Harmeet Sandhu, Scottish Courts and Tribunals Service – In the age of AI, can justice be smarter?

National Event 2025 Special

At ‘In the Age of AI, can justice be smarter?’ expert speakers: Claire Feasey, Harmeet Sandhu, Dr Susie Alegre and Shami Chakrabarti gave an overview of the potential contained within AI for improving efficiency and driving change, whilst keeping a clear-eyed view on the rights and well-being of individuals. The talks were followed by an audience Q&A.

In this episode:

Harmeet Sandhu, a specialist AI consultant working with the Scottish Courts and Tribunals Service explains how using technology and AI can make complex work simpler and more reliable, and to drive measurable efficiencies across business workflows.

Transcript

Uh good morning everybody and thank you for having me. I've spent 25 years building and shaping technology. For the last decade I've been working very closely with the Scottish public sector. More recently I've been my focus has been in implementing AI safely, responsibly and with real impact. During my journey working with many public and private sector organisations, I know that we've made big strides in digital transformation. The services are faster, more transparent, um, compared to the paper black hole that we used to have, um, new mobile apps, shiny portals, but transformation rarely reaches the end journey. It doesn't. Inside our offices, my colleagues are still copy and pasting between two tabs. And that's real. They are moving information between teams. They are handling thousands of emails, chasing spreadsheets. It's repetitive work. And we all know about it. You know what I'm talking about. It's repetitive, but it's important high stakes work and it must be done right every time because people's life depends on it. And this is what slows the journey. Now let's imagine

an episode of Black Mirror. A world where someone commits a crime. The entire investigation, trial, and sentencing are run by AI bots. No judges, no jury, no defence council. Just an automated decision. Guilty. That's not going to happen. That's that's not going to happen. You know, just it's it's let's let's let's keep that in a Netflix episode of Black Mirror. But let's talk about a typical trial. Everyday judges, prosecutors, and defence teams handle huge amounts of information. We're talking about evidence files, statements, CCTV footage, transcriptions, reports from everywhere. Prosecutors must build a clear and fair case. Defence teams must review every detail. And judges must absorb all this information to reach a reason and lawful decision. This is where the new tools of AI can really make a genuine difference. For prosecutors, AI can help search and summarize large volume of information and evidence, identifying patterns and inconsistencies faster without missing the real detail. For defence team, AI can support document review automatically putting out case points and contradictions so that they can focus on the argument that really matters. And for judges, AI transcription and summarisation tools can turns hours of audio into searchable and reliable records, helping them review submissions quickly and concentrate on the case itself. Used this way, AI is isn't replacing judgment. It is removing all the noise around a case. It lets people spend less time finding information and more time thinking. But and this part really matters. It has to be done carefully. Our justice system carries a very heavy responsibility. It holds power to account, protects rights and ensures a fairness for everybody. And to keep doing that, it does need to evolve. The new wave of AI lets us lift our load. But we must be careful and careful. I'm going to stress on careful. Our rule is simple. AI is a co-pilot, not a pilot. AI is all around us. And we we've talked about this chip. It's in the banks, phones, Netflix episodes. It's everywhere. AI is everywhere. And I want to present some facts of some research done by LinkedIn and Microsoft recently. And some of these facts are shocking. 75% of workers are already using AI at workforce. 79% of leaders say adopt AI or lose competitiveness. It isn't optionally more. It's its survival. Productivity gains are massive. AI users handle 66% more work. And with programs programmers coding 126% faster. This is this is amazing. Time has been liberated. 90% of the workers report AI saves them time and tasks freeing them from strategic work. 75% of the companies are implementing AI within the next 5 years. And this is the best one. AI talent is scarce. 68% of business leaders struggle to find skilled AI talent. Creating a competitive advantage for the early movers. Introducing new technology in justice is hard work. And believe me, I know it's hard work. Never mind something as fast moving as AI. We have to be careful because the stakes are much higher. We are dealing with people's rights, their liberties and their livelihoods. And this is very important. So our approach has been deliberate and grounded in two principles. First, the purpose. Every project clearly serves a need. It needs to make access easier, decisions fairer, and processes faster. Ideally, all three. And the second principle is trust. It has to be understandable, auditable and lawful so people know what it is doing and why with humans in control. All of that sets us on the right path. But the reality of implementing something as new as AI is very different. And I'm talking from experience. You have to bring stakeholders with you on a journey. navigate legislation, data protection, lawful basis, and you have to watch security and emerging threat vectors. These are the ones we haven't discovered yet, so we don't know what we're playing with. The list is very long and there is a huge policy gap. The government and the legislation hasn't caught up with the AI policies yet. So that's that's a huge problem. There is a clause in the criminal procedure act of 1995 7 A and B and C that the person who prepares an AI well who prepares a transcript must sign it and certified that it's an accurate record of processing but I can't ask an AI bot to do that. So we need to work with the times. So let me show you what that looks like in practice and the bumps we hit along the way when we were implementing um these technologies. We are currently running two AI pilot projects. Um the first one is the AI transcription. Guided by the purpose and trust, we started lowrisk and high benefit work. We began AI speech to text and converting audio recordings into text. a transcription. This evidence evidence by commission is an evidence where a witness evidence of a witness in a trial so that the person doesn't actually go to court. Um these are vulnerable witnesses and children. We chose EBC because of the tightly controlled environment that we have and the audio recording because garbage in is garbage out and that was the reason we chose before the soft launch. We ran hundreds of hours of hearing through the system checking for accuracy with a a lot of testing and pre-processing and post-processing. We have now achieving 98% over 98% accuracy with all those even accents. Because of the criminal data, we could not send any of that to the cloud. So we had to build our own model and we do it all on prem. So there is no data going to the cloud. Um the project is still in its pilot phase. Um and the solution has produced 837 transcripts which equates to 100,000 transcription minutes. If we were to do that by an external company, they charge £2 a minute and take up to a week and maybe more. So we saving annually about £200,000 and we're doing it in minutes. This was a real change and the judges gave some feedback um during the pilot and some of this I will quote judges feels it allows them to focus on things like keeping an eye on the jury and accused you would assume they would do that but they keep taking notes. Judges reported back that the transcripts are useful for preparing notes. It also able to assist with them in any submissions made by parties regarding what was and what was not discussed in the commissions by checking the details in the transcript. And prior to the production of these transcripts, the judiciary had to rely on the crown in providing the notes taken by the crown not judges were also required to take notes during the playing of this disc. But the last one is what I really like. I have found the provision of this AI transcripts are something of a game changer. The transcripts are there for the judges but we also have the recorded audio and video. So if there is any problems the judge finds in the transcript they can always go back and that's what the AI transcription. Next is document intelligence and this is in civil cases. These are formal documents that start a civil case. They still arrive in this day scarce in PDFs and by email from all solicitors across the country. They all use their own case management system and they draft the forms in what they like. So the form you see will be no never the same if when it comes from a different solicitor company. And so they still arrive by email in different formats from all different solicitors. The case workers, they retype names, addresses, case details into our case management system. Imagine that. Imagine doing that. Copy and paste tap by tap tap by tab. This is slow and error-prone. What we did was we built an AI reader, a document intelligence. This solution pulls the key fields from document and send them into our civil case management system with acute human check. So human in the loop always 
 
 

it results in less far raking faster case registrations and fewer mistakes and less sensitive data in our inboxes all the case fonts. It wasn't straightforward though really it wasn't. Once we cleared the governance, we spent months wrestling with inconsistent layouts, maiden names, trading versus registered addresses and edge cases like companies which went to liquidation where the liquidators details are required in the form. After a lot of testing and very very very smart prompt engineering, we're now seeing 63% efficiency saving on the initial rate processing and this is in soft launch only. Just now only one solicitor company is using it in one quart. We set out for a 20% target efficiency and we we are gaining 66%. Looking ahead, we've got three ideas in the pipeline. Um, this is still in ideation and we are piloting is piloting this and I'm going to be doing set some demos of this. So we've got live transcriptions um real-time captions in the courts and this is something that the judges and sheriff really wanted um for accessibility uh for for the judge to to not to concentrate on making notes but to really um concentrate on the case the jury the accused and they have they know that they have the transcript there they can pause it they can annotate it they can comment on it and use it later. Um another another piece of technology that we are using right now uh or not using but we are piloting is the live translation. Again I don't know if anybody any of you have been in a in a court hearing and I I highly suggest you do go um go to Glasgow High Court and uh and and go and see some cases. Um some of the translators if they don't show up the case is dismissed and it takes weeks for the case to come back. Um so we are trying something like that. Uh which translates and you know it's always going to be a backup a backup system. Um we are also um trying AI in software development. Um we we want to speed up the changes within our legacy system. So these software development run by AI um will be reading legacy codes suggesting test drafting documents. So small fixes doesn't take weeks and doesn't cost like thousands and thousands across all the guardrails stay the same. It's a clear purpose a human in the loop auditable and lawful. We're also shortly going to be starting an organisation wide AI opportunities review um meeting each part of the business and understand the needs and the risks. We're going to work out on where AI can be used safely for the biggest gains for staff, judiciary and citizens. Main purpose of this exercise will be to co-design to build with our teams and not for them. It was you know we we've been on the journey for two years and and majority of the time was in the legislation the data protection impact assessments and these are very very important points AI goes really fast but we as as public sectors we to really really think about the legislation we have an idea we like to put it in but then legislation lawful basis that comes up if you are building something take that into account. Some of the lessons we learned, start small and useful. Write the purpose in one line and do governance early. Lawful basis, threat modelling, keep humans in the loop with an audit trail. Expect the slow lane to be data. It will be data. I can assure you, I know it. finding it and then once you've found it, the permission to use it, cleaning it, and then wiring it into the leg legacy systems. AI can sprint. It really can. But delivery moves at the speed of data and governance. And I'm talking from practical experience. Measure real outcomes. Measure how much time is saved, the error rates, and the user experience. That's what we did when I was listening to this those those recordings. Co-design with users. That's very very important. Design something that the user going to use. And if you co-design with users, they will trust it. If they will trust it, they will really use it. They need to know what the AI is doing. And if they were in the journey designing that, they will have a lot of faith in it. And pilot and learn and then scale and stop really quickly. Fail fast if you want to. But above all, move at the speed of trust. That's very important. If you're starting your journey, there are three things I'm going to finish with. Be ambitious with this AI journey of yours, but be very cautious and be accountable. Thank you.