Google and Accel sifted through more than 4,000 applications for their joint Atoms AI accelerator in India and rejected roughly 70% of them as wrappers, startups that bolted chatbots or LLM features onto existing software without rethinking the underlying workflow. The five companies that made it through are building AI agents for ERP systems, voice automation for call centers, and industrial manufacturing tools. Not a single consumer product.
The headline number is striking, but the composition of the applicant pool is more telling. Accel partner Prayank Swaroop told TechCrunch that about 62% of submissions targeted productivity tools and another 13% focused on software development and coding. Three-quarters of India's AI startup energy, in other words, is flowing into enterprise software for offices and dev shops.
The outsourcing hangover
That ratio should worry anyone hoping India produces the next wave of original AI companies. What it actually suggests is that founders are still thinking inside the frame of IT services and outsourcing, the model that built India's $300 billion tech sector over the past four decades. Build a tool that makes an existing enterprise workflow slightly faster or cheaper. Pitch it to the same Western clients who already buy Indian IT services. Repeat.
The rejected applications weren't reimagining workflows using AI, Swaroop said, which is a polite way of saying most founders treated GPT-4 or Gemini as a feature to tack onto a SaaS dashboard rather than a reason to rebuild from scratch. Many of the non-wrapper rejections fell into crowded categories like marketing automation and AI recruitment, areas where differentiation is nearly impossible.
This tracks with a broader anxiety in Indian tech right now. In February, Indian IT stocks lost $50 billion in market value in a single week after Anthropic released its Claude Cowork tools, which automate exactly the kind of repetitive knowledge work (contract reviews, compliance tracking, sales forecasting) that millions of Indian IT workers perform for Western clients. Nasscom president Rajesh Nambiar said publicly that new engineering graduates should be worried. Investment bank Jefferies projects Indian call-center revenues could drop 50% over five years.
So when thousands of Indian founders pitch AI startups that are essentially wrappers around someone else's model, doing the same work as the outsourcing industry but with an API call instead of a person, it's not hard to see the pattern.
What Google gets out of this
Jonathan Silber, co-founder and director of Google's AI Futures Fund, described a feedback mechanism that's worth paying attention to. The program doesn't require startups to use Google's models exclusively. Silber told TechCrunch that many companies combine multiple models depending on the workflow, and if a startup picks a competitor's LLM for a specific task, that tells Google where its own models fall short.
"If a company is using an alternative model, that means Google has work to do to build the best model in the market," Silber said, which is a remarkably candid admission from someone whose fund is writing checks up to $2 million per startup and handing out $350,000 in cloud credits.
The Atoms program, announced in November, is structured as a co-investment between Accel and the AI Futures Fund, which Google launched in May 2025. Selected startups get early access to Gemini and DeepMind models, plus direct mentorship from Google's research teams. It is, functionally, a market research operation that pays its subjects.
The five that survived
The finalists are K-Dense, building an AI co-scientist for life sciences and chemistry research; Dodge.ai, making autonomous agents for enterprise ERP systems; Persistence Labs, doing voice AI for call centers; Zingroll, a platform for AI-generated films and shows; and Level Plane, applying AI to industrial automation in automotive and aerospace manufacturing.
Zingroll is the only one that could loosely be called consumer-facing. The rest are deep enterprise plays. I find the Dodge.ai and Level Plane picks particularly interesting because they're going after workflows that are genuinely hard to automate with a thin wrapper: ERP systems are notoriously tangled, and manufacturing automation requires dealing with physical processes, sensor data, and safety constraints that a chatbot can't handle.
Swaroop said he'd hoped to see more healthcare and education startups. He didn't get them.
What's missing from this picture
This year's Atoms program received nearly four times the applications of previous cohorts. That surge came with a flood of first-time founders, which partly explains the wrapper problem. But it also points to something the 70% rejection stat doesn't capture: the sheer number of people in India who see "AI startup" as the next viable career path and default to the business model they already know.
India produces about 1.5 million engineering graduates a year. The country has deep programming talent and a mature outsourcing infrastructure. What it doesn't have, at least not yet, is a strong tradition of building original product companies that compete globally on innovation rather than cost. The Atoms results are a small data point, but they rhyme with that larger story.
The accelerator received 4,000 applications and picked five. That's a 0.125% acceptance rate, worse than any Ivy League school. Whether those five can break out of the services-plus-AI pattern is a question that $2 million and some cloud credits probably won't answer on their own.
