After years of experimentation, financial firms are beginning to put generative artificial intelligence into production, but some applications are proving easier to deploy than others.
Executives from Bank of America, BNP Paribas, Nasdaq and Société Générale discussed what is working, where the technology falls short and how they are managing the risks at FIA Expo in Chicago on 3 November.
Société Générale has used generative AI to accelerate software development and it has applied it to trade-break investigations. BNP Paribas has found a use for it in reconciling structured and unstructured information, while Bank of America has used it to turn email-based workflows into structured data. Nasdaq, meanwhile, is developing agents that could help back-office employees decide which tasks to tackle first.
Dev Patel, head of post-trade services technology at Société Générale, described a project involving the conversion of FIX messages from a legacy format into a new format. Rather than keep paying a vendor for mapping functionality, the bank used AI to build its own. The resulting production asset was conventional code owned by Société Générale, with the LLM removed from the critical path, he said.
“What is interesting about this specific use case is that it is not really the AI itself, but AI enabled us to produce conventional software significantly more efficiently,” Patel said. “The production asset at the end is code that you own,” he added. “We are using AI to dramatically accelerate the engineering required to remove the vendor from the critical path.”
Patel said the bank was also looking at applying AI to trade breaks in prime services, using upstream and downstream information to support investigations and recommend actions. He said the bank hopes to prevent breaks rather than simply deal with them after they occur.
At BNP Paribas, Marianne Maurel, head of markets operations, transversal services Americas, pointed to reconciliation between structured and unstructured data as an area where AI had made “a real difference very quickly”.
A booking system might contain one version of a counterparty name, for example, while an email contains another. AI can extract information from the text and recognise that both refer to the same entity, helping automate a process that might otherwise require human interpretation.
Bank of America has found similar value in turning information designed for people into data that machines can use. Ian Grubb, director of EMEA client solutions for futures and options and OTC clearing, described one mortgage-market application as a “glorified email parser”.
It takes a legacy email flow of bids wanted in competition and turns the information into a structured data set that can be processed more efficiently. “It has radically accelerated their workflow,” Grubb said.
Another Bank of America application gathers information on spreads, pricing and positions across bonds, credit default swaps, ETFs and indices to build a more coherent source of information for credit markets.
The bank now has more than 100 agents in beta or development. Grubb described beta as a form of “cautious production”.
But getting an experiment to work is only part of the challenge. Firms also need to establish whether an application delivers enough value to justify taking it further.
Sophie Marnhier-Foy, vice president of financial technology and head of digital asset solutions at Nasdaq, said measuring success and ensuring users would adopt the technology had become an important part of its approach. Nasdaq is looking at measures including how many working hours an application could save and what additional data or metrics it could generate.
One application under development is a pre-production “morning brief”. Before the working day begins, agents would gather information and suggest which tasks a back-office employee should prioritise.
Where AI struggles
As with any new technology, however, not every use case has worked. At BNP Paribas, Maurel said AI had proved less satisfactory when asked to work with large sets of numerical data, such as profit-and-loss or independent price verification figures, to identify outliers or trends. The problem is not simply getting an answer but not seeing the work and understanding how the model reached it, she said.
“If you have a text and you ask a question on the text, you can have a quote and say, well, that is where it comes from in the text,” Maurel said. “If you say, I am going to manipulate a lot of numbers and output a result, it is going to be harder to sometimes have the full reasoning and the full justification.”
Alternatively, she said developers at the bank will ask AI to produce code that will carry out the calculation. Then they will check the code before implementing it.
But AI-generated code creates another problem. While AI is invaluable for producing short pieces of code, Grubb questioned what happens when it is used to generate a full application or a system.
He said firms could be approaching a point where AI produces so much code that the humans in the loop do not have time to read the entire codebase.
“I feel that we are probably now well past the cusp of that, meaning we are going to transition to a point where we are going to end up with code bases that no one has ever actually read through completely as a human,” Grubb said.
He said relying on testing to verify the inputs and outputs of code that nobody had fully reviewed represented “a massive leap of faith”.
From recommendation to action
Firms are also distinguishing between AI that assists employees and AI that acts on their behalf.
Marnhier-Foy described a progression from chatbots that answer questions to more independent agents capable of investigating problems, accessing data and preparing recommended actions.
“[The agents] will stop at the point where they recommend an action to a user,” she said. “It can be very detailed, it can save a lot of work, but they are never going to be able to actually take actions.”
She described a later stage in which agents could “work autonomously, independently”, with humans orchestrating them, but described that as a future development rather than something already operating today.
Bank of America draws a similar distinction. “Anything goes” for activities such as reading, discovery, composition and suggestion, Grubb said, “but it stops at that point of execution”.
“I am not aware of anything in our stack universally that basically allows autonomous taking of any sort of active risk decision that is not supervised by a human,” he said, adding that trading algorithms remain based on hard-coded deterministic logic.
That caution is built into the bank's deployment process. Grubb said Bank of America had spent roughly 18 months to two years developing and refining a framework for deploying and governing AI, with significant involvement from its information security organisation. Employees can only deploy agents within a framework he referred to as “Orchestra”, which places technical controls and constraints around them.
The controls introduce friction when AI is supposed to accelerate development, he acknowledged, but said they reduced the risk of rogue or unwanted behaviour and potential reputational damage.
At Société Générale, AI-generated code still must pass through the bank's normal software development lifecycle, including source control, code review, testing, security controls, deployment processes and production monitoring.
“We are not saying that AI generated it, that therefore we trust it,” Patel said. “The point of it is that AI can accelerate the engineering process without actually bypassing your engineering controls that you already have in place.”
Changing the build-versus-buy equation
As AI makes software development faster and cheaper, it could also change the calculation financial firms make about whether to build technology themselves or buy it from a vendor.
Patel argued that falling engineering costs could challenge some of the economics behind the software-as-a-service industry.
For much of the past two decades, he said, it made more sense to licence an established platform than spend a year with a team of engineers recreating the same functionality. AI is starting to change that calculation.
“If I use 20% of a vendor platform to pay for 100% of it, and AI makes reproducing that 20% dramatically cheaper, does buying the platform still make economic sense?” Patel asked. “It is not going to.”
He does not expect software-as-a-service to disappear, however. Vendors with differentiated intellectual property, data, scale, network effects or deep domain functionality would remain valuable, he said.
The same shift could change the skills firms value, the panellists agreed. As AI makes implementation faster and cheaper, Patel said greater value could lie in understanding the underlying problem, architecture, domain knowledge, system behaviour, interfaces, data and validation.
“The LLM does not know that you are asking it a specific task of what it needs to do, and it is just doing it without knowing why it is doing it,” he said. A strong engineer, by contrast, understands the purpose behind the task.
Asked what a similar panel might discuss in two years’ time, Maurel pointed to accountability. She said agents at different institutions could eventually interact without instructions from either side.
“Who is going to take the credit? Who is going to own whatever output is produced out of that?” Maurel asked. “Who do you hold responsible for something that happened?”
For now, firms are working out which generative AI applications are best suited to production. As development becomes faster and agents grow more capable, the harder questions may increasingly be about what firms still need to buy, what they are prepared to let AI do, and who is accountable when it does it.
