TL;DR: Adalat AI is a nonprofit speech-to-text tool that transcribes Indian court proceedings as they happen, in English and several Indian languages, and drafts the deposition record automatically. By the end of 2025 it ran in roughly 4,000 courtrooms, about one-fifth of India’s district courts, with formal partnerships across nine states and five more in talks, as its co-founder told the Carnegie Global Technology Summit. Kerala became the first state to make it compulsory in every trial court, from 1 November 2025. Courts using it report recording six to eight witness statements a day against two earlier. The promise is real, and so are the risks: word accuracy in Indian accents and code-mixed speech, what happens to sensitive testimony stored in the cloud, and whether an AI transcript can ever be the official “court of record.” This is what the tool does, how far it has spread, where it strains, and what it changes for you.

On this page

What Adalat AI actually does

Strip away the branding and Adalat AI is a courtroom microphone wired to a speech recognition model trained on Indian legal speech. As a witness deposes, or a judge dictates, the words appear on screen as text within seconds. The system handles Indian accents, dialects, and legal terminology, and it produces the running transcript that becomes the deposition record. According to the Stanford Social Innovation Review profile of the nonprofit, it compresses the documentation of a single answer from “5 to 10 minutes” of dictation and re-typing down to seconds.

The transcription is the headline feature, but the tool is wider than that. Its own site describes a stack that adds case-flow management, document automation through OCR, litigant chatbots, and judicial research and summarisation. In plain terms, it tries to take over the clerical labour around a hearing, not just the typing.

It is run by a nonprofit, not a commercial vendor, which matters for how courts have been willing to adopt it. The organisation was founded in 2023 and launched its India work in early 2024. Its co-founder and chief executive Utkarsh Saxena is a Harvard Law graduate and Supreme Court lawyer; the co-founder and chief technology officer, Arghya Bhattacharya, is an AI engineer. The Stanford profile records modest annual operating costs in the range of roughly 240,000 to 360,000 US dollars, funded by philanthropic backers including MIT Solve, Echoing Green, Fast Forward, and The/Nudge Institute. This is closer to public-interest infrastructure than to a SaaS contract, which is part of why High Courts have let it inside the courtroom so quickly.

One distinction is worth fixing early. Adalat AI is a transcription and workflow tool. It does not decide anything. It does not draft judgments, score the credibility of a witness, or predict an outcome. That line, tool versus decision-maker, runs through everything that follows, and it is the same line the courts are drawing in their own AI rules for India’s courts.

The bottleneck it targets: pendency and the stenographer gap

India’s case backlog is the reason any of this exists. As on 31 December 2025, district and subordinate courts alone had 4,76,57,328 cases pending, around 4.76 crore, according to the National Judicial Data Grid maintained by the Department of Justice. Across all courts the figure is above five crore. Uttar Pradesh by itself carried over 1.13 crore pending matters in its district courts. Pendency in the subordinate judiciary has grown, not shrunk, rising by close to six percent over three years.

Backlog has many causes, and clerical drag is one of the more fixable ones. “Court delays is a complex multi-dimensional problem,” Saxena said at the Carnegie summit, as reported by The Tribune. “There are many reasons why it happens, but one big contributing factor is the fact that we have a lot of manual and clerical processes, paper-based processes that slow the system down.”

The sharpest version of that bottleneck is the stenographer shortage. Recording evidence in a trial court has traditionally meant a judge dictating each answer while a stenographer or typist captures it, or a clerk taking it down by hand. The trouble is supply. ThePrint’s ground report found that fewer than 10 percent of Indian courtrooms have skilled stenographers. When a court has no steno, depositions crawl, witnesses are sent home, and dates get burned on the mechanics of writing things down rather than on the dispute itself.

That is the gap the tool fills. Courts running it report jumping from about two witness statements a day to six or eight, the kind of throughput change that moves a trial from a year of dates to a few months. Karnataka High Court judge N.G. Dinesh put the mechanism plainly to ThePrint: “Transcripts are generated in real time as the judge speaks.” The constraint shifts from how fast someone can type to how fast the court can actually hear evidence.

How big the rollout really is

The numbers are the part most worth getting right, because hype inflates them fast. Here is what is actually on the record as of late 2025 to mid-2026.

ClaimFigureSource
Courtrooms using the toolAbout 4,000 (around one-fifth of district courts)The Tribune / ANI
Formal state partnershipsNineThe Tribune / ANI
States in advanced talksFive moreThe Tribune / ANI
Witness statements per day6 to 8, up from 2ThePrint
Demonstrated accuracy95% in training sessionsThePrint

So the honest headline is roughly 4,000 courtrooms across nine partner states, which is meaningful but not yet national. One-fifth of district courts is a serious footprint for a tool that did not exist in courts before 2024, and the stated ambition is to keep climbing. The deployment has also crossed borders: the Stanford profile records a Ghana rollout in November 2024 and pilots in Nigeria, Kenya, and Bangladesh, which tells you the underlying problem, courts without enough typists, is not uniquely Indian.

Treat the efficiency claims with the same care. The “30 to 50 percent reduction in case timelines” and “two to three times judicial output” figures come from the tool’s own reporting and early adopter accounts, not from an independent audit. The throughput jump in evidence recording is well attested by judges on the ground. The downstream claim that whole cases finish that much faster is plausible but harder to verify, because a trial speeds up only if every other stage keeps pace. Believe the steno-replacement gain. Hold the end-to-end timeline claims a little looser.

Kerala goes first: the mandate that set the template

Most states have adopted the tool as an option. Kerala did something different: it made it compulsory. That makes Kerala the test case everyone else is watching.

The sequence is precise. The pilot began on 1 February 2025 in select courts in Ernakulam. On 27 September 2025 the Kerala High Court issued the directive, and from 1 November 2025 the tool became mandatory for recording witness statements and evidence across all district courts in the state, as reported by Elets eGov. Kerala is the first state to mandate such a system across every trial court. The version deployed there runs on a dedicated legal speech model in Malayalam, built after roughly nine months of testing across nine states.

The High Court’s memorandum did not just switch the tool on. It described the workflow: a real-time transcript, a witness review screen inside the testimony box so the deponent can see and confirm what was recorded, digital signing, and cloud storage that registered advocates can access. The recorded justification is the familiar one. Depositions used to rely on handwritten notes or dictation to a typist, which, with limited staff, “often” led to adjournments and delays.

Kerala matters beyond Kerala for two reasons. First, a mandate removes the comfortable middle ground where a judge can simply decline to use a system that misbehaves; reliability now has to be good enough to be the only option. Second, the same court paired the mandate with a governance policy, which is the part other states will copy or ignore at their peril. More on that below.

How the courtroom changes, step by step

It helps to see the before and after side by side, because the change is concrete rather than abstract.

In the depositionOld way (steno or dictation)With real-time transcription
Capturing the answerJudge dictates, typist or steno records✓ Spoken words appear as text in seconds
Throughput per dayOften 2 witnesses✓ 6 to 8 witnesses
Dependence on scarce staff✗ Needs a skilled stenographer✓ Runs without one
Witness sees the record✗ Rarely, before signing✓ Review screen in the witness box
Turnaround for a copy✗ Days, sometimes weeks✓ Same-day digital file
Risk of transcription errorHuman mishearing, fatigueMachine mishearing on accent or overlap

The workflow is not “press record and walk away.” A judge still controls the proceeding, corrects the text, and owns the final record. The witness gets a chance to see the transcript and confirm it, which is a genuine procedural improvement over a system where many deponents signed a deposition they never read. The signed, verified transcript is then stored digitally and pulled up far faster than a hand-written or typed record ever could be, which feeds straight into how quickly you can later get a certified copy of the record.

The recording of evidence and statements is governed by statute, and that statute changed recently. The procedural rules for recording witness testimony and statements now live in the new criminal codes, the BNSS and the Bharatiya Sakshya Adhiniyam, which also push audio-video recording of certain steps. A transcription tool sits comfortably inside that statutory shift toward recorded, retrievable evidence, but it does not replace the legal requirements for how that evidence is taken, read over, and signed.

The accuracy and language problem nobody should gloss over

The 95 percent accuracy figure is from training sessions, and a courtroom is not a training session. Real depositions have crosstalk, a witness who mumbles, a judge who interrupts, code-mixed Hindi-English-Malayalam in a single sentence, dialect, technical jargon, and proper nouns the model has never seen. Every one of those degrades a speech model.

This is not theoretical. The most instructive failure is the Supreme Court’s own, covered in the next section, where overlapping voices repeatedly broke the transcription. The lesson generalises. A demo number measured on clean audio overstates what you get when a frightened witness in a property dispute answers a leading question while the opposing counsel objects over them.

Language is the harder half of the problem. India runs courts in many languages, and a model that works in Malayalam has to be rebuilt, not merely translated, for Tamil, Bengali, or Marathi. Building a legal speech model per language is exactly why Kerala’s deployment needed a dedicated Malayalam model and months of tuning. The tool reportedly relies on native language experts to extend coverage, which is honest engineering but also a reminder that “supports Indian languages” is a roadmap, not a finished fact, and quality will vary sharply by language and dialect.

There is a deeper accuracy concern that has nothing to do with typos. The Chief Justice of India has publicly acknowledged that AI systems can “amplify discrimination” when trained on data that reflects existing social inequality by caste, gender, class, or religion, a point captured in Global Voices’ survey of AI tools entering India’s courts. For pure transcription, the worry is narrower than for a tool that recommends precedents, but it is not zero: a model that hears some accents worse than others records some witnesses worse than others. Whose words get captured cleanly is itself a fairness question.

The Supreme Court’s own transcription experiment

India’s apex court tried live transcription before the district courts did, and its experience is the most useful warning we have.

On 21 February 2023, the then Chief Justice D.Y. Chandrachud announced that live transcripts of Constitution Bench hearings would be published with the help of AI, on an experimental basis. The court used software called Teres, built by Bangalore-based Nomology Technology, starting in Courtroom 1 during the Maharashtra MLA disqualification hearings, as reported by LiveLaw and carried by the government’s own IndiaAI portal. The ambition was significant. Justice P.S. Narasimha framed it as turning the court into a true “court of record,” with arguments “accessible for all times to come.”

Then deployment met reality. A Supreme Court Observer review titled, pointedly, a “failed experiment” found that transcripts were published for only 36 percent of the Constitution Bench hearings held since launch. February managed three of four hearings. March produced transcripts for none of its eight hearings. The Maharashtra case itself got transcripts for three of nine sittings. The review attributed the gap to the slow human work of vetting and proofreading, unclear courtroom audio, and overlapping dialogue that the model could not separate.

Read the 36 percent figure carefully, because it is easy to misuse. It measures how often a usable transcript was actually published, not the word-level accuracy of the lines that were. But that is exactly the point for a litigant. A transcript you cannot rely on being produced, on time and complete, is not yet infrastructure. The district-court rollout is a different product in a different setting, with verified review screens and judge control, but it inherits the same hard constraint: the distance between a working demo and a dependable official record is where these projects live or die.

Data privacy and the open-court question

A microphone that captures every word of every witness creates a data trail that did not exist before, and the law on what happens to it is still catching up.

Start with storage. Adalat AI says it uses full encryption with user-level security keys, per the Stanford profile, and Kerala’s setup stores transcripts in the cloud for access by registered advocates. The unresolved question is governance: who can pull a deposition, for how long it is retained, where the servers sit, and what happens to recordings of in-camera matters such as sexual offences, matrimonial disputes, or proceedings involving children. The Digital Personal Data Protection Act, 2023 applies to automated processing of personal data, and a verbatim deposition is dense with personal data. How that statute maps onto judicial records is not yet settled by any clear, court-specific rulebook.

To its credit, the Kerala High Court anticipated part of this. Its 2025 AI policy treats AI strictly as an administrative tool, prohibits generative AI from drafting judgments or predicting outcomes, requires judges to rigorously evaluate any AI output, and, importantly, bars uploading confidential judicial data to external public platforms, with vetted fallback platforms allowed only if the primary system fails. Those guardrails, summarised in the Global Voices analysis, are sensible. They are also one High Court’s policy, not a national standard, and the rest of the country has no uniform rule.

Then there is the open-court principle, which cuts the other way. Indian courts are presumptively public, and an accurate, searchable transcript is a gain for transparency: it lets the press, researchers, and parties see exactly what was said, rather than relying on memory or a clerk’s paraphrase. The tension is that the same searchability that aids transparency aids surveillance. A permanent, indexed record of every witness’s words is a powerful thing in the wrong hands, and the safeguards around it are thinner than the technology. Both things are true at once, and pretending otherwise is how good tools get misused.

What it changes for litigants and lawyers

For a litigant, the most tangible change is speed and a fairer record. If your matter is in a court running real-time transcription, evidence recording should move faster, which means fewer wasted dates and, in principle, a quicker trial. The witness review screen is a real protection: your deposition is shown back to you before you sign it, so the “I never said that” problem, where a signed record drifts from what was actually spoken, shrinks. And a same-day digital transcript is far easier to obtain and check than a typed record you wait weeks for.

For lawyers, the shift is subtler and demands a new habit. The transcript is now a machine-produced first draft of the official record, so reading it against what you heard becomes part of the job, not an afterthought. A misheard “not guilty” as “guilty,” a mangled date, a witness’s qualification dropped: these are the errors a speech model makes, and they live in the document your appeal will rest on. Verifying the transcript at the time of recording, while correction is still easy, is now basic competence. The broader principle, that a lawyer remains responsible for what an AI produces, is the same one that governs research and drafting, set out in the lawyer’s duty to verify AI output.

It also pays to keep two different AI risks separate. A transcription tool can mishear; it does not invent. A general-purpose chatbot used for research can invent, and Indian courts have already been burned by it. The Delhi High Court in 2023 declined to rely on arguments built on ChatGPT precisely because large language models can fabricate case citations and facts, the cautionary tale told in our piece on AI-hallucinated citations in India. Transcription is the lower-risk end of courtroom AI. The high-risk end is anywhere a model is asked to state the law, which is why how you vet legal-AI citation accuracy is a separate discipline from checking a transcript.

Where this sits in the law: eCourts Phase III and court AI policy

None of this is happening in a policy vacuum, and understanding the scaffolding helps you predict where it goes next.

The money and mandate come from the eCourts project. On 13 September 2023 the Union Cabinet approved Phase III of the eCourts project for four years with an outlay of 7,210 crore rupees, a roughly fourfold jump over Phase II, per the Press Information Bureau and LiveLaw’s report on the approval. Phase III is run jointly by the Department of Justice and the eCommittee of the Supreme Court, deliberately in a decentralised way through the respective High Courts. That decentralisation is why you see Kerala mandate a tool while another state merely pilots it: the centre funds the direction, the High Courts choose the pace.

On the AI side, the architecture is still patchwork. The Supreme Court already runs other tools, SUPACE for research assistance and SUVAS for translating judgments into Indian languages, and has an AI committee studying integration, with reported collaboration with IIT Madras. The recurring official line is human supervision, ethical oversight, privacy protection, and that only a judge may sign an order. What does not yet exist is a comprehensive national AI law for courts. As Global Voices notes, the rules are scattered across court circulars and the data protection statute, which is exactly the kind of gap that lets good and bad implementations grow side by side. For the bigger picture of where these tools fit, see our overview of AI in Indian courts and the Supreme Court’s own AI rules and what tools must meet.

The direction is set. Recorded, retrievable, searchable proceedings are the future of Indian trial courts, and transcription is the thin end of that wedge. The open questions are quality, privacy, and accountability, not whether it happens.

How Niyam helps you work with AI-transcribed records

Adalat AI changes how the record is made. Niyam helps you work with what comes next: turning that record, and the law around it, into something you can argue from.

Once a deposition or order exists as clean digital text, the bottleneck moves to research. You have a faster transcript, and now you need the controlling judgments, the current position on a point, and a check that the precedent you are about to cite is still good law. That is what Niyam is built for. Ask a question in plain English, such as “what is the procedure for recording witness evidence under the BNSS” or “can an AI transcript be the official court record,” and Niyam answers with relevant Indian judgments, every proposition tied to a real case you can open and read, which is the opposite of the fabrication problem that sinks generic chatbots in court.

That distinction is the whole point of a native Indian legal AI versus a generic GPT. A transcription tool and a research tool are different jobs, and both demand the same discipline: the machine drafts, you verify, and every citation must resolve to a source. Niyam is built so the verification is the easy part, because the source is always one click away. If you want to see how grounded AI legal research in India compares with what you use now, that is what our comparison lays out.

Start for 100 rupees

Try Niyam on your next research question. For 100 rupees you get credits to run real research grounded in Indian judgments, with every answer cited to a case you can read. Create your account and start for 100 rupees.

Frequently asked questions

What is Adalat AI?

Adalat AI is a nonprofit legal-technology tool that provides real-time speech-to-text transcription of court proceedings in India, in English and several Indian languages, plus related workflow features like case-flow management and document automation. It transcribes witness statements and evidence as they are spoken, replacing the older method of dictation to a stenographer or typist. It was founded in 2023 and began its India deployment in early 2024.

How many courts use Adalat AI?

As reported at the Carnegie Global Technology Summit in December 2025, the tool runs in roughly 4,000 courtrooms, about one-fifth of India’s district courts, with nine formal state partnerships and five more states in advanced talks. That is a large footprint for a tool this new, but it is not yet nationwide.

Is Adalat AI mandatory anywhere?

Yes, in Kerala. The Kerala High Court issued a directive on 27 September 2025 making the tool compulsory for recording witness statements and evidence in all district courts from 1 November 2025, after a pilot that began on 1 February 2025 in Ernakulam. Kerala is the first state to mandate such a system across every trial court.

How accurate is the transcription?

The reported figure is 95 percent accuracy, but that was demonstrated in training sessions, and real courtrooms are harder, with crosstalk, accents, dialects, and code-mixed speech that lower accuracy. The Supreme Court’s separate live-transcription experiment using different software managed to publish complete transcripts for only 36 percent of its Constitution Bench hearings, which shows how far a demo can be from a dependable record. Always read the transcript against what was actually said.

What languages does it support?

It is built to handle Indian accents and multiple Indian languages. The Kerala deployment uses a dedicated legal speech model in Malayalam, and the tool also works in English and Hindi, with other languages extended using native language experts. Quality varies by language, because each language needs its own trained model rather than a simple translation.

Does the AI decide cases or write judgments?

No. Adalat AI transcribes and manages the record. It does not judge, score credibility, or predict outcomes. The Kerala High Court’s 2025 AI policy expressly treats AI as an administrative tool only and prohibits generative AI from drafting judgments or predicting case results. Only a judge can sign an order.

Can an AI transcript be the official court record?

In practice the process keeps a human in charge: the transcript is produced in real time, the witness can review it on a screen in the witness box, and the judge corrects and finalises it before it is signed and stored. So the official record is a human-verified transcript, not a raw machine output. The recording of evidence is still governed by the BNSS and the Bharatiya Sakshya Adhiniyam.

What are the privacy concerns?

A verbatim transcript of every witness is dense with personal data, and the Digital Personal Data Protection Act, 2023 applies to automated processing of such data. Open questions include who can access stored depositions, how long they are retained, where servers are located, and how in-camera matters are protected. Kerala’s policy bars uploading confidential judicial data to external public platforms, but there is no uniform national rule yet.

How does this affect open court and transparency?

Both ways. An accurate, searchable transcript improves transparency, because parties, press, and researchers can see exactly what was said. The same searchable, permanent record also raises surveillance concerns if access is not tightly controlled. The technology has outpaced the safeguards, which is the heart of the debate.

Will this actually reduce case backlog?

It clearly speeds up evidence recording: courts report rising from about two witness statements a day to six or eight. Whether whole cases finish faster is harder to verify, because a trial only accelerates if every other stage keeps pace. The steno-replacement gain is well attested; the broader 30-to-50-percent timeline claims come from the tool’s own reporting, so treat them with more caution.

Who funds Adalat AI?

It is a nonprofit with relatively modest annual operating costs, funded by philanthropic backers including MIT Solve, Echoing Green, Fast Forward, and The/Nudge Institute, according to the Stanford Social Innovation Review. Its nonprofit status is part of why High Courts have been willing to adopt it inside the courtroom.

How does the government’s policy support this?

The eCourts Phase III project, approved by the Union Cabinet in September 2023 with a 7,210 crore rupee outlay over four years, funds the digital-courts push and is run in a decentralised way through the High Courts. That is why one state can mandate a tool while another only pilots it. There is, as yet, no comprehensive national AI law for courts.

What should lawyers do differently now?

Treat the transcript as a machine-produced first draft of the official record and verify it at the time of recording, while correction is easy. Watch for misheard negatives, wrong dates, and dropped qualifications, because those errors can change the meaning of testimony your appeal depends on. And keep transcription separate in your mind from research chatbots: transcription can mishear, but it does not invent citations the way a general AI can.

How is Adalat AI different from the Supreme Court’s transcription system?

They are different products. The Supreme Court’s live-transcription experiment, launched in 2023, used software called Teres by Nomology Technology for Constitution Bench hearings and struggled with reliability. Adalat AI is a separate tool aimed at district trial courts, with witness review screens and judge-controlled correction built into the workflow. They share the same core challenge: making a machine transcript dependable enough to be an official record.

Where can I read more about AI in Indian courts?

Start with our overview of AI in Indian courts, then the Supreme Court’s AI rules and what those rules require of tools. For the risk side of courtroom AI, see AI-hallucinated citations in India and the lawyer’s duty to verify AI output.