Deccan AI just closed a $25M Series A on March 27, 2026 — four days ago — betting that the AI industry’s real problem isn’t intelligence, it’s accuracy. And the fix isn’t better models. It’s throwing cheap human expertise at bad training data.
The pitch sounds boring until you see the numbers.
AI annotation — the invisible labor of labeling, evaluating, and refining training data — is a $2.51 billion market in 2026. Most people assume this work is automated or niche. It’s neither. By 2033, it’s projected to hit $13.11 billion at a 31% compound annual growth rate. That’s not a side hustle. That’s infrastructure.
The $2.51 billion industry hiding behind every AI model
Deccan AI’s funding round — led by A91 Partners, with Susquehanna and Prosus Ventures participating — validates something the industry doesn’t advertise: enterprise AI pilots are moving to production, and the accuracy gap is killing them. Google DeepMind is a client. So is Snowflake. The majority of the Magnificent 7 reportedly pay for this work.
Why? Because RLHF — reinforcement learning from human feedback — doesn’t run on vibes. It runs on domain experts labeling edge cases, evaluating outputs, and catching the failures that break models in production. The shift from pilots to production AI has created a market most people don’t know exists.
Deccan’s revenue grew 10x over the past year, according to Moneycontrol. That’s not hype. That’s enterprises paying real money because their models ship broken without this layer.
India’s 1M expert pool is the feature — and the risk
Deccan’s model hinges on domain expertise from IIT and NIT alumni, PhDs, and students. The company claims access to 1 million+ experts globally, with 5,000 to 10,000 active monthly. It employs 125 people, mostly in Hyderabad.
Do the math. That’s a 200:1 contributor-to-employee ratio.
Founder and CEO Rukesh Reddy positions this as “super accuracy” — the antidote to the super intelligence hype cycle. But the execution raises questions. Deccan vets contributors using a “human + AI” process, reportedly covering 500,000+ specialists. The other half million? Unclear. And when you’re QA-ing evaluations for Fortune 500 clients at that scale, quality drift isn’t a risk. It’s a certainty.
The pitch works because India’s workforce is both massive and cheap. But cheap at scale has hidden costs: 100+ in-house QA staff, messy enterprise workflows, and integration pains turning evaluations into operational deployments. Nobody’s publishing error rates. Nobody’s tracking contributor fraud. Nobody’s documenting vetting failures.
The data we don’t have tells the real story
Here’s what’s missing: pricing. Deccan doesn’t publish enterprise-grade annotation costs per data point or per hour. Neither do Scale AI, Labelbox, or Appen — at least not in any 2026 source I could find. That opacity is strategic. If clients knew the real cost of “super accuracy,” they’d question the ROI.
Also missing: failure data. 88% of AI agent projects fail before production across 2024-2025, with data quality issues accounting for 27% of those failures. But nobody isolates annotation errors. Nobody names the companies that shipped broken models because of bad training data. Nobody admits the human-in-the-loop failed.
That’s not because everything’s fine. It’s because the industry doesn’t measure it publicly.
Deccan’s $25M validates that human annotation is essential, not temporary. But the entire model depends on quality control nobody can verify at scale. If super intelligence requires super accuracy, and super accuracy requires cheap human labor, what happens when the labor pool can’t keep up with the model size?
