AI Data Annotation in India: What Fusion CX’s 400 Expert Hides Really Signal

Fusion CX, a Kolkata-headquartered customer experience provider preparing a ₹1,000 crore public offering, says its AI data annotation arm Annotera is hiring more than 400 domain experts across Kolkata, Howrah and Bengaluru.

The press release presents this as a jobs story riding a market wave. It is better read as a positioning story. A 22-year-old CX and BPO operator is telling investors, weeks before it intends to list, that it is no longer only a call-centre company — that it now builds the human layer underneath AI models, from judging chatbot answers to recording the physical world for robots.

Both things are true at once. The hiring is real, the work has genuinely changed, and the market context is real. But nearly every number that makes the story sound big is either undated, best-case, or belongs to a future that competing providers are already fighting over.

Why This Matters Now

Fusion CX received SEBI approval for its IPO on December 11, 2025 — a ₹1,000 crore offering comprising a fresh issue of ₹600 crore and a ₹400 crore offer for sale by promoters. In August 2026, co-founder and CEO Pankaj Dhanuka told BusinessLine the company hoped to launch the issue in September or October, subject to market conditions. SEBI approvals expire after twelve months, so the window matters.

The company enters it with momentum. According to figures from an addendum to its draft prospectus, Fusion CX reported revenue of ₹1,818 crore in FY26, up 36.8 percent year on year, with net profit up 128.3 percent to ₹170 crore.

Against that backdrop, an announcement about annotation hiring is not routine recruitment. It is part of a broader repositioning visible across the company’s public footprint: a ₹100 crore, 11-storey campus coming up at Bengal Silicon Valley in New Town, an E-Commerce Centre of Excellence in Siliguri, and an Annotera Centre of Excellence in the same Kolkata campus. The company’s own materials describe West Bengal as its largest offshore delivery cluster.

What the Hiring Actually Looks Like

The press release says the new hires will be domain experts — people who know “medicine, law, finance or engineering” well enough to catch a wrong answer that sounds right.

The company’s careers portal, which TechRecast reviewed, tells a more textured story. The live openings include domain experts in politics, art, geography, history, music and sports for large language model evaluation; a clinical data annotator role open to freshers and medical students; web research specialists; annotation subject-matter experts and QA leads for LiDAR and 6DoF work; and a Python and Docker engineer on a pay-per-task contractor model.

Most of the evaluation and research roles are contractual, full-time work-from-home positions. The clinical annotator role is listed as full-time but temporary. The portal currently shows the AI roles concentrated in Kolkata, with Howrah and Bengaluru named in the announcement as additional locations.

None of this makes the announcement misleading. It does mean the “400 domain experts” span everything from advanced-degree evaluators writing adversarial prompts to entry-level annotators on temporary contracts. The distance between those two jobs is one of the most under-reported facts in AI labour.

How the Work Changed

The shift the press release describes is real, and it matters technically.

First-generation annotation was mostly perception work: drawing bounding boxes around cars, tagging products, transcribing speech. Generative AI changed the job description. Training and evaluating large language models now requires people who can judge whether an answer is factually right, logically consistent, complete and safe — the reinforcement-learning-from-human-feedback and evaluation work that the careers listings describe in detail, including identifying hallucinations and building adversarial test cases.

Physical AI added a second frontier. Robots cannot learn to fold cloth or grasp a tool from text; they need first-person video of humans doing those tasks — “egocentric” data — and teleoperation, in which a person remotely guides a robot while its movements and sensor readings are recorded. The Observer Research Foundation documented this emerging data economy across Indian workplaces and homes in July 2026, and Stellaris Venture Partners estimates leading robotics labs need between 100 million and 1 billion hours of egocentric data over the next two to three years.

Annotera sits inside an AI portfolio the company calls Omind AI, alongside Roborax, the brand it uses for robot training data — teleoperation and demonstration capture for humanoid, surgical, industrial and warehouse robotics. The company says its first AI data deal, roughly four years ago, was annotation for autonomous checkout-free stores, and that clients have since moved from asking it to label data to asking it to generate data.

That arc — from labelling what exists to producing what does not — is the genuine technology story inside a routine hiring release.

Where It Fits in the Market

Fusion CX is entering a market that is consolidating around it.

EXL, through its Clairvoyant AI subsidiary, is acquiring iMerit — the Kolkata- and Bengaluru-anchored expert-in-the-loop leader with more than 6,000 specialists and clients that include three of the top seven generative AI companies — for up to $310 million, with the deal expected to close in Q3 2026. Turing nearly tripled its revenue run-rate to about $300 million in 2024. Innodata, listed on Nasdaq, employs more than 11,500 in-house subject-matter experts. Cogito Tech runs roughly 2,500 full-time staff in India with a consulting network of 7,000–8,000.

The demand side shifted too. Scale AI’s entanglement with Meta — a $14.3 billion investment for a 49 percent stake in 2025 — pushed major AI labs toward neutral providers, benefiting specialists like iMerit.

Annotera’s differentiation is essentially distribution: it inherits a BPO’s workforce machinery, delivery infrastructure and enterprise relationships, and points them at AI training data. The company says it already runs 1,500-plus dedicated annotation specialists — the 400 hires would expand that by more than a quarter — across a claimed 12 delivery locations, and claims a 99.2 percent first-pass accuracy benchmark. These figures are company-reported; no client is named anywhere in the announcement, the website or the prospectus-based reporting, and TechRecast could not independently verify any of the named project types, from bronchoscopy video to forest-fire imagery.

The Numbers Behind the Narrative

The press release leans on two NASSCOM figures. The first holds up: the Strategic Review 2026 puts direct tech-sector employment at roughly 6 million in FY26, a net addition of 135,000 people. The second deserves an asterisk.

The claim that India’s annotation market “could exceed USD 7 billion by 2030, with a potential workforce of up to 10 lakh people” comes from a NASSCOM report published in February 2021 — built on a market then worth about $250 million, and presented in the report itself as a best-case outlook. It is a five-year-old projection, not a current forecast, and the release does not date it.

The present is smaller. TeamLease data reported by the Economic Times put India’s full-time managed-services annotators at around 20,000, plus roughly 50,000 freelancers on international platforms. India may well service billions of dollars of annotation work by 2030 — but the distance between 70,000 workers and a promised million is the difference between an industry and an aspiration.

There is also a pricing warning label on the newest work. Inc42 reported that Indian data-collection rates fell about 30 percent between January and June 2026 as more suppliers entered, with annotated video selling to robotics labs at $15–50 per hour while collectors earn ₹250–400 per hour — and synthetic data generation improving fast enough to erode the recurring need for human footage.

The Question the Announcement Doesn’t Answer

What happens to the experts when the training is done?

Annotation for evaluation is durable as long as models keep being updated. But data work for physical AI is, as researchers keep noting, self-terminating: once a model performs a task adequately, the demand for footage of that task ends. A workforce hired to teach robots to fold clothes is building the thing that ends the assignment. The Observer Research Foundation and others have also flagged that this economy is running ahead of India’s data-protection rules — consent, purpose limitation and worker protections for body-worn, always-on capture remain unsettled under the DPDP Act.

For a company heading to market, the honest version of the pitch is that AI data is a fast-growing, competitive, partially commoditising services business in which India’s cost and talent advantages are real but not a moat — and Fusion CX’s CX relationships are the distribution edge. The press release’s framing — a massive, under-populated job market waiting to be staffed — is the version investors should read more carefully.

What This Means for Buyers and Builders

For AI teams evaluating providers like Annotera, the useful diligence questions are now clear: who exactly staffs a project (full-time, contracted or gig), at what qualifications, under what quality framework, with what neutrality commitments relative to competing labs, and what happens to data rights and provenance when the work spans Indian collection teams and foreign model owners.

Editor’s Note