GenAI: Custody’s new clothes?

DEEP LOOK | While financial markets embrace AI as must-have tech, the securities services industry faces a sizing problem. Targeted GenAI use cases are emerging across post-trade processes, but broader adoption lags behind – constrained by legacy systems, fragmented infrastructure, and data quality challenges that enterprise deployments rarely face.  

One heading from Citi’s latest annual Securities Services Evolution whitepaper stands out a mile from a post-trade perspective. In the section on generative AI or Generative AI (GenAI), the sub-heading reads ‘‘Everyone is using it – but not in post-trade’.

Only 7% of the firms surveyed said they were feeling the impact of GenAI in post-trade, middle office and operations, despite 57% and 24% respectively claiming that the technology was either being piloted or had gone live in the back office.

To reconcile these apparently contradictory findings we asked some of the leading securities services providers to outline their GenAI strategy and particularly how it impacts post-trade processes.

Advertisement

GenAI deployment in post-trade processes

BNP Paribas is experimenting in selected operational domains such as manual input automation, exception analysis and documentation workflows, observes Yulia Shamsudinova, the bank’s head of AI transformation & process intelligence, securities services.

“As the technology matures and regulatory clarity around GenAI in securities services continues to strengthen we expect progressive carefully supervised expansion into broader post-trade processes,” she adds.

GenAI is already being applied to some aspects that touch post-trade reconciliation, reporting, clearing and settlements at Citi, for example to support MT202 processing review explains Tod McKenna, head of AI client engagement & solutions for Citi Services.

CACEIS’s initial tests focused on streamlining its tender management process to draft business proposals before progressing to optical character recognition combined with robotic process automation to facilitate data entry tasks and then on to handling operational tasks such as dealing with reconciliation management issues. Pirum is also focused on optical character recognition improvements and structured data extraction.

“Automation around new GenAI tools relies on the availability of accurate, real-time, golden source data so we continue to be focused on providing real-time APIs that enable clients to develop their own agents and GenAI models,” observes Mack Gill, head of securities processing at FIS.

Finastra has initiated several back office projects including AI. For example, it is now able to transform the payment notification that clients receive from their counterparty into structured data that can be imported and automatically validate settlement with the counterparty before releasing the payment.

Another project relates to the management of settlement instructions, which can be a lengthy process for banks that still have part of their settlement instructions created manually from various sources, such as Excel spreadsheets.

A third example facilitates the integration and interpretation of external data, such as paper confirmations from counterparties. It enables clients to move away from manual validation and automate the matching of paper confirmations.

Obstacles to further adoption

Various factors contribute to the lag in GenAI adoption in post-trade compared to wider enterprise adoption identified in the Citi whitepaper, including system fragmentation, regulation and policy and lack of data connectivity and access across multiple systems.

“The biggest obstacles we currently see are a dependence on the right foundations and needing to incorporate agentic [self-governing] design principles into process re-engineering and re-platforming initiatives,” says Ying Ying Tan, global head of product management for financing and securities services at Standard Chartered.

She suggests it is misleading to say ‘all you need is good data for good AI’ and that to leverage AI – especially agentic AI – across the asset servicing lifecycle and operational processes requires a service orchestration layer that can act as an environment for artificial intelligence to operate within.

“Even with the right data and service foundations and the right AI models and infrastructure, if existing processes and platforms are not designed to support the agentic workflows of the future, that will remain a major obstacle for the scalable deployment of these AI capabilities,” adds Tan.

Kevin Welch, who leads the chief administrative and transformation office within service delivery at Brown Brothers Harriman adds people and organisational issues to the list of obstacles to deployment.

“A common trap is the ‘scattered pilot problem’,” he says. “Individual teams run successful proofs of concept but without a centralised AI function, those wins never scale. The firms that we see moving the fastest are using hub-and-spoke models where a central hub sets governance, guardrails and reusable components and business teams adapt them into workflows and feed improvements back.”

Welch adds that firms are generally discovering that their pilots are effective from a technology perspective but adoption may lag if their teams aren’t trained to use AI effectively. He makes the point that without strong change management, AI becomes a set of experiments rather than living up to its potential as a transformative technology.

The legacy systems that still dominate the industry pose a significant hindrance to AI adoption notes Ben Challice, CEO Pirum. “Clients that have real-time, standardised, enterprise data can enter the next AI augmented stage, where it is integrated into workflows, speeding things up while humans are still in the loop,” he says. “The quantum leap will be when autonomous AI agents are trained and begin to make decisions within pre-defined parameters and human oversight but the condition for getting there is standardised, real-time data. For that, automation is the first and most important step.”

McKenna also refers to fragmented and legacy financial infrastructures as well as the inherent black box nature of GenAI models, which complicates transparency, accountability and the identification of security vulnerabilities.

“Stringent regulatory compliance requirements – particularly concerning data privacy, security and explainability – and bias detection also pose substantial hurdles,” he says. “Furthermore, challenges related to ensuring high data quality and preventing bias amplification, alongside the operational complexities of integrating GenAI into existing systems, contribute to cautious adoption. Addressing these multifaceted challenges necessitates robust risk management frameworks and careful strategic planning.”

The requirement to meet high standards of accuracy, data integrity, explainability and regulatory compliance has encouraged institutions to adopt a staged approach to manage controlled pilots under robust validation and alignment with supervisory expectations according to Shamsudinova.

“We value GenAI as part of a wider set of tools to enable automation alongside traditional AI, agentic, robotic process automation and other available technology,” she says. “Gen AI might not necessarily be the right solution for everything so therefore we constantly and carefully assess each technology to ensure that we bring forward the most appropriate solution.”

Future uses of GenAI in post-trade

Looking at potential future developments, Xavier Crepin-Leblond, head of product management back office for Finastra Treasury & Capital Markets says there will be an opportunity for the technology to be used to help reduce settlement timeframes in a T+1 environment – firstly by leveraging AI to validate the settlement instructions and secondly, by helping banks make adjustments to the settlement instructions attached to the trade.

Leveraging AI to interpret chat and email between parties will help speed up the resolution of trade breaks.

Tan expects more tangible use cases for GenAI to emerge over the next 6-12 months in the form of AI agents or across larger and more complex workflows. “However, it will take more than 12 months to scale these solutions beyond pilots or limited scope initiatives, as both clients and regulators will need to be brought along this journey and the necessary governance frameworks must be extended to support such usage of AI,” she acknowledges. “Furthermore, operating models must be revamped to ensure these agentic designs and workflows are sustainable,” adds Tan.

As firms gain confidence, the technology is moving deeper into the post-trade lifecycle in areas such as reconciliations, execution management, reporting and clearing and settlements, says Welch.

“GenAI now has the capabilities to perform tasks like reconciling structured versus unstructured data sources, extracting key fields from prospectuses or notices and pairing them with transaction data, which has the potential to dramatically reduce manual breaks,” he says. While more tightly governed and interconnected across market infrastructures, Welch notes that early prototypes are emerging in areas such as instruction validation, mismatch explanation and anomaly detection.

“The timing of broader roll out will depend on data readiness (GenAI scales best on top of high quality, well-structured data and connected systems and many firms are still modernising these foundations) and governance maturity,” he continues.

Regulatory expectations are driving firms to strengthen oversight, model controls and responsible use frameworks before expanding into systemically critical functions, adds Welch. “Given this combination, we should expect to see broad but phased adoption, where automation of complex post-trade workflows becomes mainstream within the next few years.”

Arnaud Misset, chief digital officer of CACEIS Group says his organisation is already there, making in-house IT developments using available large language models or proceeding with the acquisition of a vendor’s existing solution.

“For example, our middle office reconciliation tool was the result of a close collaboration project with Nephelai, a SaaS solution that uses AI for tasks such as  automating confirmations, settlement and CSDR penalty tracking,” he says.