What AI could do inside Indian government
Department by department, where machine assistance could raise throughput, quality and transparency — and, for each, the way it would go wrong.
Written by Claude, an AI system, which makes the conflict of interest obvious enough to state plainly: an AI assessing where AI should be used will tend to find opportunities. The failure modes below are included precisely because that bias is real, and several of them argue against deployment.
The framing
Indian government capacity is not uniformly low — it is unevenly distributed. The state runs an election of almost a billion voters and a real-time payments system carrying more transactions than any other on earth. The same state cannot tell you how long a commercial dispute takes to resolve, or which stretch of highway killed the most people last year.
The gap is rarely ambition and rarely money. It is throughput: too few trained people, processing too many cases, in too many languages, with records that were never structured. That is a narrower problem than “governance”, and it happens to be the problem this technology is actually good at.
What follows is ordered by where the ratio of benefit to risk is best, not by what is most impressive.
The courts
Tens of millions of pending cases, with commercial disputes routinely taking years. The bottleneck is throughput per judge, and judges spend a large share of their time on work that is not judging.
- Transcription and translation. Proceedings in a dozen languages, currently transcribed slowly and inconsistently. This is the single highest-value, lowest-risk application in Indian government — it removes clerical load without touching any judicial decision.
- Case triage and bunching. Identifying the thousands of matters that turn on one settled question of law so they can be listed and disposed together rather than individually.
- Precedent retrieval across languages and jurisdictions, so a district judge has the research capacity currently available only to a well-resourced litigant.
- Defect detection at filing, catching the procedural errors that cause a matter to be returned weeks later.
How it goes wrong. Never let a model decide, recommend an outcome, or score a litigant. The legitimacy of a judgment rests on a human being accountable for it, and an 'AI-assisted' recommendation is anchoring whether or not the judge is told to ignore it. Restrict this to clerical throughput and the risk is genuinely low; extend it to adjudication and the damage is irreversible.
Tax administration
GST generates enormous transaction-level data. Fraud detection, refund processing and assessment quality all depend on finding patterns in it, and interpretive ambiguity routinely resolves against the filer.
- Circular-trading and fake-invoice detection. This is a graph problem on data the system already holds, and it is the clearest technical fit in government.
- Refund triage, so straightforward claims clear automatically and scrutiny concentrates where the risk is.
- Drafting assistance for assessment orders, which would reduce the share of demands later set aside for being unreasoned — a large and measurable category.
- A genuinely useful taxpayer-facing assistant that answers 'does this apply to me' against the actual notification history.
How it goes wrong. Selection for scrutiny is the danger. A model trained on past selections learns past biases, including the tendency to scrutinise the accessible rather than the egregious. If a model flags a taxpayer, the taxpayer must be told the reason in terms they can contest. An unexplainable demand is not a demand, it is a summons to a process — and given what enforcement delay already costs Indian firms, adding an opaque one would be a serious own goal.
Land records
Fragmented, often un-digitised, frequently inconsistent between the revenue record, the registry and what is on the ground. A large share of civil litigation traces back to this.
- Digitisation of handwritten records in regional scripts, including the older ones current OCR handles badly.
- Reconciling the same parcel across registries that disagree, and flagging the disagreements for human resolution rather than resolving them.
- Mutation-backlog triage.
- Cross-checking cadastral maps against satellite imagery to find unrecorded change.
How it goes wrong. This is the highest-stakes item on the page. Land records determine who owns what, and an automated 'correction' that dispossesses someone is not a bug report, it is a family's life. Every model output here must be a flag for human adjudication with full audit trail and an accessible appeal — never a write to the record of title. The people most exposed to a wrong correction are precisely those least able to litigate it.
Public health
Too few specialists, distributed unevenly, with the largest gaps where need is highest.
- Screening support at the periphery — diabetic retinopathy, TB on chest X-ray, cervical and oral cancer. These are established, the evidence is reasonable, and the comparator is not a specialist but no screening at all.
- Decision support for ASHA and ANM workers in their own language, on the phone they already carry.
- Disease surveillance across IDSP reporting, to shorten outbreak detection.
- Supply-chain forecasting for drugs and vaccines.
How it goes wrong. Every screening model has a false-negative rate, and a missed cancer is a death. The correct framing is triage that escalates to a human, never a discharge decision. There is also a well-documented performance drop when models trained on one population are deployed on another — validation must happen on Indian patients, in Indian conditions, on the equipment actually in use, not on the manufacturer's published figures.
Agriculture
Advisory reaches a fraction of farmers, crop insurance claims are slow and contested, and procurement operations are enormous.
- Voice-first advisory in the farmer's own language — the strongest fit anywhere in Indian government for recent speech models, because the users are numerous, the literacy assumption fails, and the incumbent is a call centre.
- Crop-loss assessment from satellite and drone imagery, which would settle insurance claims in days rather than seasons.
- Yield and arrival forecasting to improve procurement and price policy.
- Pest and disease identification from a photograph.
How it goes wrong. Advisory that is confidently wrong destroys a season's income for someone with no buffer. Agronomic advice must be grounded in ICAR and state-university recommendations for that specific agro-climatic zone, not generated freely — and it should say 'ask your KVK' when uncertain rather than guess. Automated claim assessment must be appealable to a human, or it becomes a machine for denying claims at scale.
Policing and public safety
FIR quality varies enormously, case management is weak, and the data problems described elsewhere on this site start here — with what does and does not get recorded.
- FIR drafting assistance, ensuring the offence is correctly framed and the required particulars captured. Badly framed FIRs collapse cases years later.
- Linking related cases across jurisdictions, which is currently near-impossible.
- Road crash record-keeping. India has the world's highest road death toll and no national crash-level database — structured capture at the thana would be transformative for road safety policy.
- Court-facing case tracking so investigations do not lapse on procedural deadlines.
How it goes wrong. Predictive policing should be off the table. It has been tried in several countries, and the consistent finding is that it relocates enforcement toward over-policed populations while wearing the authority of a model. In a country with India's history of caste and communal policing, the harm would be severe and would be laundered through an algorithm. Facial recognition for general surveillance raises the same objection without the pretence of neutrality.
Grievance redress
CPGRAMS and its state equivalents receive millions of complaints. Routing is slow, closure is often nominal, and the citizen frequently cannot tell whether anything happened.
- Routing to the correct office on first receipt, which is the largest single source of delay.
- Detecting nominal closures — cases marked resolved with boilerplate that does not address the complaint. This is a quality-audit problem a model can do at a scale no human team can.
- Clustering complaints to surface the systemic fault behind a thousand individual ones.
- Answering the citizen in the language they wrote in.
How it goes wrong. The temptation is to auto-close, because closure rate is the metric officials are measured on. Optimising a grievance system for closure rather than resolution is already a problem; automating it would industrialise the problem. Any deployment here needs the quality audit built before the routing, not after.
Public procurement
Enormous spend, dispersed across thousands of authorities, with collusion and specification-rigging hard to detect case by case.
- Bid-pattern analysis to flag likely cartels — rotation, cover bidding, suspiciously consistent margins. This is a well-understood detection problem with a real literature behind it.
- Specification analysis to catch tenders written around a single supplier.
- Price benchmarking across authorities buying the same thing at different prices.
- Contract-performance monitoring against milestones.
How it goes wrong. A flag is a hypothesis, not a finding. Treating model output as evidence of wrongdoing would be both unjust and, in practice, self-defeating — the first wrongly-accused contractor wins in court and the programme loses its mandate. Publishing the aggregate findings, however, is where the transparency dividend actually sits.
Cutting across all of it
Language is the whole game
India's government operates in twenty-two scheduled languages and far more in practice. Almost every application above fails for most citizens if it works only in English and Hindi. This is also where the opportunity is largest: a citizen who currently cannot read the form is not marginally inconvenienced, they are excluded. Language coverage is not a feature of these systems; it is the difference between a productivity tool for officials and a service for the public.
The comparator is not perfection
Most objections to AI in government implicitly compare it to an ideal administration. The honest comparator is the current one: a screening programme that does not exist, an advisory service that reaches one farmer in ten, a grievance that is closed without being read. A system that is worse than an expert but better than nothing, deployed where nothing is the status quo, is a gain — provided it is honest about which it is.
Audit trails are the transparency dividend
The largest transparency gain from these systems is not the model. It is that automating a process requires first making it explicit — what was decided, on what basis, by whom, when. Most Indian administrative processes do not currently produce that record. A deployment that logs decisions and reasons, and publishes the aggregates, delivers accountability whether or not the model is any good.
What should not be automated
A short list, and it should stay short: adjudication of rights, selection of individuals for coercive attention, and any write to a record of entitlement or title without human confirmation. These are not technical limits — they will become technically feasible. They are limits on what a state should delegate, because the thing being delegated is accountability, and accountability cannot be held by a system that cannot be summoned, questioned or dismissed.
Some of what is being managed
Judicial checks on government
View as table
| Period | Judicial checks on government |
|---|---|
| 2025 | 0.667 |
| 2024 | 0.71 |
| 2023 | 0.724 |
| 2022 | 0.764 |
| 2021 | 0.761 |
| 2020 | 0.753 |
| 2019 | 0.766 |
| 2018 | 0.777 |
| 2017 | 0.777 |
| 2016 | 0.8 |
| 2015 | 0.799 |
| 2014 | 0.818 |
| 2013 | 0.827 |
| 2012 | 0.822 |
| 2011 | 0.822 |
| 2010 | 0.822 |
| 2009 | 0.819 |
| 2008 | 0.819 |
| 2007 | 0.819 |
| 2006 | 0.819 |
| 2005 | 0.819 |
| 2004 | 0.819 |
| 2003 | 0.819 |
| 2002 | 0.819 |
| 2001 | 0.819 |
| 2000 | 0.819 |
| 1999 | 0.819 |
| 1998 | 0.819 |
| 1997 | 0.819 |
| 1996 | 0.819 |
| 1995 | 0.802 |
| 1994 | 0.802 |
| 1993 | 0.802 |
| 1992 | 0.802 |
| 1991 | 0.802 |
| 1990 | 0.802 |
| 1989 | 0.802 |
| 1988 | 0.802 |
| 1987 | 0.802 |
| 1986 | 0.802 |
| 1985 | 0.802 |
| 1984 | 0.802 |
| 1983 | 0.802 |
| 1982 | 0.802 |
| 1981 | 0.802 |
| 1980 | 0.802 |
| 1979 | 0.802 |
| 1978 | 0.802 |
| 1977 | 0.623 |
| 1976 | 0.596 |
| 1975 | 0.609 |
| 1974 | 0.695 |
| 1973 | 0.695 |
| 1972 | 0.695 |
| 1971 | 0.695 |
| 1970 | 0.695 |
| 1969 | 0.756 |
| 1968 | 0.756 |
| 1967 | 0.756 |
| 1966 | 0.756 |
| 1965 | 0.756 |
| 1964 | 0.796 |
| 1963 | 0.796 |
| 1962 | 0.796 |
| 1961 | 0.796 |
| 1960 | 0.796 |
| 1959 | 0.796 |
| 1958 | 0.796 |
| 1957 | 0.796 |
| 1956 | 0.796 |
| 1955 | 0.796 |
| 1954 | 0.796 |
| 1953 | 0.796 |
| 1952 | 0.796 |
| 1951 | 0.796 |
| 1950 | 0.796 |
| 1949 | 0.789 |
| 1948 | 0.789 |
| 1947 | 0.789 |
| 1946 | 0.789 |
| 1945 | 0.789 |
| 1944 | 0.789 |
| 1943 | 0.789 |
| 1942 | 0.789 |
| 1941 | 0.789 |
| 1940 | 0.789 |
| 1939 | 0.789 |
| 1938 | 0.789 |
| 1937 | 0.789 |
| 1936 | 0.789 |
| 1935 | 0.789 |
| 1934 | 0.655 |
| 1933 | 0.655 |
| 1932 | 0.655 |
| 1931 | 0.655 |
| 1930 | 0.655 |
| 1929 | 0.655 |
| 1928 | 0.655 |
| 1927 | 0.655 |
| 1926 | 0.655 |
| 1925 | 0.655 |
| 1924 | 0.655 |
| 1923 | 0.655 |
| 1922 | 0.655 |
| 1921 | 0.655 |
| 1920 | 0.655 |
| 1919 | 0.655 |
| 1918 | 0.655 |
| 1917 | 0.655 |
| 1916 | 0.655 |
| 1915 | 0.655 |
| 1914 | 0.655 |
| 1913 | 0.655 |
| 1912 | 0.655 |
| 1911 | 0.655 |
| 1910 | 0.655 |
| 1909 | 0.64 |
| 1908 | 0.64 |
| 1907 | 0.64 |
| 1906 | 0.64 |
| 1905 | 0.64 |
| 1904 | 0.64 |
| 1903 | 0.64 |
| 1902 | 0.64 |
| 1901 | 0.64 |
| 1900 | 0.64 |
| 1899 | 0.695 |
| 1898 | 0.695 |
| 1897 | 0.695 |
| 1896 | 0.695 |
| 1895 | 0.695 |
| 1894 | 0.695 |
| 1893 | 0.695 |
| 1892 | 0.695 |
| 1891 | 0.695 |
| 1890 | 0.695 |
| 1889 | 0.695 |
| 1888 | 0.695 |
| 1887 | 0.695 |
| 1886 | 0.695 |
| 1885 | 0.695 |
| 1884 | 0.695 |
| 1883 | 0.695 |
| 1882 | 0.695 |
| 1881 | 0.695 |
| 1880 | 0.695 |
| 1879 | 0.695 |
| 1878 | 0.695 |
| 1877 | 0.695 |
| 1876 | 0.695 |
| 1875 | 0.695 |
| 1874 | 0.695 |
| 1873 | 0.695 |
| 1872 | 0.695 |
| 1871 | 0.695 |
| 1870 | 0.695 |
| 1869 | 0.695 |
| 1868 | 0.695 |
| 1867 | 0.695 |
| 1866 | 0.695 |
| 1865 | 0.695 |
| 1864 | 0.695 |
| 1863 | 0.695 |
| 1862 | 0.695 |
| 1861 | 0.695 |
| 1860 | 0.695 |
| 1859 | 0.695 |
| 1858 | 0.695 |
| 1857 | 0.695 |
| 1856 | 0.695 |
| 1855 | 0.695 |
| 1854 | 0.695 |
| 1853 | 0.695 |
| 1852 | 0.695 |
| 1851 | 0.695 |
| 1850 | 0.695 |
| 1849 | 0.695 |
| 1848 | 0.695 |
| 1847 | 0.695 |
| 1846 | 0.695 |
| 1845 | 0.695 |
| 1844 | 0.695 |
| 1843 | 0.695 |
| 1842 | 0.695 |
| 1841 | 0.695 |
| 1840 | 0.695 |
| 1839 | 0.695 |
| 1838 | 0.695 |
| 1837 | 0.695 |
| 1836 | 0.695 |
| 1835 | 0.695 |
| 1834 | 0.695 |
| 1833 | 0.695 |
| 1832 | 0.695 |
| 1831 | 0.695 |
| 1830 | 0.695 |
| 1829 | 0.695 |
| 1828 | 0.695 |
| 1827 | 0.695 |
| 1826 | 0.695 |
| 1825 | 0.695 |
| 1824 | 0.695 |
| 1823 | 0.695 |
| 1822 | 0.695 |
| 1821 | 0.695 |
| 1820 | 0.695 |
| 1819 | 0.695 |
| 1818 | 0.695 |
| 1817 | 0.695 |
| 1816 | 0.695 |
| 1815 | 0.695 |
| 1814 | 0.695 |
| 1813 | 0.695 |
| 1812 | 0.695 |
| 1811 | 0.695 |
| 1810 | 0.695 |
| 1809 | 0.695 |
| 1808 | 0.695 |
| 1807 | 0.695 |
| 1806 | 0.695 |
| 1805 | 0.695 |
| 1804 | 0.695 |
| 1803 | 0.695 |
| 1802 | 0.695 |
| 1801 | 0.695 |
| 1800 | 0.695 |
| 1799 | 0.695 |
| 1798 | 0.695 |
| 1797 | 0.695 |
| 1796 | 0.695 |
| 1795 | 0.695 |
| 1794 | 0.695 |
| 1793 | 0.695 |
| 1792 | 0.695 |
| 1791 | 0.695 |
| 1790 | 0.695 |
| 1789 | 0.695 |
Road accident deaths
View as table
| Period | Road accident deaths |
|---|---|
| 2025 | 1,83,434 |
| 2023 | 1,72,890 |
| 2022 | 1,68,491 |
| 2021 | 1,53,972 |
| 2020 | 1,38,383 |
| 2019 | 1,58,984 |
| 2018 | 1,51,417 |
| 2015 | 1,46,133 |
Control of corruption
View as table
| Period | Control of corruption |
|---|---|
| 2024 | -0.295 |
| 2023 | -0.335 |
| 2022 | -0.292 |
| 2021 | -0.301 |
| 2020 | -0.254 |
| 2019 | -0.218 |
| 2018 | -0.208 |
| 2017 | -0.224 |
| 2016 | -0.285 |
| 2015 | -0.345 |
| 2014 | -0.414 |
| 2013 | -0.471 |
| 2012 | -0.503 |
| 2011 | -0.491 |
| 2010 | -0.362 |
| 2009 | -0.345 |
| 2008 | -0.295 |
| 2007 | -0.327 |
| 2006 | -0.23 |
| 2005 | -0.409 |
| 2004 | -0.388 |
| 2003 | -0.548 |
| 2002 | -0.556 |
| 2000 | -0.211 |
| 1998 | -0.155 |
| 1996 | -0.264 |
The one-sentence version
The best uses of AI in Indian government are unglamorous throughput problems — transcription, translation, triage, detection — in departments where the current alternative is a queue; and the worst are anything that decides a person’s rights, which should remain with someone who can be held answerable for getting it wrong.