Models for nothing

Nvidia and almost every major AI company this week went to bat to keep both open models and closed models at the AI buffet. Banks, ever-hungrier consumers of this tech, are still figuring out how much of each to load onto their plates.

JPMorganChase will “be smart about open source where appropriate,” CFO Jeremy Barnum said during this month’s earnings. NatWest has Meta “building us a set of proprietary models built on open source” that are smaller and cheaper to run, CIO Scott Marcar said this month.

Lenders have raced to make their platforms model agnostic, meaning they’re built so the bank can swap Claude for an open-weight alternative on the fly. But having that choice only makes AI cheaper if a bank can figure out how much of each job to give each model so that the cost of a task falls while quality doesn’t.

The basic rules are getting clearer: “You really don’t need the latest cutting-edge, incredibly expensive model to summarize an analyst report,” JPMC’s Barnum said. In practice though, banks are still searching for the tipping point between quality and price. Morgan Stanley, for example, starts with the strongest model and works its way down the ladder. “Let’s build to the best model,” said Mike Pizzi, global head of technology and operations, last month. “And then let’s bring the models down to the most efficient where there’s really no deterioration.”

CommBank this week illustrated the opposite approach: Start with cost and see how far smaller models can take you. The Australian lender was sitting on 12,000 hours of video that would’ve taken its five-person archive team at least a year to catalog. Instead, the bank built a searchable database of all the footage using two open-weight models. OpenAI’s Whisper transcribed the audio and Google’s Gemma described what could be seen in each frame. “Cost efficiency was the key point of optimization,” the bank’s engineers Lachy Berry, Xun Wang and Yossi Frenkel wrote this week. Their open approach meant the job cost about $1.31 per hour.

But their result also showed why banks can’t just default to the cheapest model everywhere. Whisper, an American model, committed a cardinal Australian sin when it struggled to parse the phrase “Crocodile Dundee.” The engineers said they’re looking for newer alternatives for future uses – likely at a higher price.

On an archive project, that miss is harmless. But in core banking work, an error could cost clients real money and shred the bank’s reputation. “We believe that fitting the model size to the problem unlocks performance at price,” CommBank’s researchers wrote. The hard part for banks now is doing that fitting every place AI is used.

This week, Italy’s UniCredit revealed that it completed the first cross-border deal sourced through DealSync, its AI-powered M&A platform. We sat down with Alessio Sulpizi, the bank’s head of digital innovation for advisory and financing solutions, and Stefania Godoli, head of mid-market strategic advisory, to see how the matchmaker platform works.

Why it’s interesting: Smaller M&A deals can require as much work as large ones while generating a fraction of the fees. That made them hard for investment banks to justify, even as UniCredit’s local bankers kept hearing from business owners who wanted help buying, selling or raising capital. The bank built DealSync to make the economics of those smaller deals work. AI takes on the costly and time-consuming work of searching for counterparties for a deal and preparing the paperwork, leaving bankers to handle judgment and introductions.

How it works: A relationship manager logs that a business may want to buy a company or sell itself. Once the client formally hires UniCredit by signing a mandate, DealSync ranks potential counterparties based on their financial fit, their products and services and what each side wants from a potential deal. The custom-built model combines UniCredit’s internal information with public market data. It returns a ranked list and explains why each match could work. An M&A banker checks the results, cuts the list to a narrower set of names and takes it to the client before UniCredit contacts anyone. “It doesn’t just write you the name,” Sulpizi said. “It tells you why it has a high probability of being a good match.”

How they did it: DealSync grew out of UniCredit’s experimental Moonshot team. Its matching engine was custom-built in part by the bank’s internal team, Sulpizi and Godoli said. Previously, leads for smaller deals arrived via email or phone calls and often went nowhere. With the platform, they’re able to turn those tips into a deal pipeline. “What we really did was create a funnel,” Sulpizi said. “It’s not just using the tool to produce lists. It’s about understanding which clients we should talk to, producing the content the network needs to engage them, and then producing the pitch again using AI.”

By the numbers: DealSync now has more than 6,000 businesses registered on the platform and has led to the bank signing more than 600 client mandates. That’s double the number of registered firms and 150 more mandates than the firm reported at the beginning of this year. So far, the bank has closed five deals. “It’s 100% new revenue,” Godoli said. “We couldn’t have tackled such transactions with a traditional model.”

Bigger picture: UniCredit is targeting €500 million ($576 million) in cost savings from AI in the next five years, CEO Andrea Orcel said in March. Tools like DealSync add a revenue boost on top of productivity. “Our ambition is not to close five deals. It is not to close 10 deals,” Godoli said. “It is to close hundreds.” The bank’s focus on smaller deals comes as competitors move in on the space: A Greek ecommerce business used a chatbot to answer initial due diligence questions from prospective buyers earlier this year. And OffDeal is working to automate M&A deals from start to finish.

That’s how many HSBC plans to hire in Singapore, 100 of whom will be AI specialists to staff its first global AI center of excellence, the bank announced this week. In addition, the bank plans to hire 100 relationship managers for its wealth business, Kee Joo Wong, the bank’s Singapore CEO wrote. The new center itself has a specific mandate: Teams will focus on “enhancing customer wealth journey conversations, introducing agentic treasury solutions and developing AI-enabled digital payments,” the release said.

Zoom out: HSBC’s even split of tech hires and bankers captures the new shape of bank hiring as firms prioritize, as they often call it, creating capacity through new AI tools. This month, Singapore-based OCBC also announced its own plans to hire some 600 relationship managers for its wealth unit alongside its rollout of two avatars that could automate some wealth tasks (see: “Wayne’s world,” The Brief, July 2). Citi is hiring hundreds of people in its wealth unit, particularly in Asia, as it rolls out its own avatar, Citi Sky.

This week we learned that OpenAI’s latest models had autonomously hacked multiple other companies during routine testing. AI agents can now search huge quantities of software for security weaknesses and exploit them far better than ever before. It seems we may be headed for…

Michael Cembalest, chairman of market and investment strategy at J.P. Morgan Asset & Wealth Management, gave a name to the expected flood of AI-discovered weaknesses: Patchmageddon. As AI models get better at cybersecurity, bad actors will get better at using them to sniff out weak spots. It means software is in what Anthropic security bigwig Nicholas Carlini calls a dangerous “transitionary period.”

The bad news: Patchmageddon is already upon us. The number of “critical” vulnerabilities reported this year is up 10-fold compared to last year. And those are just the ones being reported by the good guys.

Banks may see a lot of that as someone else’s problem. But modern software gets built a bit like LEGOs: JPMC estimates that the average commercial application is built on around 1,200 pieces of open-source code, much of it maintained by lone hobbyists and ripe for exploitation. That leaves firms with a grim spring cleaning job: mapping every dependency they have and working out which bits of borrowed code can be trusted, and which have to be replaced. Even finding the holes doesn’t mean they can fix them fast enough. In roughly 60% of breaches, a fix was already available but had not yet been installed, the report said.

So is there any good news? Probably. At a certain point, banks may well fix the flaws in their codebases, and AI will scan new software for vulnerabilities before it goes live. But that’ll only be the case after “a couple years of total chaos,” former Meta security chief Alex Stamos said.

BBVA’s Jon Ander Beracoechea and Álvaro Martín both jumped to Madrid cross-town rival Santander. Martín was head of AI transformation for commercial client solutions and Beracoechea was chief scientist. Martín will perform similar duties at Santander, while Beracoechea’s exact title is still being determined, Cinco Días reported.

Oscar Garcia Ramos is now head of the AI platform adoption team at BBVA, “responsible for driving the enterprise-wide adoption and value realization of BBVA’s AI Banking Platform,” he wrote on LinkedIn. His role fits within broader changes the bank has made to its AI organization: The bank this summer established its AI Transformation team, which is made up of its architecture teams (now called AI Tech), its global process team (renamed AI processes) and its data platforms teams, Antonio Bravo, global head of AI transformation, said in an interview last week.

NatWest hired Triona O’Keeffe to be its chief data and analytics officer, starting in January 2027. In the role, she’ll report to CIO Scott Marcar and will lead the bank’s data and engineering teams. She’s currently CIO for data and analytics at the London Stock Exchange Group.

HSBC hired Val Riabtsev to be its head of AI and data capabilities for the international wealth and premier banking units. He was most recently head of AI platforms at Revolut, where he built the neobank’s internal AI agent platform, he wrote on LinkedIn.

Citi promoted Shahzad Alam to be its chief data officer for Singapore. He was most recently director of data analytics for banking and international.

Alisha Lehr, Morgan Stanley’s former COO of firmwide AI, joined OpenAI to work with private equity firms to help them “translate AI’s potential into measurable, durable value,” she wrote. It’s the latest hire the ChatGPT-maker has brought on board to bolster its Wall Street offering: Last month, the firm brought on Scott Mullins to be head of financial services following a 12-year stint at AWS.

“The speed at which innovation is happening in terms of capability is, for argument’s sake, say 5x. The speed of innovation of governance and control and management of these agents is at 1x.”

– Nimish Panchmatia, chief data and transformation officer at DBS, in an interview, July 28

Singapore took AI’s center stage this week with the country’s two largest banks sharing two of their most advanced agentic tools. DBS detailed what it’s calling one of its most advanced AI deployments: an agentic system that uses between 70 and 80 agents to comb through reports, financial documents, news and competitive analysis and compile it into corporate credit memos. It’s a task that used to consume days of bankers’ time, the bank said. Across town, OCBC rolled out an agentic KYC tool that it says will halve the time it takes to onboard its wealthiest customers, from 30 days to 15. OCBC’s private bank, Bank of Singapore, was already using agentic AI to automate one part of the process: writing a source of wealth memo (see: “Follow the money,” The Brief, Nov. 6). Now it’s using five agents to research prospective clients, spot gaps and prepare the due diligence picture before a relationship manager even approaches them.

More than 90% of customers who start an Aviva insurance quote through ChatGPT make it to the quote stage, Jason Storah, the firm’s UK and Ireland CEO of general insurance, told lawmakers this past week. That’s up from 60% on its normal online platform, he said. It’s an important data point, given some experts, like Visa’s former chief data officer Bob Hedges, say that “AI agents could well become the default interface for financial decisions within three to five years.” Banks, for their part, haven’t been so willing to cede the front door to their products to consumer chatbots: NatWest debuted a mortgage tool in ChatGPT in April and BBVA launched an app as well (see: “Curb app-eal,” The Brief, May 14). Neither has shared any upside the way Aviva has and others haven’t followed suit.

Lloyds is launching a four-year £2 billion ($2.7 billion) cost cutting program centered around using new tech and agentic AI, CEO Charlie Nunn told investors during the bank’s earnings call Thursday. That will include new agentic assistants that help staff be more productive. The bank also said it would roll out AI-powered advice tools for its wealth division. Though the focus is cost cutting, Nunn said the new program would “require us to continue to re-skill people and hire new people.” So far this earnings season, seven of the top-10 lenders in the Evident AI Index for Banks have reported updated headcount numbers, and five have shown growth over the last year. It’s not the same everywhere in financial services. Visa announced this week it was cutting roughly 2,600 jobs, or 7% of its workforce, with technology and product teams taking most of the hit.