Maharashtra Puts Sarvam's Sovereign AI on 2,500 Officials' Desks as Indian Governments Turn to Home-Grown AI

MUMBAI, Oct 29 – About 2,500 officials across the Maharashtra government can now use an artificial intelligence workspace built by Bengaluru-based Sarvam AI to draft documents, analyse reports and summarise files, the company said on Thursday, making the state one of the largest Indian adopters of generative AI inside day-to-day administration.

Chief Minister Devendra Fadnavis launched the platform on Wednesday at Sahyadri Guest House in Mumbai. Sarvam calls the product Indus Cowork; the state government's release calls its version MahaINDUS.

"Indus Cowork is now live for 2,500 officials across the Maharashtra government," Sarvam said in a post on X. Officials can also search Government Resolutions across departments, find the latest rules and check which ones have been replaced, it said.

The agreement

The Chief Minister also presided over the signing of an Enterprise Agreement between the state's Commissionerate of Electronics, Information Technology and Artificial Intelligence and Axonwise Private Limited, the company behind Sarvam, the Chief Minister's Office said. The agreement covers "specified AI-based software and/or services" to the state government. Minister Nitesh Rane and Sarvam co-founder Vivek Raghavan attended.

The CMO post did not give the value or duration of the agreement.

Earlier figures point to the likely scale. Media report: a Government Resolution in August, cleared by a High-Power Committee headed by the Chief Secretary, approved Indus for around 2,500 users at an indicative first-year cost of ₹5.63 crore and a two-year outlay of ₹11.26 crore (report). It set a subscription of ₹1,200 per user per month for the first 12 months, against a standard rate of ₹2,500. It is not confirmed whether the agreement signed on Wednesday carries the same terms.

What officials will use it for

According to the state's release, as carried in media reports, the platform supports drafting, translation, summarisation and data analysis in 22 Indian languages, including Marathi (Media report). It has three headline tools:

• GR Expert, to search Government Resolutions and track changes to them
• Email Summariser, to condense official correspondence
• RFP Bid Evaluator, to compare and assess tenders

Sarvam's partnership page adds research across policies and precedents, first drafts of notes and replies, analysis of data for patterns and exceptions, and conversion of scanned records into searchable text.

Media report: the August Government Resolution also listed file-noting help, RTI response drafting, optical character recognition across Indian scripts, and integration with eOffice and NIC e-mail where API access is available. The same package covered Samvaad voice and WhatsApp agents for citizen outreach, including a one-month pilot calling tuberculosis patients in local languages, and Sarvam's Saaras speech-recognition and Bulbul text-to-speech models for the state's MahaAI platform.

"Launching is not enough"

The Chief Minister said AI, integrated databases and digital systems would help the government identify eligible beneficiaries, verify claims and assess whether welfare schemes work, according to media reports of the official release (Media report). He said Maharashtra was the first state in the country to use sovereign AI in this manner. MoCA could not independently verify that claim; Sarvam describes the state more cautiously as "one of India's first" to do so at this scale.

He told departments that setting up digital platforms was not enough and that they must use them actively, so that technical shortcomings surface and can be fixed in time.

MahaINDUS was launched alongside MahaDBT 2.0, which added 36 schemes, Aaple Sarkar 2.0, which added 71 services, and new services on the MAITRI portal. Chief Secretary Rajesh Agrawal and Information Technology and AI Department Secretary Virendra Singh were among those present, the release said.

Governments turning to Indian AI

Maharashtra's rollout is part of a wider move by Indian governments to put home-grown AI into public services.

• IndiaAI Mission: Sarvam is one of 12 organisations the Centre is supporting to build sovereign foundation models, with financial and compute support of ₹246.72 crore, the Press Information Bureau said in February. Sarvam, BharatGen and Gnani models were launched at the IndiaAI Impact Summit 2026, and Sarvam and BharatGen models are available on the government's AIKosh platform, PIB said in March.
• Aadhaar: the Unique Identification Authority of India partnered with Sarvam in March 2025 for voice interactions, real-time fraud alerts and multilingual support in 10 languages, on a GenAI stack hosted on-premise in air-gapped UIDAI infrastructure, PIB said.
• Odisha and Tamil Nadu: Odisha is working with Sarvam on a 50 MW sovereign AI capacity hub, and Tamil Nadu and IIT Madras on a sovereign AI research park anchored by a 20 MW data centre, according to PIB.

Why a sovereign solution is needed

The case for a home-grown system rests on a simple point: whoever runs the AI platform can, in principle, see what goes into it, and is bound by the laws of the country where it is based.

Foreign AI providers say openly that they watch for misuse of their platforms. Anthropic, the US maker of the Claude models, said in November 2025 that it had detected and disrupted what it called the first reported AI-orchestrated cyber espionage campaign, run by a group it assessed with high confidence to be Chinese state-sponsored. The group used Anthropic's Claude Code tool against roughly 30 targets, including government agencies, and Anthropic said it banned accounts, notified affected entities and coordinated with authorities (Anthropic). "We only have visibility into Claude usage," the company noted.

Its August 2025 threat report described a data-extortion operation, a North Korean remote-worker fraud scheme and AI-built ransomware, all found on its platform (Anthropic). In its September 2026 report, Anthropic said it can often spot an influence operation "on Claude while the operation is still being built", because actors use AI to plan campaigns, choose targets and write material (Anthropic). OpenAI publishes similar reports; in October 2025 it said it had disrupted more than 40 networks since February 2024 (OpenAI). One case study described banning users who "appeared to be linked to Chinese government entities" (OpenAI).

That monitoring is aimed at criminals and hostile states, and it has stopped real attacks. But it shows that providers have the technical means to review activity on their systems.

Leading US AI companies are also closely tied to their own government. OpenAI launched OpenAI for Government in June 2025, with a Department of Defense pilot through the Chief Digital and Artificial Intelligence Office carrying a $200 million ceiling, and custom models for national security offered on a limited basis (OpenAI). Its ChatGPT Gov product, launched in January 2025, lets US agencies run the service in Microsoft Azure's commercial or government cloud (OpenAI). In August 2025 the US General Services Administration made ChatGPT Enterprise available to participating federal agencies for $1 per agency for a year (GSA).

Jurisdiction is the second issue. The US CLOUD Act of 2018 requires US service providers to preserve or disclose data "within such provider's possession, custody, or control, regardless of whether such communication, record, or other information is located within or outside of the United States" (US Department of Justice; Act text). Storing data in an Indian data centre does not by itself take it outside that reach if a US company controls it.

Courts can also override a provider's own retention rules. In 2025, a US court order in The New York Times' copyright case against OpenAI required the company to keep consumer ChatGPT and API content indefinitely. OpenAI said its obligations under that order ended on Sept. 26, 2025, and that enterprise customers were not affected (OpenAI).

This article does not suggest that any Indian government data has been misused. Enterprise products from US providers also promise not to train on customer data. The point is control: once government files sit on a foreign provider's systems, that provider's monitoring and that country's courts become part of the risk.

A model built in India, hosted in India and run on systems the state can inspect, as Sarvam and the Maharashtra government describe Indus, is designed to keep those files under Indian law.