A bank's collections agent in Indore speaks to customers in a Hindi that carries English loan terms, Marathi place names and the occasional Gujarati idiom. A district court in Odisha files pleadings in Odia and receives judgments from the High Court in English. A tyre plant in Tamil Nadu runs its shift handover in Tamil and its maintenance system in English. In each case the organisation has an AI programme, and in each case the programme was built and evaluated in English, because that is the language in which the models are strongest, the benchmarks are written and the vendors sell.
Indic language AI, the set of models, datasets and tools that make artificial intelligence work in India's languages, is where this gap gets closed or does not. It matters for enterprise because most of India's customers, workers and records are not in English. It matters for public service because the state's obligation to serve a citizen in a language she reads is not optional. And it is, in 2026, a field in which the public investment is large, the model performance is uneven and the commercial decisions are being made with very little discipline about what depth of language capability a given task actually requires.
That last point is the subject of this article. Not every Indic use case needs a frontier-grade generative model in twenty-two languages. Some need a transliteration layer, some need translation of structured documents, some need a model that understands one sector's vocabulary in one language, and a few need open conversational competence across dialects, code-mixing and speech. The article sets out what the evidence says about model performance in Indian languages, what the IndiaAI Mission now supplies, and a use-case map that sorts enterprise and public-service tasks by the depth of language they need, so that a buyer or builder can match the task to the tool.
What the evidence says about the language gap
The clearest independent measure is IndicGenBench, a benchmark covering 29 Indic languages across 13 scripts and four language families, published by Google researchers in April 2024 and presented at ACL. It evaluates cross- lingual summarisation, machine translation and cross-lingual question answering, and its headline finding is unambiguous: "there is a significant performance gap in all languages compared to English", with the largest PaLM-2 models performing best on most tasks. The gap is not a property of one vendor. It is a property of the training data.
A January 2025 analysis by researchers at Tattle Civic Tech, "Analysis of Indic Language Capabilities in LLMs", explains why. Of more than 400 large language models surveyed, the authors identified 28 with some Indic capability and 10 that were Indic-specific. Indian-language articles make up about 2.05 per cent of Wikipedia; Indian languages make up about 0.47 per cent of the Common Crawl sample the authors examined, with Hindi at 0.20 per cent, despite Hindi being the third most spoken language in the world and ranking sixty-second by number of Wikipedia articles. The authors rank twelve languages into tiers of relative model support, with Hindi, Bengali, Marathi, Telugu, Tamil and Urdu in the highest tier, Gujarati, Kannada, Malayalam and Assamese in the middle, and Odia and Punjabi lowest, while cautioning that the ranking "should not be treated as a universal ranking". They also note a sharp decline for very low-resource languages such as Sindhi and Santali, and a methodological problem that matters for anyone commissioning an evaluation: most Indic evaluation datasets are direct translations of English datasets, which tests whether a model can handle translated English rather than whether it understands how an Indian speaker frames a question.
There is a cost dimension that enterprise buyers often miss. Sarvam AI's December 2023 note on its OpenHathi series reported that Hindi text requires roughly three to four times as many tokens in the GPT and Llama tokenisers as equivalent English content; a custom tokeniser trained on Hindi closed most of that gap, leaving about 70 per cent of Hindi words unbroken into sub-word tokens. An Indic use case served by a general-purpose API model is therefore paying three to four times the per-word inference cost of the same workload in English, before any quality difference is counted. For high-volume tasks such as customer support or document processing, that multiple decides whether the economics work at all.
What the IndiaAI Mission supplies
The public response to this gap is the IndiaAI Mission, approved in March 2024 with an outlay of ₹10,371.92 crore over five years. For a builder of Indic language systems its relevant components are compute, data and models.
On compute, a PIB backgrounder of October 2025 reported that India had reached 38,000 GPUs under the mission, offered to startups, researchers and students at a subsidised rate of about ₹65 per hour. A PIB release of 17 February 2026, timed to the India AI Impact Summit, announced an additional 20,000 GPUs to be added in the following weeks, taking the stated total above 58,000. Subsidised compute matters disproportionately for Indic work because training and fine-tuning in low-resource languages is exactly the kind of experiment that commercial economics do not justify at frontier prices.
On data, the AIKosh platform held 7,541 datasets and 273 models by the February 2026 PIB backgrounder, up from over 3,000 datasets and 243 models reported in October 2025, when the platform had around 265,000 visits, 6,000 registered users and 13,000 downloads. Bhashini, the national language technology platform, was reported in the same October 2025 document as supporting 20 Indian languages, integrating more than 350 AI models, having crossed one million downloads and serving over 450 active customers.
On models, the mission has selected twelve foundation-model startups and consortia, including BharatGen and Sarvam, under its foundation-model pillar; an earlier October 2025 PIB document listed Sarvam AI, Soket AI, Gnani AI and Gan AI as selected from over 500 proposals. BharatGen, a government-backed consortium, launched on 2 June 2025 with a stated scope of 22 Indian languages across text, speech and image understanding. Its first published model, PARAM-1, described in a July 2025 technical report by the BharatGen team, is a 2.9 billion parameter decoder-only text model trained from scratch on a bilingual Hindi and English corpus with 25 per cent of the data allocated to Indic content and a SentencePiece tokeniser adapted to Indian morphology. The authors position it as a baseline for India-centric applications rather than a frontier competitor, and the report gives no quantitative benchmark results in its abstract. That is a fair description of where sovereign Indic models stood in 2025: credible, small, Hindi-first, and a long way from the breadth the mission's 22-language ambition implies.
What the mission supplies is not a finished Indic AI stack but the inputs to one: cheap compute, a growing public data repository, a translation and speech platform with a customer base, and small open models that can be fine-tuned. Each is useful at a different depth of the map below. None removes the need to evaluate.
Indic language AI use cases, mapped by language depth
The map is this article's original contribution. It sorts tasks by the depth of language competence they require, from the shallowest, where the model needs only to handle script, to the deepest, where it must hold an open conversation in a regional dialect with code-mixing and speech. For each depth it names the enabling technology, what the IndiaAI Mission supplies, representative enterprise and public-service use cases, and the evaluation a buyer should insist on before deployment. Depth levels are an analytical device; real tasks often sit at a boundary.
Reading the map produces three practical conclusions.
The first is that most of the value available to Indian enterprises today sits at D1 to D3, and most of the vendor conversation happens at D4. Name matching, document translation and domain extraction are tasks where the models are good enough, the evaluation is tractable and the cost advantage of an Indic-tuned tokeniser is realised at volume. Open conversation in dialect is where the language gap is widest, the evaluation hardest and the cost of failure most visible, because the failure is heard by a customer or a citizen in real time.
The second is that the IndiaAI Mission's contribution is strongest at D2 and D3. Bhashini is a translation and speech platform with 450-plus customers; AIKosh is a data repository; PARAM-class models are fine-tuning baselines. A builder at D2 or D3 can assemble a working system from public inputs and a modest fine-tuning budget. A builder at D4 still depends on frontier or well-funded private models, and on speech data that largely does not exist in public form for most dialects.
The third is that evaluation must be done in the buyer's language on the buyer's data. The Tattle researchers' observation that most Indic benchmarks are translations of English tests means that a vendor's reported benchmark score tells a buyer almost nothing about performance on a Marathi grievance letter or an Odia court order. The evaluation column of the map is therefore not an afterthought; it is the condition on which the rest of the row depends. A separate PanScience Research piece on evaluation before deployment sets out the general harness; the Indic-specific addition is that the harness must be run per language, because performance at D3 in Hindi says nothing about performance at D3 in Punjabi.
Indic language AI in the courts: a public-service case at D2
The judiciary is the sharpest public-service test of the map, because its language obligation is structural. Pleadings are filed in state languages, High Courts and the Supreme Court work largely in English, and a litigant's right to understand the judgment in her own case is not a convenience.
The scale is large. eCourts Phase III, approved by Cabinet in September 2023 with an outlay of ₹7,210 crore, builds on a system that had computerised 18,735 courts and placed 27.64 crore orders and judgments on the National Judicial Data Grid by December 2024. Pendency, per a Lok Sabha reply of February 2026, stood at 4.77 crore cases in district courts, 63.66 lakh in High Courts and 92,101 in the Supreme Court. Every one of those cases generates documents in at least one language, and many in two.
Translation of judgments into regional languages is the use case that has been running longest, and its history is instructive. The Supreme Court's SUVAS translation tool launched in November 2019 to translate judgments into nine languages. The Supreme Court Observer's analysis of November 2021 found that 232 translated judgments had been published between January 2019 and February 2020, with Hindi accounting for 49 per cent, and only 31 between March 2020 and October 2021, concluding that the project had "progressed slowly". The next three years were different. An IndiaAI news item of August 2024, citing official figures as of 5 August 2024, reported 36,271 Supreme Court judgments translated into Hindi and 17,142 into 16 other regional languages, available on the e-SCR portal, with eight High Courts having started e-High Court Reports. AI-assisted transcription of oral arguments in Constitution Bench matters had been in use since February 2023.
Two readings follow. First, the task is firmly D2: structured documents, a fixed legal vocabulary, a human review loop, and a translation platform rather than a conversational model. The jump from 31 translations in twenty months to tens of thousands in three years is what happens when a D2 task is given D2 tooling at scale. Second, the quality question is unresolved in public sources. Neither the 2021 analysis nor the 2024 figures report an accuracy measure, and a translated judgment that misrenders a legal term is worse than no translation. This is the evaluation gap the map's D2 row identifies.
The practice is moving beyond translation. PanScience Innovations, the New Delhi venture studio, reports that its venture Nyaay AI, which it describes as judicial workflow automation, has deployments in 14-plus Indian High Courts and three national supreme courts. These are company-reported figures and this article has no independent measure of them, but the category illustrates the D2-to-D3 boundary: workflow automation over court documents requires classification, extraction and summarisation within legal vocabulary, in the languages the files are actually written in. A separate PanScience Research piece on AI in the Indian judiciary covers the court workflow in detail.
Indic language AI in the enterprise: voice at D4, documents at D3
Enterprise demand concentrates at the two ends of the map. At D3, the volume is in documents: claims forms, grievance letters, field reports, dealer invoices and shop-floor logs that arrive in regional languages and must become structured data. This is where the token penalty bites hardest and where Indic-tuned small models and tokenisers earn their keep. Evaluation is tractable because the buyer has labelled historical data and the output is a classification or extraction that can be scored.
At D4, the demand is voice. Collections, customer service, field-force support and conversational commerce in regional markets all require speech recognition and synthesis in the customer's language, a model that holds a conversation, and handling of the code-mixing that is the actual register of Indian commercial speech. PanScience reports that its venture OpticAll offers voice AI for enterprise support; the company-reported description illustrates the category rather than evidencing its performance. The evaluation burden at D4 is the heaviest on the map: word error rate per language and dialect on real call audio, task completion, escalation to a human, and fairness across languages, because a voice system that performs well in Hindi and poorly in Kannada is discriminating against a defined set of customers.
Two cautions apply across both ends. The Reserve Bank's FREE-AI committee report of August 2025, whose seven sutras include Fairness and Equity and Understandable by Design and whose 26 recommendations include indigenous financial-sector AI models, is not yet binding regulation, but a bank deploying an Indic voice system should expect its language fairness to be examined under it. And the DPDP Rules notified in November 2025 apply to the personal data in every regional-language call recording and grievance letter; the language of the data does not change its legal status.
Conclusion
The question an Indian enterprise or public body should ask of any Indic language AI proposal is not "which model is best in Indian languages" but "what depth of language does this task actually need, and has the system been evaluated at that depth, in this language, on our data". Most tasks sit at D1 to D3, where the models are adequate, the public inputs from the IndiaAI Mission are genuinely useful and the cost advantage of Indic-tuned tooling is large. The courts' translation record shows what a D2 task delivers when given D2 tooling at scale, and also what remains unmeasured. D4, open conversation in dialect and speech, is where the ambition and the evidence are furthest apart, and where a buyer should demand the most before deploying.
The language gap in AI is real, documented and narrowing slowly. The disciplined response is to match the depth of the tool to the depth of the task, to evaluate per language, and to treat a benchmark score in translated English as what it is: a measure of something other than the job.
Methodology and sources
The depth map is the author's original classification; depth levels, use-case placements and evaluation requirements are analytical and illustrative, not measured. Benchmark findings are from IndicGenBench (arxiv.org/abs/2404.16816), submitted April 2024 and revised August 2024, and from Vaidya, Prabhakar, George and Shah, "Analysis of Indic Language Capabilities in LLMs" (arxiv.org/pdf/2501.13912), January 2025. Tokeniser figures are from Sarvam AI's OpenHathi announcement (sarvam.ai/blog/announcing-openhathi-series), 12 December 2023. PARAM-1 details are from the BharatGen technical report (arxiv.org/abs/2507.13390), July 2025. IndiaAI Mission figures are from the PIB backgrounder "Transforming India with AI" (static.pib.gov.in, 12 October 2025), the PIB release "India's AI Moment" (static.pib.gov.in, 17 February 2026), and the PIB backgrounder of February 2026 for AIKosh and the twelve foundation-model selections. Court figures are from the PIB Cabinet approval of eCourts Phase III (13 September 2023) and the PIB December 2024 backgrounder, the Lok Sabha reply of 13 February 2026 as reported by IANS, the IndiaAI news item "Supreme Court of India harnesses AI for judicial translation and research" (indiaai.gov.in, 9 August 2024), and the Supreme Court Observer's "The Supreme Court's translation project is slowing to a halt" (scobserver.in, 12 November 2021). The RBI FREE-AI report is as released in August 2025 and summarised by KPMG; DPDP Rules details are from PIB on the notification of 14 November 2025. PanScience Innovations facts are company-reported from panscience.xyz, read 24 October 2026; no client of any PanScience venture is named.