India’s AI push needs better data first — and the numbers behind the warning are the government’s own
India has enrolled more than 1.4 billion people in Aadhaar, processes over 20 billion UPI transactions a month, and has issued and verified nearly 10 billion documents through DigiLocker. By almost any measure, it has built digital public infrastructure at a scale few countries can match.
A new report from the Tony Blair Institute for Global Change (TBI) argues that this scale is not the same as usable data — and that unless India fixes the data, its AI ambitions will rest on shaky ground. The report, “From Digital Scale to Data Power: A Data Operating Model for India’s States,” was published this week and carries a striking headline number: poor-quality data causes fiscal leakage equal to four to seven percent of India’s annual welfare spending.
It is worth being precise about where that number comes from, because it is not TBI’s own. The report’s case for India government data quality rests on that provenance. It is an estimate from NITI Aayog’s 2025 report “India’s Data Imperative: The Pivot Towards Quality.” The TBI report cites it. The same is true of the eye-catching savings examples it repeats. That provenance matters: this is a think tank amplifying the Indian government’s own warning, which is both a reason to take the number seriously and a reason to read it as an estimate rather than a measurement.
What the report argues
TBI’s central claim is that India’s next digital challenge is not building more infrastructure but making the data it already generates accurate, connected and usable across departments. It puts the problem in blunt terms: states produce enormous volumes of administrative data — beneficiary lists, health records, land registers, school enrolments, tax filings — but collect it “to satisfy one department’s reporting requirement, not to answer the cross-cutting questions that actually drive policy.”
Its proposed remedy is a “state data operating model” with seven elements: leadership and mandate; data ownership and stewardship; common standards; shared data capabilities; independent quality assurance; legal and procurement safeguards; and a “State Data Balance Sheet” to help leaders manage data assets and risks the way they manage finances. The report points to state initiatives — Karnataka’s Kutumba social registry, Odisha’s Social Protection Delivery Platform, Rajasthan’s Pehchan portal, and work in Andhra Pradesh, Uttar Pradesh, Tamil Nadu and Telangana — as examples worth generalising.
Where the 4–7% figure comes from
The savings figures the report repeats trace to NITI Aayog’s June 2025 paper. Removing 1.71 crore ineligible names from the PM-KISAN farmer support scheme saved an estimated ₹9,000 crore; eliminating 3.5 crore bogus LPG connections saved about ₹21,000 crore over two years; and dropping 1.6 crore fake ration cards is saving roughly ₹10,000 crore a year. NITI Aayog, the government’s own think tank, published those numbers — and the 4–7% leakage estimate — alongside its own tools for measuring data quality.
That is the strongest part of the story for a reader: the warning about India’s data is not an outside critique imposed on the government. It is the government’s own assessment, and it is specific about the money at stake.
Why bad data blocks AI
The report’s more consequential argument is about artificial intelligence. Deploying AI on fragmented, unreliable data, it says, will not fix public services — it will “amplify existing blind spots” faster. TBI’s India country director and report co-author, Vivek Agarwal, framed the question plainly: “Is the data good enough to trust an AI system, or a human official, to act on it? Right now, in most states, the honest answer is not yet.”
That is a technically sound point, and a familiar one in data engineering: a model is only as good as its inputs, and errors in records — a duplicated identity, a mismatched name, a stale address — propagate into automated decisions with the appearance of authority. For government, where a wrong decision can deny someone a pension or a ration card, the stakes are higher than in most commercial settings. The report’s conclusion is that data governance is an operational precondition for AI, not an administrative afterthought.
The government is already building the fix
The report lands inside an active government programme, which the release notes but does not emphasise. MoSPI, the Ministry of Statistics and Programme Implementation, has issued a National Metadata Structure (NMDS 2.0), a Statistical Quality Assessment Framework based on the UN’s quality framework, and an AI-readiness self-assessment for departments; it launched a Quarterly Progress Report portal on data harmonisation in September 2026. Separately, MeitY’s National Framework for Data Sharing, approved in May 2026, created a National Data Governance Committee and State Data Governance Committees in every state and union territory, with a roadmap of milestones stretching to 2028.
In other words, much of what the report recommends — common standards, quality assurance, custodians, data-sharing frameworks — is already official policy in some form. TBI’s contribution is to pull it into a single state-level operating model and to argue that it must come before AI, not alongside it. That is a useful framing, and it is also the reason to treat the report as advocacy informed by evidence rather than a discovery.
One design point is easy to miss: the report does not call for a single national data lake. Its aim is a federated model — departments keep ownership of their data, but common standards, identifiers and APIs let authorised systems discover, exchange and correct it. That is the same principle behind MeitY’s framework, which explicitly favours a “federated, security-conscious” approach over centralisation. It is a sensible architecture for a country of India’s size and federal structure; it is also harder to implement than one database, because it depends on every department adopting the same standards.
The counterpoint: leakage vs exclusion
There is a tension in the report that deserves more attention than it gets. Its framing emphasises leakage — money saved by removing ineligible records — and its examples are savings figures. But the report also names “exclusion from welfare schemes” as a consequence of bad data. Those two things pull in opposite directions, and the debate is not settled.
A substantial body of research argues that the “savings” attributed to Aadhaar-linked clean-ups are partly the result of eligible people being denied benefits they are legally entitled to. Reetika Khera’s work has argued that what passes as savings “is often the result of denial of legal entitlements.” A review of the evidence for NCAER concluded that biometric authentication reduced leakage in several settings but that its effect on beneficiaries “depends crucially on the protocols and details of implementation,” and that exclusion — though a lower share than non-Aadhaar-related exclusion — still affects millions and falls disproportionately on the most vulnerable.
None of this invalidates the report’s core argument; if anything it strengthens it. Data quality is not only a fiscal issue. A record that is wrong can leak money to someone who should not get it, or deny money to someone who should. A serious data-quality agenda has to measure both errors, not just the ones that show up as savings. The report’s State Data Balance Sheet idea — treating data as an asset with risks — is one place that both halves could track, but the report does not develop it in those terms.
Who is making the argument
Transparency matters here. TBI is not a neutral observer. It is a not-for-profit policy institute that works with governments in more than 40 countries, and it has recently launched a global initiative, CentreAI, aimed at how AI can transform government processes. A report urging states to make data AI-ready is, unsurprisingly, aligned with that agenda. That does not make it wrong — its factual basis is the Indian government’s own — but readers should weigh it as the argument of an interested party.
What to watch
For state governments and their data teams, the milestones to watch are the MoSPI harmonisation roadmap through 2028 and whether State Data Governance Committees move from constitution to practice. For AI-in-government units, the test is whether datasets are genuinely machine-readable and quality-assuring before models deploy on them. And, for policy watchers, watch whether the savings-versus-exclusion tension addresses in how success is measuring. And for anyone tracking India’s digital state, the honest question is the report’s own: is the data good enough to act on?
What the report does and does not establish
It establishes a well-evidenced argument, drawn from the government’s own assessment, that India’s data quality is a binding constraint on its AI ambitions, and it offers a concrete, state-level model for fixing it. It does not establish a new fact about India’s data, a measured figure for the true cost of poor records, or a reconciliation between the savings it cites and the exclusion it acknowledges.
The report’s value is that it says, clearly and with the government’s numbers, what is easy to skip: before India deploys AI across its public systems, it has to make sure the records those systems run on are true. Whether it also makes sure those records stop denying people what they are owed is the harder, and less-quantified, half of the same problem.
Editor’s Note: This article draws on the Tony Blair Institute‘s report and the government data it cites, including NITI Aayog’s June 2025 assessment, plus MoSPI and MeitY documents and independent academic research. TBI is a policy institute with an interest in AI-enabled government; that interest is disclosed above. No interview was conducted, and the underlying estimates have not been independently reproduced. See the source list in the publication package.