One Year of the Kerala Mandate: Impact, Challenges, and Learnings

In September 2025, the High Court of Kerala issued a notification making Adalat AI mandatory for recording witness depositions in all trial courts across the State’s district judiciary from November, 2025, as part of an effort to “reduce delays in the process of recording evidence and to modernise the system of recording witness depositions.” With this, Kerala became the first Indian state to mandate statewide adoption of Adalat AI.

The mandate brought Adalat AI to 786 courtrooms across all 14 districts. Over the past year, Adalat AI has been used to transcribe more than 25,600 hours of audio and generate 4,00,000 legal documents. While the mandate directed judges to use Adalat AI to record witness depositions, many judges have gone beyond the mandate, using the platform to dictate orders and judgments from their chambers, and to research and edit documents.

For us, the Kerala mandate was more than just a milestone. It was a defining test. For the first time, an Indian state was introducing AI into how its courts record evidence, and it was doing so in partnership with a non-profit. The stakes were high, and so was the resistance. Change is never easy, and it is harder still when it means learning a technology that many judges had never used.

This blog is the story of that journey. A year on, we take stock of what it took to bring a new technology into every court in Kerala. We begin with the mandate in numbers: how widely Adalat AI is used, how usage varies across Kerala’s districts, and how the way people use it has changed. We then describe the challenges Kerala’s courts presented and how we adapted the platform in response. Next, we trace how our training evolved as we learned how best to support the Kerala judiciary, before ending with our plans for the year ahead.

A Stenographer logging into Adalat AI platform in Ernakulam District Court

The Mandate in Numbers

We draw on a year of usage data to examine how widely Adalat AI has been adopted, how usage varies across districts, and what judges use the platform for.

Usage Growth and Steady State

Figure 1: Minutes transcribed and active users per month

Following the announcement of the mandate in September 2025, Adalat AI launched an intensive training programme across Kerala’s court complexes. In October 2025, we conducted 22 training sessions, including 12 in person and 10 online sessions. Together, the announcement of the mandate and the training drove a steep rise in usage. Between September and November, monthly active users grew fivefold and minutes transcribed tenfold.

Since then, usage has settled into a steady pattern. As Figure 1 depicts, since November 2025, around 1,100 people have used Adalat AI each month, transcribing about 150,000 minutes of audio. Minutes recorded dip in December, April, May and August, when the courts are on vacation.

District-wise Variation in Usage Intensity

Figure 2: Monthly minutes transcribed per active user, by district

The primary metric used to measure usage intensity is average monthly minutes transcribed per user. Broadly speaking, districts which see the most use are in the centre and south of the state, with usage being lightest in the north. In Kottayam, which sees the highest usage, an active user transcribes about 145 minutes a month; while Wayanad is only at a third of this number at 51 minutes recorded each month. So far our in-person training has concentrated on south and central Kerala, where most of the larger court complexes are, and this may partly explain the difference. In the coming year, we will extend in-person training to court complexes in the north.

Figure 3: Gap in habitual use between the highest and lowest districts

Another way of measuring usage intensity is by gauging whether the use of Adalat AI is becoming an embedded part of the judicial workflow for judges across Kerala. We track this through ‘habitual’ users, i.e., active users who use the platform at least 3 out of 4 weeks in a month.

The share of habitual users in each district has varied over the year. Just before the mandate, in October 2025, Thiruvananthapuram district had the highest share (65%) and Thrissur district the lowest (29%). By September 2026, Idukki district led at 94% and Palakkad district trailed at 78%.

The gap between the districts with the highest and the lowest percentage of habitual users has seen a steady decline. As our team has expanded training sessions across all districts, the gap in habitual use between the highest and lowest districts declined from 36.1 percentage points in October 2025 to 15.7 percentage points in September 2026.

Changing Composition of Adalat AI’s Platform Use

Figure 4: Changing usage patterns on the Adalat AI platform

As judges and court staff have grown familiar with the platform, the way they use it has changed. In October 2025, just before the mandate took effect, batch dictation was the most common way of using Adalat AI: with 55% of users primarily using Adalat AI for this purpose, against 22% for live dictation and 18% for document editing. [1]

A year later, the mix looks quite different. Batch dictation has grown to 61%, while live dictation has fallen to 6%. Most of October’s live dictation users have since moved to batch. Document editing has risen to 29%. This change can be explained by the fact that batch dictation is more accurate than live dictation and supports AI-assisted formatting of dates, numbers, abbreviations, case law, and statutes.

Challenges and How We Responded

The mandate meant that every judge in the district judiciary had to move their daily work from a predominantly paper-based workflow to a computer. Comfort with technology varies widely and for judges who had worked a certain way for decades, the cost of switching was significant.

Change always meets resistance, and that was evident in the questions we faced during our initial interactions. Would it work? Would it add more to my workload? What if my computer crashes and I lose everything? Will it understand Malayalam and local dialects? Will this free tool turn into a paid one, with the High Court rolling back and all our learning going to waste? Some judges were openly sceptical of the decision itself.

It was hard to blame them. These are judges already carrying heavy workloads and long backlogs, and we were asking them to learn a new application, change the way they had worked for years (in some cases decades), and place their trust in a non-profit that was barely a year into the field and had no active mandates yet.

We knew a mandate notification by the High Court alone would not make this work. We needed the district judiciary’s buy-in, and we set out to earn it in two ways. The first was training: carefully designed online and in-person sessions that took judges from the basics, like operating Google Chrome, to advanced in-platform features such as shortcuts and templates. The second was co-designing the product. We actively collected feedback, and used it to improve both the platform and our speech-to-text models.

Beyond the resistance, Kerala’s courts also stretched Adalat AI in ways we had not anticipated, and each challenge pushed us to support the judiciary better. Here are some of the specific challenges we faced, and how we countered them.

Courtrooms with unreliable internet connectivity

Kerala’s courts span everything from newly built complexes to older buildings never designed with digital work in mind. In many of these court complexes, multiple courtrooms share a single server or Wi-Fi connection. For an internet-reliant platform that is built to capture every word spoken in court, an unstable connection puts the completeness of the judicial record at risk. It also fed a fear we heard from many judges: what if the connection drops, or the computer crashes, and I lose my work?

To address this, we built a safeguard into every recording called Zero Data Loss. If the internet drops, judges can keep recording offline for up to five minutes, and as long as the connection returns within 30 minutes, the full recording is saved and transcribed. We are now working to extend the five-minute window. Knowing their recordings are protected has helped ease judicial anxiety, as one judge’s feedback shows:

Internet connectivity continues to be one of the biggest challenges in digitisation of our courts. Intermittent network issues, especially during the rainy season, disrupt our use of online platforms. But the Zero Data Loss feature has been extremely helpful in overcoming this challenge.

-Additional District Judge, Palakkad

Scarcity of training data for the Malayalam ASR model

When building our Malayalam language model, we were cognisant that courtroom Malayalam is a world of its own, full of legal terms and sentence structures that rarely come up in everyday speech. For instance, the word ‘ബന്തവസ്സിൽ’ is the Malayalam term for the word ‘custody’ that frequently appears in court records and depositions, but is not commonly used in everyday parlance. A similar example is the term ‘സത്യവാങ്മൂലം’, which means ‘affidavit’. Words like these posed a challenge for training our models because publicly available data did not contain any of these unique legal Malayalam terms and phrases.

To overcome this challenge, we worked with judicial officers and court staff, cataloging the words that fill their working day and the contexts they are used in. From this, we built a court-specific dataset from the ground up, which was used to train and improve our Malayalam model. This is an ongoing process – the more the platform is used, the more we learn through user feedback, and our ML research team incorporates this feedback to iteratively improve the models.

Managing Two Languages in a Single Dictation

A deposition in Kerala is mostly recorded in Malayalam, but English is sporadically interwoven in every dictation. Terms like “seizure mahazar”, “FIR” and “medical certificate” come up regularly. So we needed a model that could recognise both languages in the same breath. On top of that, it had to get the details of the record right: dates, numbers, and the shorthand that courts rely on, all written exactly as they would appear in the official court record. For a sceptical judge, a platform that stumbled on such integral terms like these would have been reason enough to go back to paper.

To resolve this issue, we fine-tuned our Malayalam model to understand the English terms most commonly used in witness depositions. For example, the model can recognise witness labels like PW [Prosecution Witness], DW [Defence Witness], and CW [Court Witness], item labels such as M.O. [Material Object], legal abbreviations commonly used in depositions like FIR, IPC, CrPC, case type abbreviations like OS (Original Suit), IA (Interlocutory Application), and transcribe them in English, even when they are interwoven with Malayalam dictation. The following example illustrates how the model recognises and retains a commonly used English legal term in deposition while transcribing the surrounding Malayalam dictation.

While this works well as an interim solution for English terms embedded in Malayalam sentences, we are working on a more comprehensive bilingual model that is capable of moving between both languages seamlessly.

Balancing Screen Time and Witness Observation

As judges started using the Batch model on Adalat AI, they told us the platform was taking too much of their attention and drawing their eyes away from the witness. In a courtroom, words are only half the story. A pause before answering, a glance away, or a sudden change of tone can bear on a witness’s credibility, and judges may need to note it in the deposition. Noticing these cues means watching the witness, not the screen. This was co-design in action: the feedback came from the courtroom, and we delivered the fix. We made two changes to help provide a more hands-free experience.

First, we improved the speed of the Batch model, reducing the time taken to generate transcripts. This helps Judges to view the transcribed content more quickly after recording, allowing them to stay focused on the witness rather than the screen.

Second, ending a recording used to mean clicking “Stop Recording” in the toolbar, and switching languages mid-dictation meant navigating an extensive drop-down menu. Both pulled the judge’s eyes from the witness to the screen, so we introduced keyboard shortcuts for each.

How Trainings Evolved Over the Year

At Adalat AI, we have adopted a dual-engine philosophy. The first engine is ‘technology’ – building products that remove bottlenecks in court workflows and are designed for easy adoption by judges. The second and equally important engine is ‘partnerships and programming’ - that actively supports judges to drive adoption, trust and daily use.

We realise that public institutions are not always tech-forward environments, and a well-built product is only useful if it gets used. That is why training was our first answer to the resistance we met in Kerala. Our on-ground team trains judges and court staff, online and in person, and carries every piece of user feedback back to our product teams. The result is a training ecosystem that supports users at every stage, from first login to advanced mastery.

Over the past year, our approach to training changed almost as much as the platform itself, shaped at every step by what judges and court staff told us they needed.

From Online to In-Person Training Sessions

To train as many officers as possible, we initially focused on online training sessions which could provide scale by bringing in large audiences quickly. However, we soon realised that in-person training was crucial for understanding where users were getting stuck and helping them. This especially mattered because judges and court staff vary widely in their comfort with technology. For those new to it, adopting an unfamiliar tool was always going to be hard. In-person training sessions enabled us to answer every question, and every user had someone beside them guiding them the first time they tried it.

Stills from online training sessions conducted for (left) hon’ble judges of the Kerala High Court and (right) for judicial officers from Pathanamthitta and Wayanad districts.

We travelled across Kerala, demonstrating the platform live in courtrooms, walking users through it step by step, answering questions and fixing problems for court staff in their chambers. To date, we have held 64 training sessions across the state: 43 for judges, 17 for court staff and 4 for both.

We also realised that there is no substitute for hands-on experience on our platform, so we now ask judges to bring their laptops to every session. For court staff, who usually work on desktop systems in the judge’s chambers or offices and have no personal device to bring along, recreating that hands-on experience is the next frontier that we are actively working on.

In person training sessions conducted for judicial officers and court staff in (Clockwise from top right) Kollam, the Kerala Judicial Academy, Kottayam, and Idukki

Embedding Basic Computer Literacy in Training Sessions

Before judges can dictate a judgment or record a deposition, they need to feel at home with a computer. Soon we noticed that many of our training sessions had turned into computer classes, with us walking users through basic computer literacy, such as helping them install Google Chrome, log in to their email accounts, and change email passwords. We came to see those moments not as detours but as the foundation we needed to establish.

Small-group sessions give judges and court staff hands-on practice and facilitate effective doubt clearing.

Focusing on Refresher Trainings

Our products evolve constantly, shaped by the feedback we receive. Our conversations with judges and court staff tell us what slows them down and what would make the platform work better for them, and we transform these suggestions into feature updates. Refresher training closes that loop, connecting each update to the people using it, helping judges and staff adapt to changes and get the most out of the platform.

A virtual refresher training session conducted on the newly launched Legal Lens feature for judicial officers across various districts of Kerala

Cultivating Trainers from within the Judiciary

This year we launched our Master Trainer Programme, in which we trained selected judges to teach their peers how to use Adalat AI. Through an intensive online course, we trained 28 judges, with every district in the state represented.

The model is promising because judges understand courtroom workflows better than anyone. A Master Trainer can show a colleague exactly where the platform fits into a deposition or a judgment, in the language of the courtroom and with practical tips from their own experience. Colleagues are also more comfortable asking a fellow judge their questions, and the training carries a familiarity and trust that is hard to match.

The programme also reflects how we see our role. We are not a vendor shipping a product to the judiciary. We work alongside judges to improve judicial workflows, and the Master Trainers are among our closest partners in doing so.

(Left) Sri Rohit Nanda Kumar, a judicial officer from Ernakulam conducting a demo session for trainees at the judicial academy; (right) Manish, Senior Civil Judge, leading a training session for fellow judicial officers.

Continuous User Support and Co-designing the Platform

Alongside training, we have established multiple channels for user communication, feedback and support. Judges and court staff can report issues, share feedback and request features through the in-platform feedback option, a dedicated WhatsApp group, and helpline numbers that offer support by call and WhatsApp message. We also launched Office Hours: short online group sessions where users bring their questions, discuss workflow challenges and ask to see specific features in action, without waiting for the next training session.

These communication channels are not limited to resolving issues, they also allow users to share their needs and suggestions, helping us continuously improve the platform. Feedback from Kerala has already led to features such as text highlighting, tables and document duplication.

(Left) Online office hour session with judicial officers; (Right) WhatsApp community with district-wise groups of judges and court staff to facilitate resolution of doubts and peer learning.

What’s Next: Year Two and Beyond

The mandate’s first year was about onboarding and adoption. In the year ahead, we want to deepen that adoption by strengthening user support, shipping feature updates shaped by user feedback, and giving courts with low usage more targeted training and help.

Filtering out Noise

Courtrooms are never quiet, and so we want to build a model that holds up even better in low-connectivity, noisy environments. Our engineering and ML teams are working on improving the models such that they recognise and transcribe only the judge’s voice from the volley of voices in a courtroom.

Improved Trainings for Greater Adoption

A year into the rollout, we have a much clearer picture of which courts have embraced the platform and which need more help. In the second year, we plan to use these insights to guide our training efforts. This will involve conducting targeted training sessions in court complexes where platform adoption remains low, alongside one-on-one interactions with users who have yet to fully adopt the platform.

Additionally, the focus would be to augment the Master Trainer Programme. We imagine a future where master trainers take the lead in their own districts, guiding fellow officers, answering their questions and sharing what has worked for them in court.

Towards Courtroom-Ready Malayalam

Malayalam is one of the hardest languages to build speech models for, because high-quality training data is scarce. Building one for courtroom Malayalam, with its legal vocabulary and the conventions of deposition recording, is even more challenging. On this front, our newer speech model has made measurable progress: in a typical 100-word legal dictation, the number of corrections needed has fallen from around 15 to 9. In year two, we will keep pushing accuracy up and make real-time Malayalam transcription faster.

Conclusion: A Year of Learning

The Kerala mandate, in many ways, was our proving ground, and it has been delivered. In a year, Adalat AI has grown to 786 courtrooms across all 14 districts, with more than 25,600 hours of audio transcribed, over 400,000 documents generated, and around 1,100 people using the platform every month. We have built a strong relationship with the High Court and are now exploring new projects together. And many of the judges who were once sceptical of us have become our strongest champions.

It has also been a year of learning for every part of our organisation. Our partnerships team learnt which training methods work best for judges and court staff. Our engineers learnt to optimise the platform for the realities of the courtroom, from patchy connectivity to noisy halls. Our ML team learnt how to improve Indic speech models for courtroom Malayalam, and our design team learnt how to make the platform more intuitive and easier to use. We also set up a new customer support function to address user doubts and act on feedback.

These lessons have not stayed in Kerala. They are shaping our strategy in our two other active mandates, in Sikkim and Andhra Pradesh, and in the 11 other states where we are actively deploying Adalat AI.

We are grateful to the High Court of Kerala for placing its trust in us with our first statewide mandate. We also thank the officials of the Registry, the judges of the High Court and the District Judiciary, and the court staff for their continued support and valuable feedback, particularly in the early stages of implementation. This mandate would not have been possible without the cooperation and support of the Kerala judiciary.

One year in, there is more to do, and we look forward to doing it alongside Kerala’s courts.

Notes

1. The difference between these two dictation models is that ‘batch dictation’ transcribes speech at one go after the user has completed dictating whereas in ‘live dictation’ the transcribed text appears in real time while the judge is recording.