Automatable processes in the sector are often rules-based, involving collecting data, validating it against internal and external databases, and assembling required documentation. In theory, conventional software has been used to automate such work. However, Marco Argenti, Goldman’s chief information officer, argues that if a rules-based system resolves most cases, a small percentage of transactions fall outside defined parameters that can translate into thousands of individual items at the type of scale in question. He cites the example of identity verification in KYC compliance, where minor discrepancies or documents approaching expiry can create edge cases requiring judgement.

Argenti says that neural networks can address these micro-decisions as they’re capable of applying contextual reasoning where fixed rules might be missing or don’t necessarily give a clear answer. In this scenario, generative AI augments existing rules systems rather than supplanting them. Operational improvements, therefore lie in the reduction in the number of cases that require manual intervention and thus shortening time needed to resolve the exceptions.

Goldman’s prior experience with Claude models used internally for software development informed its decision to extend AI to other areas of operations. Developers use a version of Claude with Cognition’s Devin agent to aid them with programming. In this context, human developers set specifications and regulatory parameters, the agent produces code, and humans review outputs. The agent is also used to run code tests and validations. He describes this as a change to devs’ workflows, with agents operating according to defined instructions. The benefit is increased developer productivity and the faster completion of projects.s

For trade accounting and client onboarding, Goldman and Anthropic AI project owners observed existing workflows with domain experts to identify work bottlenecks. The implemented agents review documents, extract entities, determine whether additional documentation is required, assess ownership structures, and can trigger further compliance checks. Tasks automated in this way tend to be document-heavy and require individual judgement. By automating extraction and preliminary assessment, the agents reduce the time analysts spend on comparison work.

Indranil Bandyopadhyay, principal analyst at Forrester, says that reconciliation in trade accounting requires comparing fragmented data in internal ledgers, counterparty confirmations, and the perusal of bank statements, and that a typical workflow depends on accurate extraction and matching of figures and text to existing documents. Claude’s ability to process large context windows and follow instructions, he says, makes it suited to just such workflows. The labour involved in client onboarding, such as parsing passports and corporate registration documents, and the cross-referencing of all sources means AI’s ability to extract structured data and flag inconsistencies makes the technology a good fit, reducing overall workloads.

Bandyopadhyay stresses that accounting and compliance platforms remain the canonical systems of record. Claude operates in the workflow layer, handling extraction and comparison so human analysts can handle the code’s exceptions.. In his assessment, the operational value in a regulated environments like banking lies in such a division of labour.

Jonathan Pelosi, head of financial services at Anthropic says Claude is trained to surface uncertainty and to provide source attribution, creating an audit trail – reducing the effect of hallucinations. Bandyopadhyay also notes the importance of human oversight and validation, saying institutions should design systems so that errors are detected early.

Goldman’s Marco Argenti rejects the view that AI systems are inherently easier to deceive than people, arguing that social engineering exploits human vulnerabilities and that AI can detect subtle anomalies at scale, and reiterates the need to combine human judgement with automated scrutiny in teams. His claim implies a increase in operational capacity without proportional increases in staff, even with the issues known to affect AI rollouts.

AI in banking operations

In the banking sector, generative AI is a tool that improves operational performance by accelerating document processing, reducing exception handling time, and increasing throughput in high volume workflows. But the need to retain human oversight to counteract AI’s errors means the retention of and reliance on existing systems of records remains.

(Image source: “Dreams…” by noahwesley is licensed under CC BY-NC-SA 2.0)

Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and co-located with other leading technology events. Click here for more information.

AI News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.