Bhashini's Billions of Calls, Zero MSME Directory: What India's Language Stack Wasn't Built For

What Bhashini actually is, and how big it has gotten

Bhashini launched at Gandhinagar in July 2022 as the flagship product of the Ministry of Electronics and Information Technology's National Language Translation Mission. Its job, stated plainly, is to help Indian citizens communicate and transact across languages: machine translation, automatic speech recognition, and text-to-speech, offered as free, open APIs across all 22 constitutionally scheduled Indian languages. More than 300 pre-trained AI models are available to partners under that mandate, at zero marginal cost to integrate.

By 2026 the scale is genuinely large. Trade press coverage of India's AI policy landscape reports that Bhashini processed roughly 2.5 billion API calls over the course of 2026, on pace to exceed 5 billion by year-end, and that the platform is woven into DigiLocker, the successors to Aarogya Setu, and the UMANG super-app, the everyday citizen-services layer of the Indian state. Every one of those calls is, in effect, a sentence moving between an Indian language and a system that could not otherwise read it.

That is a real achievement, and it is also a precise description of the job Bhashini was built to do: move meaning between languages inside services that already know who the citizen is. A DigiLocker session already has the citizen's identity on file. A grievance portal already has the case number on record. Bhashini's role in each is translation, not identification. Nothing in that design brief asks the platform to describe a business to a machine that has never encountered it before.

The MSME layer Bhashini touches

Bhashini's reach into commerce runs mainly through the Open Network for Digital Commerce. In September 2024, ONDC launched Saarthi, a Bhashini-powered reference application that lets businesses build their own buyer-facing apps with voice and text support in Indian languages, starting with five, Hindi, English, Marathi, Bangla, and Tamil, with a stated plan to extend to all 22. On the government-scheme side, the Ministry of MSME's Trade Enablement and Marketing initiative, a sub-scheme of the World Bank-supported RAMP program, set out in June 2024 to onboard five lakh micro and small enterprises onto ONDC, half of them women-owned, backed by an outlay of Rs 277.35 crore over three years and implemented by the National Small Industries Corporation. The support on offer is concrete: help with digital cataloging, account management, logistics, and roughly 150 e-commerce workshops planned for tier-2 and tier-3 cities.

That scheme sits inside a larger 2026 push. At the MSME Day 2026 event on 27 June, the government reported that Udyam registrations, the formal MSME identity number, had crossed 8.7 crore, an enterprise base it says anchors the livelihoods of 38 crore people and contributes 31 percent of GDP, 35 percent of manufacturing output, and close to 45 percent of exports. Alongside that figure it launched or upgraded a cluster of MSME portals, PMEGP 2.0, SAMADHAAN 2.0, MSME Global Mart 2.0 (built to integrate with the ONDC ecosystem), and a new Testing Portal, all now offering services in all 22 scheduled languages through Bhashini and the National Informatics Centre.

Set against that infrastructure, the actual behavior of MSMEs online looks thin. A 2025 SIDBI-linked survey, reported by YourStory, found that while more than 90 percent of MSMEs now accept digital payments, only 13 percent actively use digital marketing or e-commerce to reach customers, with roughly seven in ten still relying on in-person referrals, print advertising, or trade events. Payment rails arrived. Discoverability did not follow at anything like the same pace.

Citizen infrastructure and commerce-identity infrastructure are not the same build

It is worth being precise about what each of these systems actually stores. Bhashini stores language models, not business records; a call to its API carries a sentence in and a sentence out, with no requirement that the sentence describe a company, its address, its hours, or its credibility. Udyam stores a registration number and classification data, built for eligibility and scheme access, not for public discovery. ONDC's registries function, by its own technical documentation, like a DNS for commerce, a directory of network participants for the purpose of routing transactions between buyer and seller apps, not a public, crawlable business directory addressed to the open web or to the systems that read it.

ONDC's own developer documentation is candid about the difficulty this creates even inside the network: existing platforms, brands, and merchants do not share a common cataloguing schema, so the protocol has had to build standardization work just to make listings comparable to each other across sellers. That is a live, acknowledged engineering problem for commerce inside the network. It says nothing about whether a listing, once standardized for ONDC's own buyer apps, is also structured in the schema.org, entity-graph terms that an external answer engine can parse, corroborate, and cite. Those are different specifications solving different problems, and nothing in the public record indicates ONDC's catalog schema was designed with the second one in mind.

None of this is a criticism of any single system doing the job it was funded to do. It is the observation that a citizen can now file a grievance, receive a scheme benefit, or place an ONDC order in her own language, and none of those interactions produces the artifact an AI answer engine actually consumes: a structured, corroborated, standing description of a specific business that exists independent of any single transaction.

What an answer engine needs, that none of this supplies

The 2024 peer-reviewed paper that introduced "Generative Engine Optimization" tested what actually moves a source into a generated answer, and found that citations, quotations, and authoritative, corroborated content raised a source's visibility inside the engines it tested. The unit being optimized is not a page rank; it is inclusion in a synthesized answer, and the inputs that earn it are structural and evidentiary properties of a business's presence on the web, sustained over time, not a single well-translated transaction.

The consequence of missing that layer is growing for every business, in every language. A 2025 Pew Research Center browsing-panel study found that users clicked a traditional search result in about 8 percent of searches when an AI summary was present, against 15 percent without one, and clicked a link inside the summary itself only about 1 percent of the time. Ranking is worth less than it used to be; being named inside the answer is worth more.

The open question no one has measured yet

There is no public dataset, to date, that measures whether Indic-language MSME businesses are actually surfaced or cited when Indian buyers ask AI answer engines a vernacular question, "मेरे पास एक अच्छा प्लंबर सुझाएं" instead of "suggest me a good plumber near me." Neither MeitY, ONDC, nor any independent academic body has published sampled-query results comparing citation rates for businesses with structured, corroborated identity against businesses without it, broken out by language. That comparison would be the rigorous way to test the thesis this piece argues: draw a matched sample of MSMEs across a few scheduled languages, some with structured schema and consistent citations across the web and some without, run the same buyer-intent queries in each language across two or three major answer engines, and measure how often each group is named. Until that work exists, the claim that vernacular translation scale and MSME AI-visibility are two separate gaps is a reasoned inference from what these systems actually store, not a measured finding, and it should be read that way.

What is measurable today, and already published, is the shape of the inputs: a translation layer moving billions of calls, an MSME base of 8.7 crore mostly relying on referrals and print rather than digital discovery, and a commerce network whose own documentation admits its catalog data is not yet standardized for its own buyer apps, let alone for external answer engines. Those three facts, taken together, are consistent with a business being fluently translatable and still functionally invisible to the systems now doing a growing share of the recommending. They do not, on their own, prove the size of that invisibility. That is the boundary of what public data currently shows.

How to read the evidence

Bhashini is not failing at its job. It is succeeding at a job that was never the job of making a business machine-legible. Conflating the two, treating vernacular translation capacity as if it were vernacular AI-visibility, is the mistake this piece is trying to head off, because it leads to comfortable, wrong conclusions: that a Hindi-speaking shop owner whose government portals now work in Hindi is thereby findable by an AI engine answering a Hindi query. Nothing in the public record supports that leap.

The corrective is not a bigger translation model. It is a structured, corroborated, standing record of who a business is, in whatever language its buyers actually use, built and maintained independent of any single government scheme or commerce network. That is a different build than the one India has funded so far, and it is the specific thing a business can check for itself rather than infer from national statistics.