Annotera: Why Frontier AI Models Depend on Enterprise-Grade Annotation

In June 2025, Meta paid $14.3 billion for a 49% stake in Scale AI, a company whose main business is labeling data. It was one of the biggest moves of the AI race, and it did not buy a model or a chip. It bought a position in the human judgment that turns raw data into training examples. 

That single deal says more about where model quality now comes from than any benchmark score. The scarce input in frontier AI is no longer raw data, and it is not only compute. It is well-labeled data, produced by people who understand what they are looking at. 

The bottleneck moved from volume to judgment 

For a decade, data annotation was volume work. Draw boxes around cars, tag emails as spam, transcribe an hour of audio. The answer was usually obvious, and the only real question was speed and cost per label. 

Once a base model has read most of the public internet, more raw labeled data stops helping much. The scarce input becomes human judgment on the questions the model still gets wrong. A category of provider grew up to supply it. Annotera is an AI data annotation company built around this newer work rather than around bulk tagging. 

The market is following the money 

The spending backs up the shift. The data annotation tools market was worth about $1 billion in 2023 and is on track to reach $5.3 billion by 2030, a compound annual growth rate of 26.3%. Count the wider data collection and labeling market and the 2030 figure climbs past $17 billion. 

Those numbers describe an industry that used to be treated as a cost center and is now treated as central to model performance. That reframing is the story. 

Reinforcement learning made the shift concrete 

Reinforcement learning from human feedback, the process used to align most current chat models, shows what changed. A person reads two model responses and decides which is better. Or rewrites a weak answer into a strong one. Or flags a fluent reply that happens to be false. 

There is no box to draw and no single correct label in a spreadsheet. The annotator has to understand the domain well enough to judge what “better” means. A clinician checking whether a model’s summary of a scan is safe is not interchangeable with an anonymous worker paid by the task. 

A data pioneer has been saying this for years 

Andrew Ng, who founded Google Brain and Landing AI, has argued since 2021 that teams overspend on models and underspend on data. He calls the alternative data-centric AI. 

Speaking at an MIT event, Ng defined it as “the discipline of systematically engineering the data needed to build a successful AI system.” 

He has also argued in Fortune that clean, consistent labels on a smaller dataset often beat a larger, noisier one, backed by experiments where improving the data outperformed improving the model. Same model, better labels, better results. It is an unglamorous idea, which may be why it took a pioneer to keep repeating it. 

The $14 billion signal, and the backlash 

Scale AI is worth studying because its clients were the frontier labs themselves. Its annotation work has fed models at OpenAI, Google, Anthropic, Meta, and Microsoft. When Meta took its stake, that shared dependence turned into a problem. 

Within days, rivals headed for the exits. OpenAI wound down its work with Scale, and Google, which had planned to spend around $200 million with the company, moved to cut ties. Nothing focuses a lab’s mind on vendor neutrality quite like watching its data supplier get half-bought by a direct competitor. 

The winners were independent providers without a competitor’s name on the cap table. “Our demand has tripled overnight after the news,” said Garrett Lord, chief executive of Handshake, in an interview with CNBC. Firms built as neutral, enterprise-grade annotation providers, Annotera among them, sit in exactly the position Scale could no longer offer. The episode proved, in dollars, that frontier labs do not just need labeled data; they need a labeling partner they can trust with their roadmap. 

OpenAI’s finance chief, Sarah Friar, made the underlying point. As models get smarter, she told Reuters, teams “need real expertise” on the data side, and most work with several vendors to get it. 

How Annotera supports enterprise teams and frontier AI labs 

Annotera sits in the independent part of this market, the neutral position the Scale episode made valuable. A client’s roadmap is not exposed to a competitor. 

The operation is large. Annotera runs more than 1,500 dedicated specialists, not a crowd, across delivery locations on six continents. Its multilingual work covers more than 28 languages, which matters when a model must sound native in each market rather than translated. 

The first differentiator is quality control. Every project runs through a three-layer review, from the annotator to a team lead to an independent check, which the company reports at over 99% first-pass accuracy. Consistency like that is what holds up on judgment-heavy work. 

The second is range. The services reach past labeling into LLM fine-tuning and preference data, along with robotics formats such as egocentric video annotation, the first-person, head-mounted footage that robots and wearable assistants learn from. A single clip carries intent, motion, and sequence at once, so it rewards a trained reviewer over a fast one. 

For a frontier lab, the last piece is trust. Data moves through access-controlled, ISO 27001-aligned workflows, with HIPAA-aware and GDPR handling for regulated clients. That is the test the frontier labs applied when they walked away from a vendor, they no longer trusted. 

For an enterprise team deciding where its training data comes from, provable quality and neutrality are the combination that matters. 

Enterprise-grade is a process, not a headcount 

“Enterprise-grade” is easy to claim and harder to run. In practice it means written guidelines that define a good answer, calibration so different annotators reach the same call, and a way to measure how often they agree. It means catching quality drift before it reaches the training set, not after. 

For a frontier lab, it also means handling data the lab cannot hand to an open crowd. Model prompts, user data, and unreleased capabilities carry confidentiality and compliance weight. A provider at this tier needs vetted staff, access controls, and an audit trail, which is a different operation from a marketplace that routes tasks to whoever is online. 

The frontier is physical 

The newest work is harder still, because it leaves text behind. Self-driving systems need sensor data labeled frame by frame, including the weather and lighting the model rarely meets on a clean test route. Robots learning to handle objects need recordings of people doing the task, often shot from a head-mounted camera and marked up for intent, grip, and sequence. 

The stakes are highest in autonomous driving, the single largest vertical in the AI annotation market in 2024 at roughly a third of revenue. Assistants meant to work across dozens of markets add another layer, because they need native speakers rather than translated English, and tone and idiom rarely survive machine translation intact. None of this is bulk work. 

Quality decides the ceiling, quietly 

When labels involve judgment, two careful people will sometimes disagree, and how a team handles that disagreement sets the ceiling on the model. Teams that treat annotation as a checkbox find out late, after a model has learned the noise in its data. A vision system trained on loose boxes places its own boxes loosely. A reward model trained on rushed preference data learns to prefer answers that look confident over answers that are correct. 

The failure is quiet. It is set before anyone runs the evaluation that would reveal it, which is why serious teams watch inter-annotator agreement the way they watch a production metric. 

Can models label their own data? 

The obvious move is to let a strong model label the data for the next one. For easy cases, that works, and it cuts the human load on routine tasks. Large models can pre-label, cluster similar examples, and flag likely edge cases for review. 

But training a model mostly on another model’s output has limits. Feed a system too much of its own synthetic data and small errors compound instead of averaging out. A model also cannot reliably flag its own blind spots; if it could see them, they would not be blind. Human review remains the step that catches the case the model was never going to catch on its own. 

What this means for teams building models 

The practical shift is easy to state and easy to underrate. Annotation is not a line item to push to the cheapest vendor and forget. It sits alongside architecture and evaluation as a place where a large share of a model’s behavior is decided, which is why Meta was willing to spend $14 billion to secure a seat in it. 

The caveat is that none of this is settled. Automated labeling keeps improving, synthetic data will cover more routine cases each year, and the market is full of vendors who promise judgment and deliver crowdwork. The teams that pull ahead will be the ones who can tell the difference, and who treat the people labeling their data as part of the model rather than a cost to grind down.