SFS: AI adoption in securities finance hinges on data foundations and human accountability
25 September 2026 UK
Image: SFT
The panel, moderated by Martin Walker, senior product manager at FIS, provided an insight into the practical uses, governance requirements, and adoption challenges for AI in financial market processes.
Andy Pardoe, professor of AI, started off the discussion by explaining the greater picture looking back on the evolution of AI. 鈥淭he technologies and techniques we have today are very different to what AI was considered in the 90s or the 70s.鈥�
He suggested that there are two main categories of AI. Symbolic AI 鈥� 鈥渢he old鈥憇chool expert systems and rule engines鈥� 鈥� and the more modern connectionist approaches, 鈥渨hich are neural networks, deep learning and LLM鈥憈ype tools鈥�. They have strengths in different areas, he said.
Pardoe explained that model selection should reflect task complexity. 鈥淵ou can often solve something with a simple AI method, and that鈥檚 usually the best option. Mapping the complexity of the problem to the complexity of the solution 鈥� you don't need to use the biggest LLM for every problem.鈥�
Walker then asked how AI relates to securities lending. Richard Colvill, project manager at Reg-X, said AI is often seen as 鈥渁 solution to a problem we don鈥檛 yet know we have鈥�, but its operational value is clear. 鈥淭raders seldom see all the components that get a transaction to market.
鈥淢ost firms focus on the headline P&L, but there are many operational costs underneath 鈥� multiple custodians, platform fees, SWIFT costs, and post鈥憈rade regulatory reporting charges.鈥� He said AI could 鈥渉elp define and analyse鈥� those costs.
Ernst Dolce, CEO and co-founder of Banqora, highlighted that client demand is focused on lifecycle management, collateral management and optimisation, visibility, and data capture. 鈥淐lients want to bring actionable insights from their data, including from intelligence from their middle and back-office operations.
鈥淚n securities finance, booking the trade was never the challenge 鈥� most of the
information is already there: accessing all the info from your email, your chats,
internal, and service provider systems is the hard part. AI is very good at capturing information that is stored everywhere and exploiting it.鈥�
He said firms want to understand the true cost of their transactions. 鈥淚f you have three counterparties with the same credit risk and credit limit, all giving you the same fee rate, they are not equivalent in terms of P&L once you factor in
settlement-fail penalties, funding costs, balance-sheet and capital treatment, and operational overhead. You can use the power of AI to see that.鈥�
Pierre Khemdoudi, CEO and co-founder of Gentek AI, expressed that mapping remains one of the biggest barriers to adoption. 鈥淭he main challenge remains having systems connected to each other.
鈥淧utting everything into a single data lake costs a lot of money and requires huge resources. Smaller organisations cannot do that.鈥�
He said firms need a connective layer before automation. 鈥淲hatever you do afterwards 鈥� transaction reporting, collateral management, decision鈥憁aking 鈥� you need to do that first.鈥�
Khemdoudi also highlighted the need for model orchestration. 鈥淪ome models are very expensive, some are very cheap. Some are good at text, some at pictures, some at numbers. You need a layer that not only connects systems but orchestrates which model you use to be efficient in terms of cost and capability.鈥�
In terms of risks, Pardoe stated that most firms鈥� data is 鈥渁n absolute mess鈥�, and AI agents require guardrails. 鈥淎I can map the data today, but it can also look at how the data evolves and remap based on what it sees.
鈥淲e need the infrastructure underneath that guides, governs, and controls the agents. We鈥檝e already seen cases where agents go out of their sandbox.鈥�
Dolce said mistakes are inevitable but scale changes the risk. 鈥淗uman beings make mistakes. What changes with AI is the scale. There is a risk, but the way you deploy it matters. You will always have human interaction with AI.鈥�
Colvill stated that oversight does not require large teams. 鈥淵ou don鈥檛 need very many people sitting on top of a bot. Fair鈥慳llocation algorithms are an area that needs to be looked at with AI. I can鈥檛 honestly say they are fair for many smaller beneficial owners. AI could offer much鈥憂eeded transparency.鈥�
The panel agreed that AI is not a 鈥渕agic鈥� solution. It requires data foundations, governance, critical thinking and domain expertise. Adoption is expected to grow, but firms must choose the right models, apply them to the right problems, and understand the workflows they are trying to improve.
Andy Pardoe, professor of AI, started off the discussion by explaining the greater picture looking back on the evolution of AI. 鈥淭he technologies and techniques we have today are very different to what AI was considered in the 90s or the 70s.鈥�
He suggested that there are two main categories of AI. Symbolic AI 鈥� 鈥渢he old鈥憇chool expert systems and rule engines鈥� 鈥� and the more modern connectionist approaches, 鈥渨hich are neural networks, deep learning and LLM鈥憈ype tools鈥�. They have strengths in different areas, he said.
Pardoe explained that model selection should reflect task complexity. 鈥淵ou can often solve something with a simple AI method, and that鈥檚 usually the best option. Mapping the complexity of the problem to the complexity of the solution 鈥� you don't need to use the biggest LLM for every problem.鈥�
Walker then asked how AI relates to securities lending. Richard Colvill, project manager at Reg-X, said AI is often seen as 鈥渁 solution to a problem we don鈥檛 yet know we have鈥�, but its operational value is clear. 鈥淭raders seldom see all the components that get a transaction to market.
鈥淢ost firms focus on the headline P&L, but there are many operational costs underneath 鈥� multiple custodians, platform fees, SWIFT costs, and post鈥憈rade regulatory reporting charges.鈥� He said AI could 鈥渉elp define and analyse鈥� those costs.
Ernst Dolce, CEO and co-founder of Banqora, highlighted that client demand is focused on lifecycle management, collateral management and optimisation, visibility, and data capture. 鈥淐lients want to bring actionable insights from their data, including from intelligence from their middle and back-office operations.
鈥淚n securities finance, booking the trade was never the challenge 鈥� most of the
information is already there: accessing all the info from your email, your chats,
internal, and service provider systems is the hard part. AI is very good at capturing information that is stored everywhere and exploiting it.鈥�
He said firms want to understand the true cost of their transactions. 鈥淚f you have three counterparties with the same credit risk and credit limit, all giving you the same fee rate, they are not equivalent in terms of P&L once you factor in
settlement-fail penalties, funding costs, balance-sheet and capital treatment, and operational overhead. You can use the power of AI to see that.鈥�
Pierre Khemdoudi, CEO and co-founder of Gentek AI, expressed that mapping remains one of the biggest barriers to adoption. 鈥淭he main challenge remains having systems connected to each other.
鈥淧utting everything into a single data lake costs a lot of money and requires huge resources. Smaller organisations cannot do that.鈥�
He said firms need a connective layer before automation. 鈥淲hatever you do afterwards 鈥� transaction reporting, collateral management, decision鈥憁aking 鈥� you need to do that first.鈥�
Khemdoudi also highlighted the need for model orchestration. 鈥淪ome models are very expensive, some are very cheap. Some are good at text, some at pictures, some at numbers. You need a layer that not only connects systems but orchestrates which model you use to be efficient in terms of cost and capability.鈥�
In terms of risks, Pardoe stated that most firms鈥� data is 鈥渁n absolute mess鈥�, and AI agents require guardrails. 鈥淎I can map the data today, but it can also look at how the data evolves and remap based on what it sees.
鈥淲e need the infrastructure underneath that guides, governs, and controls the agents. We鈥檝e already seen cases where agents go out of their sandbox.鈥�
Dolce said mistakes are inevitable but scale changes the risk. 鈥淗uman beings make mistakes. What changes with AI is the scale. There is a risk, but the way you deploy it matters. You will always have human interaction with AI.鈥�
Colvill stated that oversight does not require large teams. 鈥淵ou don鈥檛 need very many people sitting on top of a bot. Fair鈥慳llocation algorithms are an area that needs to be looked at with AI. I can鈥檛 honestly say they are fair for many smaller beneficial owners. AI could offer much鈥憂eeded transparency.鈥�
The panel agreed that AI is not a 鈥渕agic鈥� solution. It requires data foundations, governance, critical thinking and domain expertise. Adoption is expected to grow, but firms must choose the right models, apply them to the right problems, and understand the workflows they are trying to improve.
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 网爆吃瓜 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 网爆吃瓜 Finance Times
