India's courts hold more than five crore pending cases: 4.77 crore in the district courts, 63.66 lakh in the High Courts and 92,101 in the Supreme Court, according to a Lok Sabha reply of 13 February 2026. Against that backlog the government has committed ₹7,210 crore to the third phase of the eCourts project, approved by the Union Cabinet on 13 September 2023, with the stated aim of digital, online and paperless courts and the explicit intention to use "AI and its subsets Machine Learning (ML)", optical character recognition and natural language processing for scheduling, registry work and data-based decision support (pib.gov.in, 13 September 2023). The question for anyone working on AI in judiciary India is no longer whether the courts will use AI. It is which parts of court work AI can safely take on now, which parts must wait, and what the eCourts infrastructure does and does not supply to a builder.
This article answers that question with a workflow map. It breaks the life of a case into its operational steps, from filing to judgment, and for each step asks three things: how much of the work is clerical versus adjudicative, whether the data the step needs is now available in digital form through eCourts, and what the consequence of an error would be. The combination determines deployability. Steps that are clerical, digitised and reversible are deployable now; steps that are adjudicative, undigitised or irreversible are not, whatever the technology can demonstrate.
The map is an analytical construction. The figures on eCourts, pendency and the Supreme Court's own AI tools are drawn from government releases opened for this article; the judgements about deployability are ours.
What eCourts Phase III actually funds
The Phase III approval is often summarised as an AI programme. It is not. The ₹7,210 crore is overwhelmingly infrastructure, and the component table in the Cabinet release makes the priorities plain.
Source: PIB release, 13 September 2023. Total across all 24 components is ₹7,210 crore.
Two things follow. First, the line that names AI directly, future technological advancements, is ₹53.57 crore, under 1% of the outlay. The AI opportunity in Phase III is not in that line; it is in the ₹2,038 crore digitisation line, which creates the data, and in the ₹243 crore software line, which creates the integration points. Second, the programme is decentralised: it is implemented through the High Courts under the Department of Justice and the Supreme Court's e- Committee. A builder does not sell to "the judiciary"; it works with 25 High Courts, each with its own registry practices, and with the Supreme Court separately.
Progress is substantial but incomplete. As on 20 July 2026, ₹3,210.94 crore had been released to the High Courts under Phase III and ₹2,277.27 crore spent, with further central spending on NIC technical support (₹186.08 crore), BSNL connectivity (₹54.79 crore) and the C-DAC Judicial Digital Preservation System (₹9.10 crore), according to a Lok Sabha reply of 24 July 2026 (static.pib.gov.in). Roughly a third of the outlay had been spent three years into a four-year programme. The data foundation is being laid; it is not finished.
What the earlier phases built matters as much. A PIB backgrounder of December 2024 records 18,735 district and subordinate courts computerised, over 27.64 crore orders and judgments available on the National Judicial Data Grid, 3.38 crore cases heard by video conference in the district courts and High Courts as of 31 October 2024, 1,394 e- Sewa Kendras in the district courts and 36 in the High Courts, and virtual courts in 21 states and union territories that had handled over 6 crore cases and collected more than ₹649.81 crore in online fines. The orders-and-judgments corpus on NJDG is a legal text resource of a scale that did not exist in India before the eCourts programme.
What the Supreme Court has already deployed
The Supreme Court's own experience is the best available evidence of what works and what does not, because it has run AI tools in production for several years and reports on them to Parliament.
Translation is the clearest success. As on 5 August 2024, 36,271 Supreme Court judgments had been translated into Hindi and 17,142 into 16 other regional languages using AI language technology, available on the e-SCR portal, according to a Lok Sabha reply reported by IndiaAI (indiaai.gov.in, 9 August 2024). By February 2026 the Court's AI translation tool, SUVAS, and related tools built with the National Informatics Centre were translating judgments into 18 Indian languages (static.pib.gov.in, 11 February 2026).
Transcription is the second. AI speech recognition has been used to transcribe oral arguments in Constitution Bench matters since February 2023, and the February 2026 PIB note records that it is "enabling near real-time transcription of oral arguments", with transcripts published on the Court's platforms and the system being extended to regular hearing days. The same note is careful to add that final validation remains human-led.
Research assistance is the instructive failure, or rather the instructive non-deployment. SUPACE, the Court's AI research tool, "remains in an experimental stage and is not yet deployed for regular judicial use" as of February 2026. The tool closest to the adjudicative core is the one that has not crossed into routine use.
That pattern, translation and transcription deployed, research assistance still experimental, is exactly what a deployability analysis would predict. It is the empirical anchor for the workflow map that follows.
A workflow map for AI in judiciary India
The map rates each stage of a case on three criteria. Nature of work runs from clerical (rule-bound, high-volume, checkable) to adjudicative (requires legal judgement). Data availability asks whether eCourts has made the inputs digital and structured. Error consequence asks whether a mistake is caught and reversed cheaply or becomes a legal fact. Deployability is the combination.
The "now" column is dominated by work that is clerical or linguistic, where the eCourts corpus already exists and where a human sits between the system and the record. The "later" column is where legal judgement enters. The distinction is not about model capability; current models can draft a plausible order. It is about where an error becomes a legal fact rather than a correctable artefact, and about whether the institution has an evaluation and accountability structure that would let it detect the error first.
Where the data is, and where it is not
A builder reading the map should notice that data availability is high in most rows, and should then ask what "available" means. The NJDG corpus of 27.64 crore orders and judgments is available for pendency analytics and for training or evaluating translation and retrieval systems. It is not uniformly available as clean, machine-readable text: much of the historical record is scanned, which is why ₹2,038 crore of Phase III is digitisation and why the Cabinet release names OCR as a priority technology. The practical sequence for a court AI deployment is therefore OCR and structuring first, then the workflow tool that consumes it.
Decentralisation compounds the data question. Case information systems are run by High Courts, registry practices differ, and a tool that works in one High Court's scrutiny process may need substantial reconfiguration in another. PanScience reports that its venture Nyaay AI builds judicial workflow automation and has deployments in 14 or more Indian High Courts and three national supreme courts; the company-reported figure is noted here as an indicator of what multi-court deployment of clerical-layer tools looks like in practice, not as an evaluation of the product. The relevant lesson from any such deployment is that the integration work per court, not the model, is the binding constraint.
Why the adjudicative layer should wait
The case for caution at the research and drafting stages is sometimes made as a matter of principle. It can also be made as a matter of engineering, and the engineering argument is more persuasive to institutions.
A legal research tool that retrieves precedent must be evaluated against a ground truth: for a given question, which authorities are correct and controlling. Building that evaluation set requires judicial expertise, is jurisdiction-specific and must be maintained as the law changes. Until it exists, the institution cannot know the tool's error rate, and a research tool with an unknown error rate is more dangerous than no tool, because it is persuasive. The Supreme Court's decision to keep SUPACE experimental rather than put it into regular judicial use is consistent with exactly this reasoning.
Judgment drafting raises a further issue. In a translation or transcription workflow the human validator checks the output against a source that exists independently: the original judgment, the recorded audio. In judgment drafting there is no independent source; the draft is the thing being made. Human-in-the-loop review is therefore weaker, because the reviewer is checking a draft against their own unwritten reasoning rather than against a document. That is why the map confines drafting to assistance, such as formatting, citation checking and summarising the record, and excludes generation of the reasoning itself.
None of this means the adjudicative layer is permanently closed. It means it is gated by evaluation infrastructure that does not yet exist at scale. Building that infrastructure, which requires judges, legal academics and engineers working together, is itself a legitimate programme of work. PanScience reports that its PALETTE Centre with BITS Law School, launched in May 2025 and inaugurated by former Chief Justice D.Y. Chandrachud, is directed at legal AI research of this kind; it is cited as one example of the institutional form such work takes.
What a builder should take from Phase III
Four implications follow from the map and the spending data.
The addressable work is clerical and linguistic, and it is large. Scrutiny, translation, transcription, process tracking and pendency analytics cover the majority of registry effort in a court, and all of them are rated deployable now. A builder that frames court AI as "AI judges" is addressing the wrong layer and will find no buyer; one that frames it as registry productivity, with the judge untouched, is addressing the layer the courts are already funding.
The data foundation is two-thirds unfinished. With roughly ₹2,277 crore of ₹7,210 crore spent by July 2026, digitisation is still in progress across the High Courts. Tools that depend on clean legacy records should be sequenced behind OCR and structuring work, which is itself a deployable AI service.
The proof pattern is set by the Supreme Court. Translation and transcription have been accepted because they have an independent source document, a human validator and a published output. New clerical-layer tools that follow the same pattern, with the human sign-off kept and the validation visible, will meet less institutional resistance than those that do not.
Evaluation is the unlock for the next layer. The gating factor for research and drafting tools is not model quality but the absence of court-specific evaluation sets and accountability structures. Investment there, by courts, law schools and builders together, is what moves rows from "later" to "now".
Pendency is the reason this matters. Five crore cases will not be cleared by faster typing. They may be reduced by registries that process filings in days rather than weeks, by judgments that reach litigants in their own language, and by data that tells court administrators where the delays actually are. Those are the problems AI in the Indian judiciary can address now, and Phase III has, mostly without naming it, funded the foundations for doing so.
Methodology and sources
The court workflow map and the three-criterion deployability rating are an original analytical construction by PanScience Research. Ratings reflect our reading of the nature of each workflow stage, the state of digitisation under eCourts and the reversibility of errors; they are judgements, not measurements, and are offered as a framework for discussion.
Observed figures and their sources: the eCourts Phase III approval, outlay, component allocations and the statements on AI, ML, OCR and NLP are from the PIB release of 13 September 2023 (https://pib.gov.in/PressReleasePage.aspx? PRID=1956919). Courts computerised, NJDG orders and judgments, virtual hearings, e-Sewa Kendras, virtual courts and fine collections are from the PIB eCourts backgrounder of December 2024 (https://static.pib.gov.in/WriteReadData/specificdocs/documents/2024/dec/doc20241217473201.pdf). Funds released and spent under Phase III as on 20 July 2026 are from the Lok Sabha unstarred question reply of 24 July 2026 published by PIB (https://static.pib.gov.in/WriteReadData/specificdocs/documents/2026/jul/doc2026726932901.pdf). The Supreme Court translation counts as on 5 August 2024 are from IndiaAI's report of a Lok Sabha reply (https://indiaai.gov.in/news/supreme-court-of-india-harnesses-ai-for-judicial-translation-and-research). The status of SUVAS, SUPACE and AI transcription in 2026 is from the PIB note of 11 February 2026 (https://static.pib.gov.in/WriteReadData/specificdocs/documents/2026/feb/doc2026211785401.pdf). Pendency figures are from the Lok Sabha reply of 13 February 2026 as reported by IANS. Figures on Nyaay AI and the PALETTE Centre are company-reported by PanScience Innovations (https://panscience.xyz); no client of any PanScience venture is named or implied, and no court is identified as a Nyaay AI deployment.