Scale AI vs Appen: the incumbent and the one that came before it
Appen has been selling human data since 1996, which is roughly twenty years before the category had a name. Scale AI defined what the category became. Putting them side by side looks like a generational comparison, and in some respects it is, but the more useful thing they have in common is recent: both lost a large customer abruptly, both for reasons that had nothing to do with the quality of their work, and both are rebuilding around that. This comparison covers what each one actually sells today, what each publishes about its own operation, and where the differences change a purchasing decision.
Key takeaways
- Appen's advantage is linguistic and geographic reach: a stated one million-plus contributor network across more than 170 countries and over 500 locales. Nothing else in the category approaches that.
- Scale's advantage is government. FedRAMP High and DoD IL4 authorizations, plus a verifiable federal contract history, decide the comparison for any defense-adjacent program.
- Both have customer-concentration histories worth asking about. Appen's revenue fell 18.4 percent after Google ended its contract; Scale lost work from Google, OpenAI, and xAI after Meta took a 49 percent stake.
- Appen publishes far more about its workforce than Scale does. Scale publishes no contributor headcount, no delivery geography, and no pay rates.
- Appen's ISO 27001 certification is facility-scoped rather than company-wide, which matters if your project runs somewhere it does not cover.
- Neither publishes IP indemnification terms, and neither publishes a named case study with attributable metrics.
About Appen
Founded in 1996 in Sydney by Dr Julie Vonwiller, a linguist, and listed on the Australian Securities Exchange since 2015. It runs a dual headquarters in Sydney and Kirkland, Washington, with roughly 1,000 employees as of 2024 and a contributor network far larger than that.
Appen organizes its current offering into six pillars: frontier model alignment covering chain-of-thought data, subject-matter-expert RLHF and red-teaming; agentic AI covering trajectories and RL environments; speech and audio across more than 500 locales including text-to-speech and multi-speaker transcription; multimodal vision-language alignment; physical AI covering LiDAR annotation, robotics trajectories, and sensor fusion; and model integrity and evaluation covering hallucination and bias benchmarking. It also sells a pre-built dataset catalog spanning more than 80 languages.
The financial history is relevant to a buyer because it shows concentration risk playing out. First half FY2025 operating revenue fell 18.4 percent year over year to 113.4 million US dollars after Google terminated its contract, with global services revenue down 36.5 percent. First half FY2026 revenue rose 17.5 percent year over year to 119.9 million with improved profitability, led by an 80 percent increase in Appen China. Ryan Kolln has been chief executive since February 2024.
Customers appear as a logo wall rather than case studies, including Oracle, Amazon, Microsoft, Nvidia, Adobe, Bloomberg, Airbus, and Boeing. Appen self-claims that 80 percent of the world's leading large language model builders are customers, without stating a methodology.
About Scale AI
Founded around 2016 by Alexandr Wang and Lucy Guo, based in San Francisco. In June 2025 Meta invested roughly 14.3 billion US dollars for a 49 percent stake at a valuation above 29 billion. Wang left to lead Meta's superintelligence effort while remaining on Scale's board, and Francis deSouza became chief executive in August 2026.
Scale sells a Data Engine for annotation and curation across text, image, video, and 3D and LiDAR; a generative AI portfolio covering RLHF, prompt-response generation, and red-teaming; public model leaderboards; and Donovan, a decision-support platform sold into defense and enterprise.
Its government business is the most verifiable part of its record, because federal contracting data is public. Army contract W911QX20C0051 shows 105.5 million dollars total obligated across modifications. An Army indefinite-delivery contract signed in May 2026 carries a ceiling near 99.5 million. A February 2026 Air Force award of roughly 5.1 million integrates Donovan into an airborne command platform. The frequently cited 500 million dollar Department of Defense agreement is a ceiling rather than obligated spend, a distinction that tends to disappear in secondary coverage.
Named customers with case-study-level description are Mayo Clinic and British Petroleum. Scale claims 90 percent of leading generative AI model builders are powered by Scale, also without stated methodology.
What buyers report
Both review bases need reading carefully, for different reasons.
Appen holds 4.2 out of 5 on G2 across 35 reviews. Reviewers cite labeling accuracy across image, video, and text, quick project deployment, and reliable payment. Criticisms include inconsistent project availability, periodic server outages at peak load, a slow qualification process, and slow support response. The important caveat is composition: G2's review base for Appen mixes crowdworker feedback with enterprise-buyer feedback, so the buyer signal is diluted by people reviewing it as a place to work rather than a vendor to hire.
Scale's GenAI Platform holds 4.5 out of 5 on Gartner Peer Insights from two reviews, citing model quality and the combination of models, infrastructure, and tooling, with pricing structure flagged as a purchasing concern. A sample of two is an anecdote.
Neither gives you what you would want. Plan on a paid pilot.
The comparison
| Appen | Scale AI | |
| Founded | 1996 | Around 2016 |
| Ownership | Public, ASX listed since 2015 | Private; Meta holds 49 percent since June 2025 |
| Core strength | Language and locale breadth at volume | Breadth across modalities, plus government |
| Workforce | Crowd, 1 million-plus contributors stated | Not published |
| Geographies | 170-plus countries, 500-plus locales, named secure facilities | Not published |
| Pay rates | Not published on the corporate site | Not published |
| Compliance | SOC 2 Type II, ISO 27001 facility-scoped, HIPAA at UK facilities only | SOC 2 Type II, ISO 27001, DoD IL4, FedRAMP High. No HIPAA found |
| Government | None published | Verifiable federal contract history |
| Open data | None found | 36 datasets and 5 models on Hugging Face |
| Reviews | 4.2 out of 5, 35 reviews, mixed audience | 4.5 out of 5, two reviews |
| IP indemnification | Not published | Not published |
Where Appen is genuinely ahead
Locale coverage is the one dimension where Appen is not merely competitive but categorically different. More than 500 locales, more than 170 countries, more than 100 languages, with a dataset catalog spanning over 80 languages. If your problem is speech recognition across Indic languages, text-to-speech in Southeast Asian markets, or evaluation in languages that do not appear in a standard vendor's list, this is not a close call.
The second advantage is disclosure. Appen publishes its network size, its geographic reach, and the locations of its named secure facilities. Scale publishes none of these. A vendor that can tell you where its contributors are and under what conditions they work is a vendor that can usually tell you why a particular label came out the way it did, which is what you need when quality problems have to be traced rather than absorbed.
Where Scale is genuinely ahead
Government, decisively. Scale holds a DoD IL4 Provisional Authorization and FedRAMP High authorization, usable across NIPR, SIPR, and JWICS networks, backed by a federal contract record that anyone can verify in public procurement data. Appen publishes no equivalent authorizations. For a defense or federal program this is not a comparison, it is a single qualifying vendor.
Scale is also substantially more open with research artifacts, publishing 36 datasets and 5 models on Hugging Face, the largest open footprint of any services vendor in this category, alongside public model leaderboards. Appen's Hugging Face organization shows none. Open releases let you inspect task design and labeling quality yourself before committing, which is more informative than any capability claim.
Concentration risk, which both have demonstrated
This is the question to raise with either vendor, because both have lived it.
Appen's revenue fell 18.4 percent in the first half of FY2025 after Google ended its contract, with global services down more than a third. That is what single-customer dependence looks like on a public income statement. The recovery is real, with first half FY2026 up 17.5 percent, but it was led by an 80 percent increase in Appen China, which is a concentration of a different kind and one worth asking about directly if your data has residency or jurisdictional constraints. Note also that Appen's ISO 27001 certification is facility-scoped, covering named sites in the UK, the Philippines, and China, rather than the company as a whole.
Scale's version arrived through its cap table. After Meta's June 2025 investment, Google, OpenAI, and xAI were reported to have pulled back or halted projects, Scale laid off roughly 200 data-labeling staff, and by August 2025 reporting indicated Meta's own AI unit was favoring other vendors. Unlike a lost contract, this does not resolve with time, because the cause is who owns the counterparty.
The practical question for either: what share of revenue does your largest customer represent, and what happens to my delivery team if that customer leaves.
Which to choose
Choose Appen when the work is multilingual at scale, particularly speech and audio across locales that other vendors do not reach; when you want a vendor that publishes what its workforce looks like; or when you want a publicly listed counterparty whose financials you can read before signing. Confirm which facilities your project will run in, because the ISO 27001 scope follows the site rather than the company.
Choose Scale when you are deploying into defense or government, where the authorizations settle it; when you need 3D and LiDAR annotation at throughput; or when you want published certifications and open benchmarks you can evaluate before a sales conversation. Weigh the Meta stake according to whether Meta competes with you.
Why HumanSignal is the better choice when the data has to be created
Both companies here process material that already exists. Appen does it with enormous linguistic reach; Scale does it with enormous breadth and the strongest government posture in the category. Neither is organized around the case where the examples your model needs have never been recorded by anyone, which is now the common case in robotics, embodied AI, and any domain where the expertise lives in people rather than on the open web.
HumanSignal Services is built for exactly that. We design 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 rather than with a staffing layer or a task queue.
The quality mechanism is inspectable rather than asserted. 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 happening 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, and the HIPAA coverage is not scoped to a single facility.
The capability came from acquiring Erud AI, a physical data collection company, in November 2025. If your constraint is capture rather than throughput, the physical AI data services page sets out how a program gets scoped, and we will work through the protocol and the acceptance criteria with you before anything is committed.
Is Appen still a viable vendor after losing the Google contract
On the published numbers, yes. First half FY2026 revenue rose 17.5 percent year over year to 119.9 million US dollars with improved profitability. As a listed company its financials are public, which is more than can be said for most vendors in this category, so you can assess this yourself rather than take a salesperson's word for it. Ask about current customer concentration.
Which one is better for multilingual speech data
Appen, clearly. More than 500 locales and more than 100 languages, with a dedicated speech and audio pillar covering text-to-speech and multi-speaker transcription. Scale does not publish comparable locale coverage and does not position around language breadth.
Does Appen's ISO 27001 certification cover my project
Not necessarily, and this is worth checking explicitly. The certification is scoped to named facilities in the UK, the Philippines, and China rather than to the company as a whole, and its HIPAA claim covers UK secure facilities only. Ask which site your work will run in and whether that site is in scope.
Can either vendor collect physical or robotics data
Both annotate it. Appen's physical AI pillar covers LiDAR annotation, robotics trajectories, and sensor fusion, and Scale annotates 3D and LiDAR within its Data Engine. Neither publishes a designed collection capability with its own capture facilities, so if you need the data produced rather than labeled, look at collection specialists instead.