Section 1 Comments on institutional architecture
A. Issues with the proposed design
(i) Fragmentation of responsibility and overlapping mandates
The Draft Regulations establish an Apex Body, five standing committees, an AI Committee in every High Court, an AI Secretariat at every High Court and District/Subordinate Court, and the Centre of Research and Excellence on AI (CoRE-AI). Together these perform functions ranging from policy formulation and standard-setting to procurement, technology evaluation, research, cybersecurity, audits, infrastructure planning and implementation. The breadth reflects the importance the draft rightly places on responsible adoption, but it raises a governance question:
How should these functions and responsibilities relate to the judiciary’s existing institutions for technology and reform?
Technology-enabled reform is not a standalone activity. AI depends on digital records, case-management systems, cybersecurity, procurement, budgeting, judicial administration and continuous training, functions already undertaken, at least in part, through existing institutions such as the e-Committee, High Court Computer Committees and Committees for Digitisation. As AI becomes embedded in court processes, these functions inevitably intersect: an AI-assisted scheduling system is at once a case-management project, a digitisation project, an infrastructure project and an AI project. This creates two risks:
- Fragmented accountability: it becomes harder to identify which institution is ultimately responsible for initiating, planning and delivering a reform.
- Overlapping mandates: decision-making slows where multiple committees must each examine related aspects of the same project before implementation can begin.
(ii) Limited institutional continuity
Technology systems require continuous evaluation, maintenance, upgrades, cybersecurity review, vendor management and periodic replacement. Committees are, by design, deliberative bodies rather than permanent delivery organisations: composition changes, members rotate, and institutional knowledge is not always retained. Organisations responsible for continuous execution perform better when constituted as permanent entities: operational autonomy enables quicker decisions, long tenure builds institutional memory, and clear ownership makes success and failure easier to attribute. Three Indian examples illustrate the principle:
TNMSCThe Tamil Nadu Medical Services Corporation was set up as a specialised statutory corporation to professionalise procurement and supply-chain management of medicines, and became a model for several other states.
UIDAIThe Unique Identification Authority of India was created as a permanent institution to design, deploy, maintain and evolve the Aadhaar ecosystem, including standards, cybersecurity and vendor management.
NHAIThe National Highways Authority provides dedicated capability to plan, procure, implement and maintain the highway network, with execution kept separate from policy, which stays with the Ministry.
AI adoption would similarly benefit from a permanent professional organisation that builds technical expertise over time, with enough delegated authority to make timely operational decisions and own their outcomes.
(iii) Sub-optimal use of judicial time for operational functions
The draft envisages judges leading the Apex Body and AI Committees and overseeing approvals, audits, procurement, policy formulation, evaluation and implementation. Judicial leadership over questions of independence, fairness, due process and appropriate use is both necessary and appropriate. But not every function needs the same degree of judicial attention. At present, operational responsibilities are combined with governance and approval responsibilities in the same body. For example, Regulation 46 makes the AI Committee the final approval authority even though approval involves operational work such as evaluating systems and running procurement; the mandates in Regulations 26–30 (developing and deploying tools, monitoring research, building infrastructure) do not require judges’ attention; and Regulation 50’s living repository of best practice need not consume judicial time. Judicial time is among the judiciary’s scarcest resources: responsibilities should follow comparative expertise, with judges focused on adjudication, interpretation and policy, and permanent technical teams handling evaluation, procurement, cybersecurity and implementation under judicial supervision.
(iv) Transaction costs for innovation
The draft rightly prioritises innovation over restraint, but the proposed approval structure may raise the cost of participating in the judicial technology market, especially for smaller and newer entrants. If approvals, evaluations and pilots run independently across 25 High Courts and multiple tribunals, vendors face repeated evaluation of substantially similar technologies, and the judiciary expends significant effort duplicating work, unintentionally favouring well-established vendors. Many regulated sectors address this through nationally recognised qualification frameworks, empanelment or common technical standards that permit decentralised procurement while avoiding repetitive evaluation. Comparable approaches for judicial AI could lower transaction costs, encourage innovation and promote competition while preserving each court’s autonomy over procurement.
B. Recommendation: a specialised organisation for judicial administration
These concerns point to a clear need: a permanent professional organisation able to plan, execute and continuously manage technology-enabled reform, while judicial leadership retains policy, standards and strategic direction. This distinction between governance and execution is well recognised across public and private institutions: governing bodies set policy, approve priorities and exercise oversight, while permanent executive organisations translate those decisions into operational outcomes.
We respectfully recommend establishing a permanent Indian Courts Service: a specialised administrative and technical organisation dedicated to judicial administration and technology management. To preserve the federal character of the judiciary, its services could be shared across High Courts and tribunals, while individual courts retain the flexibility to establish their own where scale or local needs warrant. This is consistent with the National Judicial Technology Council envisaged under the eCourts Phase-III mission document, and requires not a substantial redistribution of functions but a reallocation of existing ones according to their character. The structure is guided by three principles:
- Optimal use of judicial time: policy, governance, approval of AI systems and strategic priorities stay under judicial leadership, while operational, technical and administrative work moves to permanent professional institutions.
- Functional separation: decision-making, implementation and research require different capabilities and need not sit in one body, and separating them improves accountability and lets each develop specialised expertise.
- Separation of powers: those who set standards should not procure, those who procure should not audit, and those who audit should be independent of both.
On these principles, the architecture could comprise three distinct bodies:
SC & HC AI Committee
Primary governing body for AI in the judiciary. Owns decisions on what systems are deployed.
Ensures adoption is safe and consistent with constitutional principles and the objectives of the judiciary, by providing strategic direction, formulating policy, approving AI systems, and exercising oversight over the two bodies below.
The Indian Courts Service
The judiciary’s permanent execution and implementation arm.
Provides the professional and technical capability to plan, procure, evaluate, deploy, maintain and continuously improve AI systems: maintaining an empanelled list of qualified vendors (as GeM does) while leaving procurement with each High Court; running the controlled testing environment and a permanent innovation platform for researchers and start-ups (as SEBI’s sandbox and NPCI’s shared infrastructure do); grounding planning in empirical data; and acting as first point of contact for AI-related complaints, preparing technical reports for the appropriate court to adjudicate.
CoRE-AI
The judiciary’s neutral body for research, innovation and standard-setting.
The intellectual backbone of judicial AI governance: developing and maintaining the independent standards, benchmarks and evidence base against which all court AI systems are evaluated and audited, and fostering continuous learning.
Section 2 Comments on standards for AI systems and vendors
Regulation 35 requires a Technical and Ethical Impact Assessment before any AI system is approved, and lists the criteria it must cover: purpose and architecture, training-data quality, risks of bias and hallucination, cybersecurity, explainability and incident reporting. But it does not provide the standard against which the answers are judged. Knowing a system must be assessed for bias is necessary but not sufficient: the framework must also specify what level of bias is acceptable, for which functions, and how it is measured. Without this, the assessment becomes a documentation exercise rather than a genuine qualification threshold, and the same gap applies across every criterion, leaving each High Court to arrive at inconsistent answers.
We rely on minimum standards to govern almost every system that affects public safety and rights. They exist not to constrain innovation, but to ensure it serves people reliably and consistently.
- ISO 9001: before a drug can be sold in India, its entire manufacturing process must meet defined quality standards; the product is not simply tested once, the process must consistently conform.
- BIS certification: before an electrical appliance is sold, it must carry a Bureau of Indian Standards mark defining what it must do and how it must perform under stress.
- Building codes: before a commercial unit is built it must meet structural, accessibility and fire-safety standards; the developer does not decide what “safe” means, the standard does.
- PDF/A: courts already require judgments and filings in PDF/A, a technical standard for long-term preservation that defines what the file must be capable of over time.
Standards serve two functions: they let courts make objective, evidence-based approval decisions against a minimum benchmark; and they signal to the market the capabilities the judiciary wants to encourage, giving developers the clarity and predictability to design, test and invest. Articulating minimum specifications and evaluation criteria is therefore not merely desirable; it is foundational to the integrity of the approval mechanism itself.
What must be specified?
Some of the critical standards that need to be defined for judicial AI:
- Explainability: systems must give reasons and evidence for every output, understandable to the user receiving them, and reflecting the system’s actual decision process rather than a post-hoc justification.
- Accuracy and error thresholds: the maximum acceptable error rate for each category of judicial function, the methodology for testing it, and the consequences (including suspension) when a system falls below the threshold.
- Data governance and provenance: training data documented, lawfully sourced and representative of the diversity of courts and litigants served, with clear records of collection and updates.
- Human oversight and override: for each function, the point at which a human must review or approve an output, with a guarantee that any AI-assisted output can be overridden without penalty or procedural barrier.
- Data and API standards: conformance to open data formats and API specifications so courts can migrate between providers without losing data or functionality, avoiding lock-in where approving one system becomes a permanent procurement decision.
This list is not exhaustive, and standards will evolve with the technology. Evaluation against them could be supported by common benchmark datasets for categories such as OCR, translation, anonymisation or case classification (shared reference datasets against which competing systems report standardised metrics), and by codifying administrative benchmarks, defining what good performance means for functions like scheduling or cause-list management before an AI system is asked to perform them.
Who should create these standards?
Setting standards and procuring technology must be kept institutionally separate. The body that evaluates and selects vendors cannot also define what those vendors must demonstrate; that is akin to asking a contractor to write the building code their own work must meet. The value of a standards body lies in its neutrality: no stake in which vendor wins, no procurement budget to defend, no implementation timeline to protect.
This is the model that made India’s payments revolution possible. NPCI did not build PhonePe or Google Pay; it wrote the specification and stepped back. The standard was shared; the market built everything else.
The Indian Courts Service proposed under Recommendation 1 is well placed to plan, procure and manage technology, but the standards against which it procures must be set by a separate, neutral institution. CoRE-AI is well suited to this: its role should stay focused on research, specifications and standards, with the Draft Regulations protecting its neutrality by ensuring it has no role in procurement or service delivery, and mandating it to develop and maintain the minimum specifications the approval mechanism currently lacks.
