How India Uses AI

Anthropic's India Country Brief puts India second globally in Claude.ai conversations at 5.8% of the November 2025 sample, behind only the United States. Rank the same traffic per working-age person and India falls to 101st of 116 countries with enough data. One week of Claude.ai consumer use. Two different stories.

Chatbots, hospital fraud screens, weather models and annotation jobs still get counted together. The sources, dates and limits differ by layer.

Tap to enlarge

Cash, GPU and fellowship counts for the IndiaAI Mission sit in The IndiaAI Mission.

Absolute counts and per-person ranks

IAMAI and Kantar's Internet in India Report 2025, launched at the India Digital Summit on 29 January 2026, estimates 958 million active internet users. Of those, 44% said they had used AI-enabled features: voice search, image search, chatbots, filters. Ages 15-24 came in at 57%. Ages 25-44 at 52%. The base is roughly 87,000 samples. The question catches light engagement as well as heavy use.

A Google-Kantar study from April 2025 asked more than 8,000 people across 18 cities. Sixty percent said they were unfamiliar with AI. Only 31% had tried a generative tool. About 75% wanted a daily growth collaborator. Different method, different year, different result. Merge the two into one adoption rate and the certainty looks fake.

WhatsApp's Will Cathcart told Economic Times that India has the most monthly active Meta AI users and leads the world. Meta did not publish the India total beside that claim. In July 2024, Meta CFO Susan Li had already called India Meta AI's largest usage market. The ranking is stable. The headcount is still missing.

Google's Gemini app crossed 1 billion monthly users globally by August 2026. Hindi and other Indian-language access grew. India-specific products such as Personal Intelligence and Gemini Spark shipped. Google still has not published India's share of those billion monthly users.

Airtel's 17 July 2025 note offered customers a free year of Perplexity Pro. Carrier deal. Claim windows expire. It is not an AI-user census. TRAI's December 2025 quarterly figures put total internet subscribers at 1,028.61 million. June 2026 coverage puts broadband near 1,087.50 million. Subscriber lines and IAMAI's active-user count are different series. Neither is an AI-user count.

Swiggy's FY25 annual report describes analytics-backed recommendations and Pyng, an AI-driven booking layer for verified professionals. Meta's work with the Retailers Association of India says 72% of product discovery in its surveyed retail context happens on WhatsApp. Commerce signals. They do not prove that shoppers name the experience as AI.

Tap to enlarge

Maharashtra (15.5%), Tamil Nadu (13.2%), Karnataka (12.7%) and Delhi (10.5%) together pass half of India's Claude.ai use. Anthropic notes that the concentration tracks IT geography. Software-related tasks make up 45.2% of Indian Claude use, the highest country share in that brief. Work-related use is 51.3%, above the global 46%. Personal use sits below the global average. The sample covers one week in November 2025 and Claude.ai consumer use only, so Meta AI, Gemini, enterprise APIs and offline systems fall outside it.

Silent systems behind the chat screen

A second layer of use never asks a person to open an AI app. It sits inside programme workflows.

The Ministry of Health told Lok Sabha on 5 December 2025 that a Clinical Decision Support System inside eSanjeevani had supported standardised data capture for 282 million consultations from April 2023 to November 2025. That figure is CDSS-backed capture across those visits, not a count of AI diagnoses. The same reply describes AI-based differential diagnosis recommendations and names related tools such as diabetic retinopathy detection and an abnormal chest X-ray classifier. The Hindu reported ministry plans to extend CDSS into more hospitals. Some coverage of the same reply also reports about 12 million consultations with direct AI diagnostic assistance. Capture volume, doctor-facing suggestions and diagnosis counts are different acts. The doctor stays the decision-maker.

Reporting on the National Health Authority Annual Report 2025-26 says the National Anti-Fraud Unit blocked suspicious claims worth ₹678.47 crore, penalised 2,842 hospitals ₹114.06 crore, de-empanelled 2,003 hospitals and suspended 839. Former NHA CEO Indu Bhushan has described more than 100 algorithms refined through feedback. Those agency-reported outcomes leave false-positive harm and residual fraud unmeasured.

A 1 April 2025 PIB release said Kisan e-Mitra had answered more than 92 lakh queries and was handling over 20,000 queries a day in 11 languages. Later Lok Sabha coverage in December 2025 put answered queries above 93 lakh, with more than 8,000 queries a day. The National Pest Surveillance System, in the same April reply, was used by more than 10,000 extension workers and covered 61 crops and over 400 pests. Query counts measure contact volume. Farm yield gains sit outside those tallies.

A 3 April 2025 PIB note said that as of 28 March 2025, 36,344 Supreme Court judgments had been translated into Hindi and 47,439 into other vernacular languages, for 83,783 uploaded translations on the e-SCR portal. SUVAS is the named AI translation tool. The February 2026 PIB backgrounder says AI use remains assistive and subject to judicial review. Translation volume stops short of automated judging.

The Bharat Forecast System, according to a 22 July 2026 Lok Sabha reply, runs at 6 km resolution for real-time prediction. Government statements also cite about 30% better accuracy for extreme rainfall versus the previous operational model in research mode. Separately, a Google Research post describes University of Chicago forecasts, blending NeuralGCM with other models and IMD history, reaching about 38 million farmers by SMS in 2025. That ministry delivery partnership leaves the operational monsoon model under IMD ownership.

Civil Aviation coverage in late May 2026 said DigiYatra had crossed 10 crore passenger journeys across about 38 airports, with later Lok Sabha reporting near 10.75 crore journeys and 39 airports. DigiYatra uses facial recognition. Stretching every biometric match into "AI use" overclaims the label, yet the system still belongs in any account of machine vision at this scale.

Airtel's 21 April 2025 release said its AI spam tool had flagged over 27.5 billion spam calls since September 2024, about 1,560 a second, and claimed a 16% drop in spam calls for customers. Those are company metrics. The release omits an independent false-positive audit.

The RBI FREE-AI committee report, released 13 August 2025, sets seven sutras and six pillars for financial AI. SBI executives have described AI underwriting near ₹1 trillion of MSME loans in FY26 and cheque automation for smaller amounts, in FIBAC coverage. HDFC Bank's FY26 annual report describes an in-house platform called Neev and says advanced analytics cut eligibility-search effort by about 50% for pre-approved loan limits across channels. These remain self-reported operating claims. A lot of India's AI use looks like workflow automation inside institutions that already hold the data.

Workplace surveys and firm maturity

Microsoft and LinkedIn's 2024 Work Trend Index India findings put workplace AI use among Indian knowledge workers at 92%, against a global 75%. BCG's AI at Work 2026 reports 97% of Indian managers and leaders using AI regularly, and 70% of Indian frontline AI users saying they save at least one full workday a week, against a global 42%. Microsoft's 2025 Work Trend Index India release says 93% of Indian leaders intend to use AI agents to extend workforce capacity within 12-18 months, and 59% already use agents to automate workstreams.

Those figures describe sampled desk and knowledge workers. They leave the wider labour force out, and they leave open whether pilots create durable value.

NASSCOM's AI Adoption Index 2.0, summarised on IndiaAI.gov.in, gives India an aggregate maturity score of 2.47 on a 4-point scale in 2024, barely above 2.45 in 2022, with 87% of 500 surveyed firms in the middle Enthusiast or Expert bands. The study covers seven sectors that the authors say represent about 75% of GDP. Salesforce's Agentic Workplace Study India findings, released 30 July 2026, say Indian desk workers are 45% more likely than the global average to put AI in their core workflow, and also report the highest share of unsuccessful AI pilots: 38% of Indian respondents said their organisation had an unsuccessful AI pilot in the past year, versus 28% globally. Among workers whose pilots failed, 34% cited missing business context. BCG's global AI at Work note says tool access moves impact only a little when strategy and workflow redesign stay weak.

Tap to enlarge

Reported tool use runs hot while organisation redesign lags. People bring chatbots to the desk; firms often leave the workflow vague. Salesforce's India sample covers 120 desk workers, so treat the 38% failed-pilot figure as directional.

TCS said annualised AI revenue crossed $2.3 billion in Q4 FY26. Infosys said AI services were 8.2% of revenue in Q1 FY27. HCLTech reported annualised Advanced AI revenue of $620 million in FY26. Wipro has said several hundred enterprise-grade agents are in production across its delivery platforms. Each firm defines AI revenue differently. Booked work still pulls the debate off survey optimism.

Tap to enlarge

Talent data from the Stanford AI Index 2026, as covered by Indian Express and ThePrint, places India second for identified AI authors and inventors at 50,460 in 2025, behind the United States. LinkedIn-relative AI skill penetration leads at about 3.0 times the global average. Student generative-AI use is reported near 84% in 2025 coverage of the same index. India also shows high net talent outflow in that reporting.

The labour that trains models

Models need labelled text, images, audio, video and, increasingly, human motion.

NASSCOM's data annotation work has long projected that the market India serves could exceed $7 billion by 2030, with up to about 1 million workers in full-time and part-time models. Economic Times reporting has cited industry estimates of roughly 20,000 full-time managed-service annotators and about 50,000 Indian contractors on international platforms. Those remain industry estimates rather than a labour census.

Field reporting from Tamil Nadu's Karur district makes the work concrete. AFP coverage republished in June 2026 describes Objectways workers wearing head cameras and smart glasses to record household and factory tasks for robotics training. Later TechTimes coverage of related reporting puts some pay near $2.62 an hour. The New York Times published a related August 2026 piece on annotation work in India, including Karur. The open AFP and TechTimes accounts carry the on-the-ground detail used here; the NYT URL is linked for readers with access.

India buys AI tools at scale and also sells labelled data and human judgment that make those tools work. Chatbot rankings and annotation payrolls measure different markets. Policy that counts only GPUs and foundation-model contests will miss the towns where the training data is made.

A Nomura note, as reported by Times of India and India Today, counted about 83,100 AI-related hires against about 31,921 AI-linked losses in India, drawn from a review of Asia cases mostly spanning 2022 through August 2026. The India net is positive in that ledger. Hiring often lands in AI delivery roles. Losses often hit support teams. Annotation work can expand even while some office roles shrink. Net headcount remains a weak proxy for job quality.

What the government stack claims

Government backgrounders around the February 2026 AI Impact Summit describe an India AI stack across applications, models, compute, networks and energy. The 4 February 2026 PIB note on the India AI Stack and related Significance of AI for India materials list sector deployments and infrastructure claims. They map official priorities more cleanly than they prove outcomes. The cash, GPU and fellowship ledger for the Mission sits in The IndiaAI Mission. The 13 August 2026 PIB backgrounder and 5 August 2026 Lok Sabha reply are the better recent official tallies for AIKosh dataset counts, sanctioned GPU hours and Centres of Excellence.

Lok Sabha figures reported in December 2024 put about 7.91 lakh secondary and 50,343 senior-secondary CBSE students in AI courses for 2024-25. Later ministry data put AI skill modules or subjects in 18,839 CBSE schools. YUVA AI for All is the national short literacy course. August 2026 parliamentary coverage cited by TechObserver says more than 26.5 lakh people completed it. Enrolment and completion are different words in other coverage of the same programme, so keep the verb attached to the source.

Reliance and Google announced an October 2025 partnership covering Gemini distribution to Jio users, TPU access and Gemini Enterprise go-to-market through Reliance Intelligence. Company pledges precede completed capacity. Microsoft and Amazon have also announced multi-year India cloud and AI infrastructure commitments in the tens of billions of dollars. Those envelopes mix ordinary cloud spend with AI-labelled capacity, so they should stay outside any IndiaAI cash total.

Bhashini and related translation stacks appear in summit materials, court tools and farmer chatbots. When a farm query or a judgment translation works in a regional language, the visible use often depends on earlier language and speech work that users never see.

What the numbers leave out

India posts high absolute chatbot volume, led by WhatsApp distribution and strong Claude use among software workers. Outside that core, Claude per-person use stays thin in Anthropic's sample. Meta AI leads on absolute users by company claim, without a published India headcount.

Much of the state's machine help stays off the chat screen: eSanjeevani CDSS data capture at 282 million consultations, claims screening, pest alerts, judgment translation, weather products, biometric gates and network spam filters. Event counts are not the same as chatbot habits, and capture is not the same as diagnosis.

Workplace surveys and IT-services revenue lines show intense tool use among sampled professionals. Firm maturity scores and failed-pilot rates, including a small Salesforce India desk sample, show many organisations still struggle to redesign the work around those tools.

India also supplies labelling and motion-capture labour for global models. Field reports put that work in non-metro towns. It barely appears in chatbot rankings.

Absolute volume and per-person use answer different questions. So do chatbot rankings, silent programme traffic, IT delivery lines and annotation payrolls. Mix them and India looks like one market. Keep them apart and the map gets clearer.