A Conversation with Vikas Mahendra

You graduated from NLS, pursued a combined economics and law masters, and built a career in international arbitration. What was the moment – or the accumulation of moments – that made you decide to pursue AI at the intersection of dispute resolution?

I think we started from the point of: there is a problem and we need to solve it. The AI piece came subsequently. When I was at Herbert Smith Freehills – London, Paris, Singapore – resources were not scanty. You had everything at your disposal. An integral part of that was real-time transcription. We used LiveNote, and as people spoke, it came up on screen as the official record of the matter. It suddenly accelerated the speed at which you were working, made hearings far more effective – all of those wonderful things.

When I came back to India around 2015-16, I realised there was a significant lack of infrastructure to support that kind of offering. In the high-value matters where people did use transcription, they would fly in stenographers from Dubai, Singapore, London – business class tickets, five-star hotels, the works. I thought it was a criminal waste of resources. That is really when we started asking: what can we do about this?

We realised quickly that you cannot build the human pool in India the way it exists elsewhere. Typing 200 words a minute is not easy. The training institutions were largely in Australia and the US. So we said, let us fundamentally rethink what this means. That is when we started looking at AI – natural language processing, speech-to-text engines – as the alternative. But we realised they were not accurate enough for legal proceedings and would never be, on their own. So we added a layer of human review. Instead of technology doing all the work, or humans doing all the work, we said: let the tech do the heavy lifting and let the humans review it. That combination is what gives us what we need. That is really how TERES started.

For readers who may not be familiar with TERES AI, could you give a sense of what it does, the range of services it offers, and how it has evolved?

When TERES started, it was a legal transcription offering. In proceedings – whether you are cross-examining a witness in court or arguing before an arbitral tribunal – everything that the lawyer, the witness, and the judge or arbitrator says is recorded verbatim. That sounds straightforward, but the difficulty is twofold.

First, in legal proceedings, any inaccuracy is potentially fatal. You cannot afford errors. So you need technology sophisticated enough to almost eliminate them, and when the technology falls short, you need a human in the loop doing a fine-tooth-comb review. Second, these proceedings are full of proper nouns and technical terms that standard tools handle badly. We go through every case, study it, brief our reviewers on the unique vocabulary, and work to get the accuracy level up through both upgraded technology and upgraded human review.

The benefits are real and measurable. In today’s Indian courtrooms, a question and answer through a stenographer takes roughly one to two and a half minutes. With our offering, you can do it approximately five times faster. In a cluttered legal system – where time is at a premium, judges are scarce, and the docket is enormous – savings of that magnitude make a material difference. There is also a significant cost argument. Transcription is expensive upfront, but people often compare only the cost of a stenographer versus our rate without looking at the total picture. We had one hearing that was provisioned for ten days; transcription got the parties done in two to three. That saved them, across senior Supreme Court judges sitting as arbitrators and a tier-one law firm, something in the order of two and a half to three crores.

But the argument I care about most is integrity. Today, in most Indian court hearings, the lawyer asks a question – but that is not what gets recorded. The arbitrator or judge rephrases it. The witness answers – but what gets recorded is the judge’s paraphrase or summary. You have lost the actual answer. I have seen instances where a witness said, ‘That is not what I said’, and the judge said, ‘Yes it is’. and there was absolutely nothing to be done about it. Verbatim transcription fixes that. It is also an accountability mechanism. If everyone knows that every word is being recorded, the time wasted discussing cricket and Bollywood, the politically incorrect remarks, the tactical stalling – all of that tends to disappear.

Beyond transcription, we have expanded into document management and e-bundling, electronic presentation of evidence, and linking audio-video recordings to transcripts so a click of a button opens the relevant document at the relevant moment. Our most recent offering is live translation. At ADGM, the proceedings are in English but many clients speak Arabic. So we built a pipeline, where transcription happens, our reviewers correct it in real time, and the corrected transcript is sent sentence by sentence for translation. You get a far more accurate translation than speech-to-speech systems produce, with a lag of only about five to ten seconds. The most unusual use case so far is when we transcribed Swedish court proceedings and translated them into Russian and English simultaneously.

We service more than sixty countries globally. We are the official provider at the Dubai International Financial Courts, an approved provider at ADGM, and active at Maxwell Chambers in Singapore, the Hong Kong International Arbitration Centre, Arbitration Place in Canada, and the IDRC in London, among others. We always conceived of this as a global product.

How does the onboarding process look when a court wants to adopt TERES AI? And how do questions of data sovereignty and confidentiality get addressed?

Different courts have different approaches. By far the most rigorous process we went through was with the Supreme Court of India – and I will be honest, our first inroad was entirely accidental and required a fair amount of shamelessness on my part.

I walked straight up to Justice DY Chandrachud when he was Chief Justice. He happened to be speaking about transcription at the Delhi Arbitration Week – an event we were transcribing. That was serendipitous in its own right: his court clerk at the time, who had written his speech, was an NLS alumna I had taught as a student. As a practising lawyer she had used our services, so she understood the benefits firsthand, and she had woven elements of transcription into his speech. I did not know any of this until he started speaking. So I approached him afterwards, pointed out what was happening live on stage, and he said, ‘Very interesting. What are you doing next Tuesday?’. That was on a Friday. By Tuesday we were transcribing a constitution bench hearing. They liked the pilot.

After that first hearing there was a second constitution bench hearing, and no one had arranged transcription. The Chief Justice asked his IT team why it was absent; the IT team asked why we were not doing it. I said I could not do this for free indefinitely. That conversation led to a work order from the top, and simultaneously to a very grueling tender process – detailed submissions on our data stack, our API calls, our technology chain, the whole thing evaluated by the AI head of NIC. There was a qualifying stage, a technical bid, a financial bid, product demos, interviews. Only after all of that did we get the contract.

On data sovereignty: for public hearings, the Supreme Court’s ultimate position was that there are not as many data-sovereignty requirements as one might expect. They were comfortable with API calls to foundational models, provided the accuracy standards were met and they had full visibility of the chain. For other use cases – such as the Karnataka High Court deployment, where we assisted judges with dictating judgments – the requirements were far stricter. Those were draft orders. We had to run an air-gap arrangement where data was processed entirely on the server side, we never retained it, we had no visibility into the content. The confidentiality requirements scale with the sensitivity of what is being processed.

Constitution bench hearings are famously chaotic environments. What did operating at that scale teach you about the limits of your technology, and how did it change the product?

Constitution bench hearings remain, to this day, among the most complicated things we transcribe. The core challenge is speaker identification. If you look at any video of a Supreme Court hearing, there are typically around ten lawyers standing near one microphone. Identifying who is speaking is the biggest technical problem, and I do not think there is an easy fix.

At the tender stage, there were providers who came in convinced it would be a purely technical solution – Amazon, Google, Jio, among others. They saw the video and asked: how do you identify speakers? Can we get voice samples, voice signatures? It quickly became obvious that we were not going to get any of that. That is exactly why the human review layer is so crucial and why we were able to deliver what pure-AI providers could not.

That experience also deepened our preparatory work considerably. Before any hearing, we now build a full dossier for our transcribers – for example, photos of every judge, their names, a detailed brief on the case, all the unique provisions and case-law references, the names of the lawyers arguing and the positions they are taking. All of that feeds into accuracy. People see the clean transcript on screen and assume not much is happening. In reality there is an enormous amount happening behind the scenes.

Could you walk us through exactly how the human verification layer works in real time?

There are five people reviewing the transcript at any given moment. Two are reviewing in real time – their job is primarily to ensure speakers are correctly identified and attributed, and to catch obvious errors as they arise. Two more are reviewing with a slight lag, going through fifteen-minute chunks: they read and listen to the audio at their own pace, rewinding and re-listening as needed to push the accuracy level up towards a hundred percent. One person sits above all of this reviewing everything from start to finish for consistency. We also have several AI review layers built into the pipeline. The combination of that entire model is our differentiator. It is how we produce something meaningfully different from everything else in the market.

How has the broader reception to TERES AI been within India’s court system, which remains largely paper-based?

Honestly, it has been challenging. We have had some acceptance in specific use cases – the Karnataka High Court is a good example, where we deployed the technology to assist judges with dictating judgments rather than for live hearing transcription. But for actual hearing transcription across the high courts, reception has been limited.

I have some sympathy for why. It is an expensive process. Courts are cautious about spending. People are still working out sustainable funding models – whether to front-load costs and charge parties on a pro-rata basis, make it optional, and so on. None of that has been settled yet. So adoption on the speech-to-text side has been slower than I would have hoped.

There has been somewhat more openness toward our generative AI tools, which I can speak to separately. And the broader space is moving. One company doing excellent work in India today is Adalat AI – they operate a very different model, they are a not-for-profit and can effectively offer services at no cost to courts, which is something we simply cannot do. The Kerala courts are a good example of where adoption has been more substantive; transcription is now becoming mandatory for evidence recording across multiple courts there. But even there, the penetration is mostly on the side of judges dictating orders rather than entire proceedings being transcribed. It is a slow-moving wagon.

How is the international arbitration community engaging with AI more broadly? Where do you see clear legitimate use cases, and where do you think caution is warranted?

AI usage has very different flavours, and I think it is important to disaggregate them.

Transcription is, in my view, nearly uncontroversial. You are replacing an extremely tedious, high-intensity administrative task with something more reliable, and you are keeping humans in the loop. The resistance you encounter there is practical, not philosophical – existing systems work, people do not want to absorb the cost of change, there is inertia. The only real philosophical concern in that space is that people conflate AI generally with hallucinations and errors. 

The genuinely contested territory is everything else: AI usage by arbitrators, by counsel, by witnesses, by experts. I think each of those requires distinct treatment.

For lawyers, I do not think there should be any regulatory restriction on AI usage, because the existing frameworks for professional accountability already handle it. If I use AI to build a better brief, draft better pleadings, or process documents faster, that should be no concern of the arbitrator – unless I am compromising confidentiality or using private data to train other models. Set the guardrails there, and then step back.

The hallucination concern for lawyers is real but not new. Before AI, we used case digests for legal research. At least ten to twelve percent of the time, the citations in those digests were wrong! No court ever accepted ‘the digest was wrong’ as a defence. It was always the lawyer’s responsibility to check the work before submitting it. Nothing has changed. If an AI gives you a wrong citation and you submit it without checking, that is your failure. 

There was also that notorious episode – I believe it was a high court, though I cannot recall which – that spent a decade deciding cases on a particular statute, only to discover eventually that the statute had never been gazetted. It was not even law. Even in the pre-AI world, false positives were a persistent reality. We managed them through professional standards, ethics, and accountability. The same applies now.

Witnesses are a somewhat trickier category. To what extent can a witness use AI to review large volumes of documentation before giving a statement? I would say: as long as the witness statement reflects their genuine personal knowledge, as it is meant to, that is acceptable. Whether you need to disclose the use of AI is a question I am genuinely ambivalent about. I do not disclose that I use Westlaw or Microsoft Word. Why should this be different? But the extent to which a witness relied on AI output is something that may warrant more careful thought in specific cases.

Experts are similar. They have always used technology. The question is whether the use of AI contradicts the independence and personal expertise that an expert is supposed to bring. That line is for the expert to draw, and it is probably worth thinking through more carefully in the context of specific testimonies.

The most complicated category is decision-makers – judges and arbitrators. This is actually where we are building now. The reality today – and I think people need to accept it – is that judges are already using AI. Our position is: if that is the reality, why not give judges tools designed specifically for their context, with appropriate guardrails built in, rather than letting them use whatever unguarded general-purpose tool happens to be available?

The philosophy we are developing is around what I would call intentionality. The goal is not to displace decision-making – that is still core to what a judge does. It is to surface the right documents, identify points of agreement and disagreement, always with the source clearly shown and easily verifiable. And critically, to surface the areas of lowest confidence in the AI’s own output, so the judge can evaluate and probe those specifically. Today’s AI output is so cleanly formatted and delivered with such apparent confidence that people assume it is right. We want to counteract that by making the uncertainty explicit.

You mentioned San Francisco and a second phase of what you are building. Where do you see TERES AI in five years, and how does the San Francisco work fit into that picture?

The San Francisco work is actually a separate company that we are in the process of incorporating and it will run under the TERES umbrella. The focus is specifically on lawyers who handle complex disputes, and we are starting with the most intensive category of all: construction arbitration and construction litigation.

Five years is frankly too long a horizon to forecast meaningfully in AI. Let me give you my best read on the next eighteen to twenty-four months instead, because I think that is where the pivotal shift happens. AI is becoming more reliable. People are beginning to recognise that it can genuinely augment their work. But there is still a great deal of hype, scepticism, and fear – all of them with some legitimate basis. The chips will fall in the next eighteen to twenty-four months. People will change how they work very fundamentally.

To put the scale of this in context: the last comparable transformation was when lawyers moved from handwritten or typewritten pleadings to a word processor. This is that shift, multiplied by roughly a hundred.

Our particular philosophy – which I will acknowledge is contrarian, especially sitting here in Silicon Valley where every conversation is about agents, autonomy, AI doing everything in the background – is that in complex disputes with high stakes, you do not want AI in the driver’s seat. You need the lawyer to understand the case because they are the one arguing it. Firms pay significant sums for good lawyers precisely because of the wealth of experience, knowledge, and strategy they bring. That must drive the process. Agentic AI talk does a real disservice to this category of disputes, where errors are costly and the value that experienced lawyers contribute is genuine and irreplaceable.

So we are building systems that very deliberately keep the lawyer thinking, that force them to take decisions, to engage with the material, to bring their expertise to bear. We even challenge our own AI outputs. For instance, if the AI generates an analysis, we also ask it to critique that analysis – what are the biggest gaps, the biggest weaknesses, the factors that have not been adequately considered? That is what we are building for complex disputes. Bring the human back into the loop. It is going against the tide, but I think it is the right way to shape this ecosystem, and we will see in the next few months and years whether the market agrees with us.