Expert judgment instead of generalist throughput

Surge AI is the category benchmark for expert text, RLHF, and evaluation design. It has taken no outside capital, reported 1.2 billion dollars in 2024 revenue, and published 13 benchmarks with 8 arXiv papers in 2026, with OpenAI and Anthropic citing its benchmarks in dated model releases. It publishes substantially less about compliance than Appen, so expect your own assessment.

Mercor matches credentialed experts through an AI interview process and bills their hours at a markup, stating 30,000-plus active experts and 107 supported payout countries. Its published trust documentation is thinner than Appen's: SOC 2 Type 2 and nothing else.

Handshake AI recruits from a career network of 20 million students and professionals and publishes a substantial body of open benchmark research. Its fellowship requires United States work authorization, so it cannot serve non-US-language work, which is often exactly what buyers are leaving Appen to improve.

Employed workforces and long-program consistency

iMerit employs its workforce full time with benefits rather than contracting it, reporting 10,000-plus active resources across 60-plus countries and a 25,000-plus network of domain experts. It publishes the broadest compliance set in this category: SOC 2 Type 2, ISO 27001, ISO 9001, HIPAA, GDPR, and TISAX, plus dated AI ethics and environmental policies. EXL Service completed its acquisition in August 2026, so ask about post-integration continuity.

Sama runs an employee-based delivery centre model and names the locations: Nairobi, Gulu, Kampala, San Jose in Costa Rica, Montreal, and The Hague. It publishes seven named case studies with quantified outcomes, which is more delivery evidence than almost anyone here. It holds ISO 27001 and ISO 9001 and claims TISAX. SOC 2 appears on its trust page without a date or auditor, unlike its ISO certifications, so ask for the report.

Innodata has run a directly employed global workforce since 1988, reporting more than 10,000 employees and 2025 revenue of 251.7 million dollars, up 48 percent year over year.

Breadth in one vendor, and government

Scale AI covers text, image, video, and 3D and LiDAR under one contract and holds DoD IL4 and FedRAMP High authorizations that nothing else here matches. The trade is Meta's 49 percent stake and a workforce it does not describe publicly.

Multimodal work beyond text and speech

Invisible Technologies covers video annotation, audio engineering, and 3D modeling across more than 80 languages, and publishes the most substantial trust documentation in this category: a SOC 2 Type 2 updated February 2026, HIPAA policies, Cyber Essentials Plus, six named penetration-test reports, and sixteen dated security policies, alongside named case studies with metrics.

Data that does not exist yet

If what your model needs has never been recorded, no annotation vendor solves it. The collection specialists are Build AI, XDOF, MicroAGI, and HumanSignal, covered below.

Comparing the alternatives

VendorWorkforceStrongest forCompliance published
Surge AIVetted experts, contractedRLHF, evaluation design, benchmarksSubstantially less than Appen
MercorMarketplace, AI-screenedSpeed to staff across 107 countriesSOC 2 Type 2 only
Handshake AICredential network, US onlyUS credentialed supply, open researchScoped to its education platform
iMeritFull-time employeesClinical, geospatial, automotive, consistencySOC 2 Type 2, ISO 27001, ISO 9001, HIPAA, GDPR, TISAX
SamaEmployees in named centresComputer vision, autonomous driving, documented deliveryISO 27001, ISO 9001, ISO 42001, TISAX, GDPR, SOC 2 displayed
InnodataDirectly employed, since 1988Long regulated programsNot detailed in this comparison
Scale AINot publishedBreadth, government, open benchmarksSOC 2 Type II, ISO 27001, FedRAMP High, DoD IL4
InvisibleVetted experts, contractedMultimodal breadth, documented securitySOC 2 Type 2, HIPAA policies, Cyber Essentials Plus, CMMC

None of these publishes IP indemnification covering the provenance of a delivered dataset. Get it drafted explicitly rather than expecting to find it published.

How to run the comparison

Start by deciding whether your problem is the workforce or the task specification. A surprising share of crowd quality problems are specification problems, and moving to an expert vendor without fixing the spec buys an expensive version of the same result. If your agreement metrics look fine but your model does not improve, the task definition is the first suspect.

Then ask what the compliance scope covers, because this category is full of certifications that cover a facility, a parent company, or an adjacent product line rather than the entity doing your work. Appen's ISO 27001 follows named sites. Handshake's attestations describe its education platform. Ask which entity and which location your engagement runs under.

Finally, ask for named outcomes with numbers. Sama and Invisible publish them. Most of this category does not, and a logo wall tells you nothing about whether anyone was satisfied.

Why HumanSignal is the better choice when the data has to be created

Every vendor above labels, evaluates, or judges material that already exists, and they differ meaningfully in who does that work and how well they document it. None is built for the case where the examples your model needs have never been recorded, which is the binding constraint in robotics, embodied AI, biometric capture, and any domain where the expertise lives in physical practice rather than in an existing corpus.

HumanSignal Services is built for exactly that. We write the collection protocol, recruit for the physical or professional characteristics it requires, run the capture in our studios, on site at your facility, or in the field with synchronized motion, vision, audio, and sensor streams, and deliver a dataset verified against the spec we agreed. The accountability for the dataset sits with us, not with a task queue or a sourcing layer.

On the diligence questions above, we answer each of them. Review workflows run in Label Studio Enterprise, the commercial edition of the open source labeling tool we build, so inter-annotator agreement, reviewer decisions, and rework rates are reportable to you while the work is running rather than summarized once it is over. We publish contributor pay ranges, which makes our sourcing claims checkable. Our customer agreement gives you exclusive ownership of customer data and of insights produced specifically for you. We hold SOC 2 Type 2 and HIPAA compliance, documented on our security page, scoped to the company rather than to a named facility.

The capture operation came in through the acquisition of Erud AI in November 2025. If you are leaving Appen because the crowd cannot produce what you need rather than because it cannot label it fast enough, physical AI data services is where to start.

Is any vendor a direct replacement for Appen's language coverage

Not from this list. More than 500 locales with a dedicated speech and audio practice is Appen's genuine moat, and no alternative compared here publishes equivalent breadth. Teams that leave Appen for expert quality often keep it for multilingual work and run two vendors.

How do I know whether I need experts or better task definition

If inter-annotator agreement is high but model performance is not improving, the task specification is the more likely problem, because consistent annotators producing consistently wrong labels will show healthy agreement. If agreement is low on cases a specialist would find obvious, that is a credentials problem. Diagnose before you switch vendors.

Does Appen's ISO 27001 cover my project

Not necessarily. The certification is scoped to named facilities in the UK, the Philippines, and China rather than to the company as a whole, and the HIPAA claim covers UK secure facilities only. Ask which site your work will run in and whether that site is in scope.

Which alternative is strongest on published delivery evidence

Sama, with seven named case studies carrying quantified outcomes, and Invisible Technologies, with named customers including Cohere and Nasdaq and specific metrics attached. Both are vendor-published and unaudited, but they are named and attributable, which most of this category cannot offer.

Related Content