Interviewer: Hello and welcome to The AI Summit London. Today we have the pleasure of being with Sudeshna Sen, Data Product Portfolio Manager at the Financial Services Compensation Scheme. For people at home, please give us a little insight into your background and your fifteen years of experience.

Sudeshna: I am a data and AI product strategist and I have been building data and AI products for the better part of fifteen years, since the days when it was called data science rather than AI. That is how I landed in financial services. I have worked with Deloitte, PwC Strategy, private equity, and financial regulators throughout my career, building data and AI products for all of them.

Interviewer: With your experience, what drew you into the field of AI, and how has your approach evolved over time?

Sudeshna: I am an economist by training and very quantitative, with a finance specialisation. That drew me in because finance was doing much of the heavy lifting in data and quantitative science. Over the years, data science has gradually evolved into what we now call AI. The quantitative side attracted me, but what has kept me here is the human side of the technology. Today we are not limited to finance. We talk about transformation of technology and also transformation of people. The quantitative work brought me in, and the human element keeps me excited.

Interviewer: What use cases within the AI space are you seeing where businesses are achieving significant results?

Sudeshna: I will mention two from my career. One was for a financial services regulator, where we reduced customer wait times from years to months, and in some cases weeks. It is one thing to say you have a productivity gain from AI, but when that gain translates into better customer experience and higher satisfaction, you are delivering on the promise of AI. We achieved that with an assisted AI system that was reliable by design. Another piece of work began around eight years ago with BSI and ISO that has become ISO 42001. I remember waking at inconvenient hours to bring together an international committee of experts to work through the standard. I feel grateful that we could contribute to such an important piece of standard making. It is one of the most impactful projects of my career.

Interviewer: As part of the United Kingdom’s AI expert pool for BSI and ISO, what are your thoughts on the role of industry standards in ensuring safe and ethical AI deployment?

Sudeshna: There are still many conversations to have because we are at an early stage. ISO 42001 has emerged from a long period of work, and BSI will continue to certify and align the elements. Where we need more discussion is in educating leaders across industries about what AI and machine learning are, where they can go wrong, and why we need to treat AI differently from traditional software. AI is dependent on the data and the use case at hand, so every problem is different. There is no one size fits all. The right guardrails come from understanding those differences.

Interviewer: There is a growing conversation about using smaller and sharper models instead of larger scale systems. What are the advantages of smaller models and where might they be more effective?

Sudeshna: I often prefer smaller models because they can deliver disproportionate return on investment. They are less expensive to run and do not require such large infrastructure. They can also be more explainable. I am not only talking about smaller language models. There are decision tree models that can deliver excellent results, with high explainability and without a black box, and they can be deployed in regulated industries, where larger models can struggle.

Interviewer: During the panel on trustworthiness, the importance of data grounding and context was highlighted. How can organisations ensure that AI outputs remain trustworthy and contextually relevant?

Sudeshna: One theme from the panel on grounding was the risk of over trust and under trust. Models can appear so polished that users over trust them, only to discover later that the answer was not right. At the same time, if users do not have trust, engagement declines. The way to address this is to bring users on the journey. Make it clear that the system is a machine and a thinking partner, not infallible. Help users understand where the model is strong and where it is weak. Then they will know when to trust the model and when not to.

Interviewer: What are the early signs of trust erosion in AI systems, and how can teams proactively address these issues to maintain confidence in their solutions?

Sudeshna: A drop in user engagement is an early signal that trust is eroding, and that is not good for any product. To address it, show people the work. Do not only present the output. Show how you arrived at it. Many reasoning focused approaches now present the steps taken or the evidence used. With smaller models and data sets you can often go even deeper. It is important to present how you reached the answer, alongside the answer itself.

Interviewer: If you could give one piece of advice to businesses looking to adopt AI as a strategic priority, what would it be?

Sudeshna: In the RPA days we looked for repeatable actions. Apply the same idea to AI and look for repeated decisions. Where do you make the highest volume of similar decisions. In financial services that could be credit scoring, claims management, or loan disbursement. Identify those areas, then understand what is needed to make those decisions. How do humans make them. There is usually a checklist or at least a set of principles. Encode that into your AI model. That is where you can get disproportionate results.

Interviewer: Looking ahead to the next twelve months, what trends in AI do you think will be having the biggest impact?

Sudeshna: I hope to see trust and explainability catch up. AI has come a long way in the last ten to twelve years to the point that it is almost recognisable from where it began, and frontier AI labs will continue to push the technology forward. We now need to bring leaders along that curve, help them understand the critical aspects of these models and where they can go wrong, and build governance structures within organisations to manage those risks responsibly.

Interviewer: How have you enjoyed the AI Summit in London so far, what's your experience been, and what's one of the biggest takeaways you'll be taking away from today?

Sudeshna: I’ve attended The AI Summit London for the past four or five years, and it’s been inspiring to watch it grow. It’s a lot to take in over a couple of days, but what I love most echoes what I heard from the headliners yesterday: practitioners learning from practitioners.

There are no egos, no so‑called “experts”—just people learning from one another and pushing the frontier together. The networking is excellent, and everyone is friendly and helpful.

Closing

Thank you to Sudeshna Sen for sharing practical insights on building trustworthy AI. 

Her focus on explainability, user engagement, and standards such as ISO 42001 offers a clear path for organisations that want to realise value from AI while protecting users and maintaining confidence.