AI in securities finance: Hype or game-changer?
27 May 2025
More than two years since artificial intelligence became the topic of conversation across industries, Daniel Tison explores how securities finance has embraced this new technology

Since the launch of ChatGPT in late 2022, artificial intelligence has dominated headlines, reshaped workflows, and prompted a wave of investment across various industries. Financial services were quick to follow, with AI being rapidly deployed in fraud detection, client servicing, compliance monitoring, and algorithmic trading.
While early enthusiasm was also apparent in securities finance, the sector's cautious, regulation-bound nature and reliance on complex, relationship-driven processes meant that AI adoption took a more incremental path.
Two years on, the dust is starting to settle, and securities finance firms have begun integrating AI into their processes, particularly where high volumes of structured data or repetitive manual work are involved. However, questions remain about whether this constitutes true transformation or simply a digital upgrade.
Adrian Dale, head of regulation and markets at the International Securities Lending Association (ISLA), says that the increase in AI-related discussions is notable as it highlights a remarkable growth in use cases, adoption, and abilities over just the past two years.
He adds: “As with any innovative technology, the two initial focus topics are ‘what to do with it’ and ‘how to control and manage risks’.”
Ben Challice, president and chief of strategy at Pirum, notes that there has been “a curious mix of hype and hyperbole”, which fits the usual pattern of industries facing technological change.
“With AI, the hype is warranted, but the hyperbole stems from a fundamental obstacle for any meaningful adoption of AI," he adds.
Marton Szigeti, head of collateral, lending, and liquidity solutions at Clearstream, describes AI as a powerful tool, with an enormous amount of potential and a certain lack of certainty around real-world use cases. However, his company has been testing the waters for quite some time to identify the best capabilities.
“Before ChatGPT became popular, we were working on AI for about five years,” says Szigeti. “We've invested constantly over a long period of time because we viewed this as a real bottleneck in the industry.”
Jonathan Lee, senior regulatory reporting specialist at Kaizen, adds that the securities finance industry is inherently a people-oriented business, which leads to reservations, scepticism, and concerns about the potential impact of AI on the human element and job security.
“Fintechs are predominantly best-placed for early adoption, seeking ways to improve the client experience, streamline services, controls, tools, and functions,” says Lee. “This may involve service enhancements, fewer manual touchpoints, and freeing up time to provide clients with greater expert interactions and support. However, there will be adoption across all types of clients, with pockets in each that will be more advanced.”
On the learning curve
In terms of practical solutions, the industry is still “at the first part of the learning curve”, according to Challice, evaluating use cases to understand what the best space for AI is. “What is clear is that not every scenario has AI baked in,” says Lee.
According to Dale, securities borrowing and lending (SBL) represents an ideal candidate for AI use cases that increase productivity, bringing return on investment (ROI) for organisations.
“AI has been recognised by financial markets through its ability to gather broad and complex data sets to then either generate readable output or take programmed actions,” says Dale.
He notes that the most openly discussed use cases in recent months have been the processing of trade communications, analysis of legal documentation, and more expansive collateral optimisation techniques.
“A more advanced use of AI, taking the place of what was referred to as ‘algorithms’, is beginning to be seen in applications used by SBL markets as well,” Dale adds.
The real power of AI, according to Challice, is gathering, manipulating, and analysing big data, enterprise-wide, and then generating evidence-based, strategic, and tactical recommendations.
He elaborates: “Much of the post-trade workflow constitutes the two parties to the trade gaining visibility of reasons why something is not matching. We are using AI to make recommendations, with embedded confidence levels, to automate the fix, thus improving the client's operating efficiencies.”
Pirum has begun the development of enterprise-wide AI solutions, led by a cross-functional team, including representation from product, legal, DevOps, and sales. This AI pilot team has been researching and developing applications across Pirum’s business and services using the latest toolsets, all within strict guardrails and sandboxes.
“We have taken an approach that first we need to learn the risks, as well as the opportunities, before we can start really leveraging this technology,” Challice remarks.
Szigeti breaks AI up into different components, with three promising use cases. The first one is focused around large language models (LLM), where Clearstream has developed the Own Selection Criteria with Automated Reasoning (OSCAR) application.
As the first collateral management tool in the market, OSCAR combines several AI techniques, including machine learning, natural language processing, and automated reasoning, to simplify collateral schedules in securities finance. According to Szigeti, it takes setting up a collateral schedule from about two weeks to around three minutes.
He explains: “You log into a terminal and just type ‘I want European government bonds with this duration’. [OSCAR] retrieves the items, which it thinks are going to be eligible, pre-populates the schedule, and then allows you to select counterparties. You can send the collateral schedule to the counterparty, who can then view it to validate eligibility. Then you are free to trade. So, it’s a straight-through process.”
Another component of AI is its ability to read and interpret documentation, as Lee explains: “There are clear sweet spots around analysis of large sets of information, particularly lengthy regulatory, legal, or contractual information.”
He continues: “We don't believe that you can throw the full end-to-end process to an AI and expect anywhere near reasonable results. However, breaking down aspects of the process and using AI as a thought partner or brainstorming tool can be useful.”
To help clients with their queries, Clearstream has created a knowledge base with an AI agent communicating in a natural human language.
“When you have thousands of clients operating in a fairly complicated triparty environment, with 10 to 15 different use cases and lots of different underlying legal agreements, it’s impossible for a single person to have any idea what’s going on,” Szigeti exclaims, “so we use an AI agent to help interrogate that, and that helps us serve our clients much better in real time..”
The third use case Szigeti introduced is still in development, but it should go live in the next few months. Thanks to a partnership with Google as a cloud and AI model provider, Deutsche Börse Group is developing a generative AI solution for collateral optimisation.
“[The solution] can digest all of the eligibility criteria and all of the optimisation rules, and it can look through that. Before doing the optimisation for the client, it can make recommendations and give a picture of what the outcome could look like,” Szigeti describes.
He hopes that this new solution will become an information dashboard for collateral managers and traders at banks, helping them to make better decisions pre execution.
“It reduces the time between running an optimisation strategy and executing the strategy because you can first what the outcome would look like, and then you can just go execute. The chosen scenario will go into the cloud for the optimisation engine, which will execute it for you,” Szigeti adds.
Following this train of thought, Lee notes: “It will be up to organisations to determine how to use these tools effectively and with an ROI. In any case, they will continue to need experts in the loop to shape the direction of the AI usage and validate the output.”
A cure for hallucinations and an extra layer of protection
In May 2024, the European Securities and Markets Authority (ESMA) published a statement for firms using AI technologies to provide investment services to retail clients. The report argues that while AI offers potential benefits to both companies and clients, significant areas of risk need to be considered.
One of the often-proclaimed challenges, when it comes to AI, is accuracy, as large language models tend to ‘hallucinate’ from time to time, with limited self-reflection. However, keeping this in mind, Clearstream has developed its AI model with the aim of removing hallucinations from the outcome.
Szigeti elaborates: “As the AI goes off and does its ‘thinking’, it revalidates the inputs with the user, and then it creates the actual algorithm. As the trader is setting the rules for optimisation, there’s an interaction between the trader and AI, where the AI validates the instructions down to a point where it’s far more certain around what it’s trying to achieve.”
He believes that AI is only as good as the input data it receives. However, based on the testing his company has done, Szigeti is confident that this AI model is “very accurate".
Other concerns about AI are often around privacy and data security. AI systems need large datasets to learn patterns, which can lead to the exposure of personal or sensitive information. This issue is amplified by the risk of the data being misused, leaked, or targeted by hackers.
“A solid data foundation that can be securely connected will be essential to leveraging the power of AI,” Lee states.
Szigeti explains that his company’s AI model is only contained within the Clearstream environment, working with data that is available anyway, and it is protected by the firm’s data security standards.
“We have built a data layer, which sits between us and our clients,” says Szigeti. “What it does on the client side is aggregation and anonymisation, and then it securely gives us the anonymised, aggregated data that our algorithm will understand and can process.”
Nevertheless, Szigeti admits that the uptake is quite slow, as every client needs to adapt its entire operating model to use the technology, which has been an “enormous challenge”.
He states: “The tool is super flexible and very dynamic in the way that it operates. That isn't the real problem. It can be configured.
“The problem is allowing it to be interfaced into the training desk in a way that it doesn't somehow collapse this delicately balanced framework that every bank has in place.”
Dale sees the main challenge in control and resilience: “Firstly, the correlation between increasing volumes through the use of technology and potential impact and risk from failure.
“Secondly, the use of AI technology to increase complexity can direct markets further away from recognised industry standards. Without the strong foundation that standards represent, built through years of consensus, markets could be navigating into a problematic future that would require perhaps more AI to untangle.”
Lee points out that the securities finance business is based upon personal relationships, which could be the key barrier to the widespread adoption of AI.
“While certain parties hold significant influence over pricing, liquidity, expertise, and service provision, the industry is also characterised by collaboration and shared goals,” he says.
A playing field of ‘first-movers’ and ‘laggards’
From data analysis to predictions, there are several promising use cases where AI can significantly enhance existing processes within the securities finance industry.
Szigeti comments: “There's an enormous opportunity for AI to achieve things that other legacy solutions have tried to achieve, but not necessarily delivered.”
Driving interaction, along with education, will be critical to facilitate a widespread adoption of AI within the industry, according to Szigeti, who also believes that human supervision will always be needed.
Looking ahead, Challice sees the adoption of AI as a playing field dominated by “first-movers”, who implement it in a structured, enterprise-wide model, and “laggards”, who will wait to see how it plays out.
“But there will also be those that see the opportunity and understand the risk, but who don’t currently have a modern enough tech stack to leverage AI beyond manually using ChatGPT,” adds Challice. “This last camp is now in the process of upgrading its tech.”
Szigeti expects to see a broad implementation of AI by banks and other financial institutions within the next few years.
He adds: “Over time, AI will shift how collateral optimisation and securities finance work because it will create transparency and speed up decision-making in the way people operate.”
On that note, Dale states: “It would be easy to say that if that trajectory continues, financial markets, in step with many aspects of daily activities, will see profound changes over the next five years. This time horizon has been echoed at several conferences this year, though we pragmatically recognise the cautious approach of the financial sector.”
He adds that ISLA is watching this space very carefully and engaging with member firms to ensure standards are followed and that communication between the market and its supervisors reflects the opportunities offered by this new technology.
Lee believes that there may be opportunities for firms to optimise securities financing across products to create a “perfect market” that arbitrages existing differences in repo rates, lending fees, and rebate rates, without differences in balance sheet treatment that exist today. This could lead to cheaper borrowing costs over time, as greater transparency reduces differences in transaction rates.
He concludes: “Ultimately, AI is likely to break down the barriers to collateral management and funding, creating a more level playing field.”
While early enthusiasm was also apparent in securities finance, the sector's cautious, regulation-bound nature and reliance on complex, relationship-driven processes meant that AI adoption took a more incremental path.
Two years on, the dust is starting to settle, and securities finance firms have begun integrating AI into their processes, particularly where high volumes of structured data or repetitive manual work are involved. However, questions remain about whether this constitutes true transformation or simply a digital upgrade.
Adrian Dale, head of regulation and markets at the International Securities Lending Association (ISLA), says that the increase in AI-related discussions is notable as it highlights a remarkable growth in use cases, adoption, and abilities over just the past two years.
He adds: “As with any innovative technology, the two initial focus topics are ‘what to do with it’ and ‘how to control and manage risks’.”
Ben Challice, president and chief of strategy at Pirum, notes that there has been “a curious mix of hype and hyperbole”, which fits the usual pattern of industries facing technological change.
“With AI, the hype is warranted, but the hyperbole stems from a fundamental obstacle for any meaningful adoption of AI," he adds.
Marton Szigeti, head of collateral, lending, and liquidity solutions at Clearstream, describes AI as a powerful tool, with an enormous amount of potential and a certain lack of certainty around real-world use cases. However, his company has been testing the waters for quite some time to identify the best capabilities.
“Before ChatGPT became popular, we were working on AI for about five years,” says Szigeti. “We've invested constantly over a long period of time because we viewed this as a real bottleneck in the industry.”
Jonathan Lee, senior regulatory reporting specialist at Kaizen, adds that the securities finance industry is inherently a people-oriented business, which leads to reservations, scepticism, and concerns about the potential impact of AI on the human element and job security.
“Fintechs are predominantly best-placed for early adoption, seeking ways to improve the client experience, streamline services, controls, tools, and functions,” says Lee. “This may involve service enhancements, fewer manual touchpoints, and freeing up time to provide clients with greater expert interactions and support. However, there will be adoption across all types of clients, with pockets in each that will be more advanced.”
On the learning curve
In terms of practical solutions, the industry is still “at the first part of the learning curve”, according to Challice, evaluating use cases to understand what the best space for AI is. “What is clear is that not every scenario has AI baked in,” says Lee.
According to Dale, securities borrowing and lending (SBL) represents an ideal candidate for AI use cases that increase productivity, bringing return on investment (ROI) for organisations.
“AI has been recognised by financial markets through its ability to gather broad and complex data sets to then either generate readable output or take programmed actions,” says Dale.
He notes that the most openly discussed use cases in recent months have been the processing of trade communications, analysis of legal documentation, and more expansive collateral optimisation techniques.
“A more advanced use of AI, taking the place of what was referred to as ‘algorithms’, is beginning to be seen in applications used by SBL markets as well,” Dale adds.
The real power of AI, according to Challice, is gathering, manipulating, and analysing big data, enterprise-wide, and then generating evidence-based, strategic, and tactical recommendations.
He elaborates: “Much of the post-trade workflow constitutes the two parties to the trade gaining visibility of reasons why something is not matching. We are using AI to make recommendations, with embedded confidence levels, to automate the fix, thus improving the client's operating efficiencies.”
Pirum has begun the development of enterprise-wide AI solutions, led by a cross-functional team, including representation from product, legal, DevOps, and sales. This AI pilot team has been researching and developing applications across Pirum’s business and services using the latest toolsets, all within strict guardrails and sandboxes.
“We have taken an approach that first we need to learn the risks, as well as the opportunities, before we can start really leveraging this technology,” Challice remarks.
Szigeti breaks AI up into different components, with three promising use cases. The first one is focused around large language models (LLM), where Clearstream has developed the Own Selection Criteria with Automated Reasoning (OSCAR) application.
As the first collateral management tool in the market, OSCAR combines several AI techniques, including machine learning, natural language processing, and automated reasoning, to simplify collateral schedules in securities finance. According to Szigeti, it takes setting up a collateral schedule from about two weeks to around three minutes.
He explains: “You log into a terminal and just type ‘I want European government bonds with this duration’. [OSCAR] retrieves the items, which it thinks are going to be eligible, pre-populates the schedule, and then allows you to select counterparties. You can send the collateral schedule to the counterparty, who can then view it to validate eligibility. Then you are free to trade. So, it’s a straight-through process.”
Another component of AI is its ability to read and interpret documentation, as Lee explains: “There are clear sweet spots around analysis of large sets of information, particularly lengthy regulatory, legal, or contractual information.”
He continues: “We don't believe that you can throw the full end-to-end process to an AI and expect anywhere near reasonable results. However, breaking down aspects of the process and using AI as a thought partner or brainstorming tool can be useful.”
To help clients with their queries, Clearstream has created a knowledge base with an AI agent communicating in a natural human language.
“When you have thousands of clients operating in a fairly complicated triparty environment, with 10 to 15 different use cases and lots of different underlying legal agreements, it’s impossible for a single person to have any idea what’s going on,” Szigeti exclaims, “so we use an AI agent to help interrogate that, and that helps us serve our clients much better in real time..”
The third use case Szigeti introduced is still in development, but it should go live in the next few months. Thanks to a partnership with Google as a cloud and AI model provider, Deutsche Börse Group is developing a generative AI solution for collateral optimisation.
“[The solution] can digest all of the eligibility criteria and all of the optimisation rules, and it can look through that. Before doing the optimisation for the client, it can make recommendations and give a picture of what the outcome could look like,” Szigeti describes.
He hopes that this new solution will become an information dashboard for collateral managers and traders at banks, helping them to make better decisions pre execution.
“It reduces the time between running an optimisation strategy and executing the strategy because you can first what the outcome would look like, and then you can just go execute. The chosen scenario will go into the cloud for the optimisation engine, which will execute it for you,” Szigeti adds.
Following this train of thought, Lee notes: “It will be up to organisations to determine how to use these tools effectively and with an ROI. In any case, they will continue to need experts in the loop to shape the direction of the AI usage and validate the output.”
A cure for hallucinations and an extra layer of protection
In May 2024, the European Securities and Markets Authority (ESMA) published a statement for firms using AI technologies to provide investment services to retail clients. The report argues that while AI offers potential benefits to both companies and clients, significant areas of risk need to be considered.
One of the often-proclaimed challenges, when it comes to AI, is accuracy, as large language models tend to ‘hallucinate’ from time to time, with limited self-reflection. However, keeping this in mind, Clearstream has developed its AI model with the aim of removing hallucinations from the outcome.
Szigeti elaborates: “As the AI goes off and does its ‘thinking’, it revalidates the inputs with the user, and then it creates the actual algorithm. As the trader is setting the rules for optimisation, there’s an interaction between the trader and AI, where the AI validates the instructions down to a point where it’s far more certain around what it’s trying to achieve.”
He believes that AI is only as good as the input data it receives. However, based on the testing his company has done, Szigeti is confident that this AI model is “very accurate".
Other concerns about AI are often around privacy and data security. AI systems need large datasets to learn patterns, which can lead to the exposure of personal or sensitive information. This issue is amplified by the risk of the data being misused, leaked, or targeted by hackers.
“A solid data foundation that can be securely connected will be essential to leveraging the power of AI,” Lee states.
Szigeti explains that his company’s AI model is only contained within the Clearstream environment, working with data that is available anyway, and it is protected by the firm’s data security standards.
“We have built a data layer, which sits between us and our clients,” says Szigeti. “What it does on the client side is aggregation and anonymisation, and then it securely gives us the anonymised, aggregated data that our algorithm will understand and can process.”
Nevertheless, Szigeti admits that the uptake is quite slow, as every client needs to adapt its entire operating model to use the technology, which has been an “enormous challenge”.
He states: “The tool is super flexible and very dynamic in the way that it operates. That isn't the real problem. It can be configured.
“The problem is allowing it to be interfaced into the training desk in a way that it doesn't somehow collapse this delicately balanced framework that every bank has in place.”
Dale sees the main challenge in control and resilience: “Firstly, the correlation between increasing volumes through the use of technology and potential impact and risk from failure.
“Secondly, the use of AI technology to increase complexity can direct markets further away from recognised industry standards. Without the strong foundation that standards represent, built through years of consensus, markets could be navigating into a problematic future that would require perhaps more AI to untangle.”
Lee points out that the securities finance business is based upon personal relationships, which could be the key barrier to the widespread adoption of AI.
“While certain parties hold significant influence over pricing, liquidity, expertise, and service provision, the industry is also characterised by collaboration and shared goals,” he says.
A playing field of ‘first-movers’ and ‘laggards’
From data analysis to predictions, there are several promising use cases where AI can significantly enhance existing processes within the securities finance industry.
Szigeti comments: “There's an enormous opportunity for AI to achieve things that other legacy solutions have tried to achieve, but not necessarily delivered.”
Driving interaction, along with education, will be critical to facilitate a widespread adoption of AI within the industry, according to Szigeti, who also believes that human supervision will always be needed.
Looking ahead, Challice sees the adoption of AI as a playing field dominated by “first-movers”, who implement it in a structured, enterprise-wide model, and “laggards”, who will wait to see how it plays out.
“But there will also be those that see the opportunity and understand the risk, but who don’t currently have a modern enough tech stack to leverage AI beyond manually using ChatGPT,” adds Challice. “This last camp is now in the process of upgrading its tech.”
Szigeti expects to see a broad implementation of AI by banks and other financial institutions within the next few years.
He adds: “Over time, AI will shift how collateral optimisation and securities finance work because it will create transparency and speed up decision-making in the way people operate.”
On that note, Dale states: “It would be easy to say that if that trajectory continues, financial markets, in step with many aspects of daily activities, will see profound changes over the next five years. This time horizon has been echoed at several conferences this year, though we pragmatically recognise the cautious approach of the financial sector.”
He adds that ISLA is watching this space very carefully and engaging with member firms to ensure standards are followed and that communication between the market and its supervisors reflects the opportunities offered by this new technology.
Lee believes that there may be opportunities for firms to optimise securities financing across products to create a “perfect market” that arbitrages existing differences in repo rates, lending fees, and rebate rates, without differences in balance sheet treatment that exist today. This could lead to cheaper borrowing costs over time, as greater transparency reduces differences in transaction rates.
He concludes: “Ultimately, AI is likely to break down the barriers to collateral management and funding, creating a more level playing field.”
NO FEE, NO RISK
100% ON RETURNS If you invest in only one securities finance news source this year, make sure it is your free subscription to Securities Finance Times
100% ON RETURNS If you invest in only one securities finance news source this year, make sure it is your free subscription to Securities Finance Times
