India’s AI startups are building homegrown alternatives to global products, yet state governments have been eyeing partnerships with foreign players, making it harder for them to find customers at home.
ElevenLabs recently signed a memorandum of understanding (MoU) with the Karnataka Innovation and Technology Society (KITS) to explore using voice AI in Kannada across citizen services, skilling, and grievance redressal. The initiative will enable mute users to communicate using a synthetic version of their lost voice, including in Kannada. Under the programme, they will receive free lifetime licences.
This partnership gives ElevenLabs a formal route to pursue public-sector applications in a state with a large technology ecosystem.
That said, India is ElevenLabs second-largest market outside the US, with more than 250 businesses and enterprises using its technology. Indian businesses recorded 100 million conversations on ElevenAgents over the past year, with more than 70% conducted in Hindi, Kannada, Tamil, and Telugu. The platform’s language support is set to grow from 14 Indian languages to all 22 official languages.
But the Karnataka government’s decision to partner with a US-based startup for government services hasn’t resonated well with Indian startups. Earlier this year, it also formed joint working groups with Anthropic to explore AI skilling across the state through Claude certifications.
“When government services rely on foreign AI platforms, we’re potentially exposing sensitive conversations between Indian citizens and their government to entities operating under foreign jurisdictions,” Dheemanth Reddy, Co-founder and CEO of voice AI startup Maya Research, tells AIM.
Reddy warns that widespread use of voice AI for government services could expose sensitive personal data, as millions of Indians may share information about their finances, healthcare, identities, and grievances through these platforms. “At that scale, the data could reveal extremely sensitive information about Indian society,” he notes
Raoul Nanavati, CEO and Co-founder of enterprise voice AI company Navana.ai, believes that choosing a foreign company for Kannada citizen helplines sends a signal to banks, insurers, and government departments that global firms are the default choice.
“A government contract is the most valuable reference an Indian AI company can earn. Foreign firms don’t need it. Indian startups do,” Nanavati says, adding that losing such contracts deprives domestic companies of the credibility they need to win more customers.
Ankush Sabharwal, CEO of CoRover.ai, says the government should prioritise Indian AI capabilities where they exist, not as protectionism but as a strategy. He cited data sovereignty, better linguistic fit, cost control, and IP and talent retention within the country as key advantages.
“The risk is that preference becomes blind protection, so Indian companies must earn contracts on outcomes, accuracy, and security,” he says.
Indian AI is Cheaper
Domestic voice AI startups have a thorough focus on Indian languages, regional accents, and enterprise use cases. Startups such as Sarvam AI, Gnani.ai, BlueMachines, Corover.ai, and Smallest.ai are developing products across speech generation, transcription, and conversational AI.
“India’s emotions, sentiments, identities, and cultural nuances live in its voice, not just in text. Voice data can reveal far more about individuals and communities than a written transcript ever could,” says Reddy.
Sasank Chilamkurthy, Founder and CEO of AI infrastructure company JOHNAIC, argues that Indian voice AI companies could deliver similar services at lower costs and with better support for Indian languages.
JOHNAIC has tested several voice AI APIs through Voicera, an open-source project by EkStep Foundation deployed in Maharashtra and other locations. Chilamkurthy says that Indian providers are better suited to support large-scale government deployment applications because of their pricing and Indian-language capabilities.
ElevenLabs’v4 text-to-speech API was priced at approximately ₹2.11 per 1,000 characters at the discounted rate till October 12, compared with its original price of ₹7.66. Its v4 Turbo model costs around ₹1.05 per 1,000 characters under the discounted rate, down from the original ₹3.83.
Meanwhile, Sarvam AI charges ₹3 per 1,000 characters for both its real-time and streaming text-to-speech API services. Sarvam AI’s latest offerings include Bulbul V4 Flash, its text-to-speech model with more than 200 persona voices, and Saaras V4, its speech recognition model supporting 22 Indian languages and English.
Chilamkurthy adds that domestic providers have an advantage over ElevenLabs when it comes to regional languages.
Sarvam AI is also working with companies such as Swiggy, Razorpay, IBM, HP, and IDFC FIRST Bank, as well as the Maharashtra government, on voice-led commerce, enterprise AI, and public-sector applications.
Gnani.ai, meanwhile, has developed several speech AI models, including Vachana STT for speech recognition, Timbre v3.0 for text-to-speech, Prisma v2.5 for speech-to-text, and Warp v2.0 for direct speech-to-speech interactions. Its customers include Tata Motors, TVS Credit, Bajaj Allianz, Bank of Baroda, Indian Bank, Fibe, and Concentrix.
These companies could also participate in government projects involving multilingual interfaces, automated helplines, and voice-based services. Their presence raises the question of whether Karnataka will create opportunities for domestic firms alongside its international partnerships.
“The opportunity in voice AI and voice commerce goes far beyond making technology sound more human; it lies in making technology work more naturally for people and businesses”, says Khilan Haria, Chief Product Officer, Razorpay.
He further adds that 20–30% of all searches and an even higher percentage of chat communication in India is voice. Interestingly, Razorpay has partnered with three voice AI startups. “Our partnerships with ElevenLabs, Gnani.ai, and Sarvam AI reflect this belief. We have already deployed voice AI solutions at scale across multiple use cases, including marketing, sales, and risk,” Kharia says.
For early-stage companies, public-sector deployments can provide opportunities to test products in real-world conditions and establish credibility with other customers. If governments primarily adopt established global platforms, domestic startups could find it harder to secure similar opportunities.
Chilamkurthy proposes two possible explanations for Karnataka’s engagement with foreign AI companies.
The first is a possible preference for foreign technology vendors over Indian companies. Drawing on his experience selling technology to government customers, he alleges that Indian businesses often faced a perception problem, with international brands sometimes perceived as more reliable.
His second hypothesis is that the engagement could be linked to efforts to persuade them to establish offices or global capability centres (GCCs) in Bengaluru. “When they want Anthropic and ElevenLabs to come set up an office here, they have to give something to them,” he remarks.
Indian Startups Need Opportunities
Nanavati says that governments in the US, China, and the EU use public procurement to strengthen domestic technology companies. He mentions India’s UPI, Aadhaar, defence, and space programmes as examples of ecosystems that grew with the state as an early customer.
“We keep saying India needs to own its AI stack, and then we hand our most important use cases to companies headquartered elsewhere,” he says.
Nanavati explains that domestic AI providers can keep data within India, offer pricing suited to local budgets and support audits and on-premise deployments, while public spending can help build local talent and intellectual property.
He acknowledges that some domestic companies may not be competitive, but open evaluations and clear benchmarks could help ensure the best companies are chosen.
Nanavati says that relying on foreign AI models could leave India dependent on other countries for technology, pricing, and data. “If the model sits abroad, India is renting intelligence, not building it,” he says, adding that decisions on product roadmaps, language priorities, and model updates would be made outside India.
He also cautions that foreign service providers could remove introductory prices once customers become dependent on a platform, with pricing set in dollars rather than for India’s low-cost public services.
On data, he says that citizens’ voice recordings could be a valuable resource for training AI models. “If that data trains someone else’s system, we give away our most valuable national AI asset for free,” he notes.
The key issue is giving domestic companies a fair chance to compete for large deployments, build credibility, and improve their products through real-world use.
“The smart approach is model-agnostic architecture, where no single vendor becomes a single point of failure,” Sabharwal says, warning that technology lock-in, rising usage-based costs, and a lack of transparency over where inference takes place could create risks at scale.
Your reaction