What good’s your intern

Interns used to earn their return offers by fetching the (metaphorical, most of the time) coffee. No more grunt work for this year’s crop, apparently.

Judging by the not-so-humble bragging on social media in recent weeks, Wall St.’s Gen Z crowd was busy this summer showing grown-ups how to take a swing at solving AI problems way above their pay grade.

At Goldman Sachs, a team of them reported on LinkedIn that they had built a tool called “Token Talk,” which scans AI usage data and adoption patterns around the bank to find places where teams can cut model costs or use AI more efficiently. It won them top marks in the firm’s AI hackathon. Elsewhere at Goldman, intern Josh Weidner built an AI agent that takes support tickets, autonomously recreates the issues they describe inside a testing environment and suggests a fix.

A team of BNY interns, meanwhile, scoped out a tool called “Eliza Green” for a case competition. It analyzes an employee’s prompt, figures out which kind of model it needs and sends it to less power-hungry options. “The architecture cuts down energy usage and compute overhead while keeping performance sharp,” wrote Fardin Rahman, one of the interns on the team. They ended the summer pitching it to CEO Robin Vince.

With potential future employers scrolling their feeds, interns have every reason to make the impact of their summer work sound bigger than it may have been. But Goldman Sachs CIO Marco Argenti said this week that the new class of junior employees really are teaching firm vets some of the better ways to use the tech.

That’s happening at JPMorganChase, too – at least if you believe the interns. “I hope the work I’ve done in demystifying LLM usage across CIB will help the firm better understand and scrutinize costs for years to come,” wrote Syed Sameer Faisal, a CIB intern this week.

Banks seem happy to keep giving juniors the chance to try. For all the predictions that AI would thin out junior ranks, Bank of America kept its campus intake broadly steady this year. JPMC brought on thousands of interns again, too. And investment banks in London are maintaining their junior hiring.

Chinese AI has found a route into Western banks – via a stopover in Singapore.

This week, Ant International – the Singaporean cousin of Jack Ma’s Chinese fintech behemoth – rolled out a new version of an AI model used for forecasting. Some of the world’s top banks put it to work predicting cash flows and exchange rates aiming to cut their currency-trading costs. In this “Model Corner,” we look under the hood at how it works.

Model: FalconTST 2.0
Vendor: Ant International
Banks: Barclays, Citi, Deutsche Bank and Standard Chartered

What it is: For decades, banks have used an arsenal of predictive models to figure out where a line graph – of, say, prices or rates or cash balances – might go next. Traditionally though, each forecasting job needed its own model, built and tuned for that particular type of data. Falcon is what’s called a time-series foundation model, a type of model that’s trained on a huge amount of data from lots of different areas to learn patterns that crop up repeatedly. It can be something of a jack-of-all-trades, saving analysts from “building, tuning and maintaining a separate model for every series,” Rob Hyndman, distinguished professor of statistics at Monash University, told Evident.

How banks use it: Falcon is being used by companies to decide how much currency – and which types – they ought to hold based on how much they expect their cash flows to change. Citi paired the model with its Fixed FX Rates product for an Asian airline, which used its predictions to decide when and how much foreign currency to buy. The airline cut its currency-management costs by 30%. Standard Chartered, meanwhile, counts Ant as a client and is using the model to predict Ant’s own cash flows and prepare better for the FX transactions the fintech may want to make. It predicted those needs with 90% accuracy, the bank said.

Yes, but: There’s reason to be skeptical. These crystal balls are still so new that the way they’re judged is pretty crude. Ant touted that Falcon topped a public forecasting leaderboard, but its competition was a niche class of models – meaning the honor is more like valedictorian of the kindergarten than a Nobel laureate. More importantly, it’s still not clear whether any accuracy gains are worth the extra costs of using a new model like Falcon. Well-built conventional models are already “cheap, transparent, fast and easy to explain,” Hyndman said. “Foundation models bring inference cost, latency, hosting/vendor dependency, and limited interpretability. Even if the accuracy gain over a well-tuned statistical model is real, the operational cost may not be worth it.”

Final word: Western banks have been cautious about Chinese AI. A Chinese-adjacent model like Falcon – which runs locally, like China’s open-weights models – could show the way for others.

That’s how much is going to Fable 5, the frontier lab’s top-of-the-line (and most expensive) model, according to new data from payments firm Ramp. That’s not very much, and, more concerning for Anthropic, it hasn’t budged since early July when the model was permanently released. It breaks from the usual pattern: Opus 4.8 – the lab’s most advanced model back in May – had captured roughly 30% of the money businesses spent on Anthropic models a week after it launched. The model, cheap by Fable 5 standards, is still capturing roughly 50% today.

What’s really going on? There’s a price war underway. Bank tech leaders are getting choosier about ordering the most expensive AI on the menu. By comparison with other options, Anthropic is starting to look expensive. OpenAI cut the price of Luna – the cheapest version of its latest family of models – last month by 80%. Last week, it knocked 20% off Sol, the priciest version. And both Anthropic and OpenAI are facing pressure from cheaper, open-weights models.

“There’s a huge danger here that in the era of AI, we outsource our reasoning to these models, and we have cognitive atrophy that stops us being able to reason from first principles ourselves.”

– Chris Churchman, head of Marquee at Goldman Sachs, on a podcast, Aug. 24

Revolut this week stood up a new AI research team “to work on ML/AI problems specific to financial systems,” Pavel Nesterov, head of Revolut’s AI department, wrote. The new unit will build on the academic work the bank did earlier this year with Nvidia – which led to the creation of its banking-specific foundation model, PRAGMA. Models like PRAGMA – which use the bank’s full arsenal of customer data to improve the accuracy of credit decisions and fraud detection over traditional ML models – are all the rage these days, including at other neobanks like Nubank (see: “Retro-chic AI,” The Brief, Aug. 20). Importantly, Revolut will “publish the results” of its research team’s work, Nesterov said. Not every firm agrees with that approach: “At this stage, I don’t believe I need another ‘leading conference’ paper to prove myself anything. Nor should our full-time researchers, or our absurdly talented interns, be judged by that metric at all,” wrote Atlas Wang, research director at trading firm XTX Markets. “Our reward function is one and only ‘business impact.’ And yes, that typically means no publication.”

Google Cloud launched Gemini Enterprise for Financial Services, a suite of products that includes a financial research agent that Deutsche Bank served as a design partner on. The German lender’s corporate banking team will have first crack at using the new agent, Joanne Hannaford, CIO of the corporate banking arm, said. The bank in the future may use the new agent to gauge the impact of tariffs on products that span geographies or forecast supply chain impacts from weather events like El Niño, she said. Other potential uses across investment banking and capital markets include market analysis and due diligence, Rohit Bhat, Google Cloud’s managing director of financial services, wrote. The offering sounds eerily similar to “Claude for Financial Services,” which Anthropic rolled out more than a year ago – but Google’s inroads into banks through cloud contracts may speed up the uptake of the new tools.

Vanguard this week announced it would acquire AI wealth management startup Altruist for more than $4 billion. The startup, founded in 2018, sells a wealth advisor platform called Hazel, which offers analysis tools that plug into firms’ CRM systems and client data. Its biggest splash came in February when it rolled out an AI tool that generates personalized tax planning strategies. Charles Schwab and Raymond James stocks sank more than 7% on its launch day. Shares of banks with big wealth management arms, like Morgan Stanley and Bank of America, dropped roughly 2%. Banks’ wealth teams – as we profiled earlier this summer – are also trying their hand at building advisor platforms stuffed with AI capabilities (see: “Not your father’s advisor,” The Brief, May 21). Vanguard’s acquisition puts a price tag on what those may be worth to the business.

Want to get a job as a chief AI officer? The University of Chicago Booth School of Business will show you how for a cool $28,000. Banks have loaded up on these executives this year, but questions remain about how long the job will exist as a standalone role (see: “Ciao, CAIOs,” The Brief, April 30). For now, the remit looks like it’s still growing: New data from Canadian model-maker Cohere out this week shows that more than half of companies have put their CAIO in charge of sovereign AI strategy. “I don’t know what the title will be [in the future], but I do think the person, the skill set is something that will be there,” said Christian Hansen, the Booth program’s faculty director.

The U.K. is playing catch-up on sovereign AI. France has Mistral; Canada has Cohere. Britain has yet to produce a homegrown lab in the same league. But it may have another route to AI nirvana: getting its biggest businesses to use the technology faster and better than everyone else.

That’s where Rohit Dhawan, group head of AI and advanced analytics at Lloyds, comes in. This year, he and Starling Bank CIO Harriet Rees were named Financial Services AI Champions by the Treasury. Their brief: Work out what needs to change so financial firms can adopt AI faster without regulators losing their grip. Last month, the pair published their Financial Services AI Adoption Plan. We traded emails with Dhawan this week to talk through what Britain’s AI rulebook ought to look like.

The following conversation has been edited for length and clarity.

EVIDENT: Why were recommendations like these needed?

DHAWAN: A lot of the financial institutions did not really clearly understand what is allowed and what isn't within the current regulatory construct. The recommendation was to provide regulatory clarity around what's permissible, what's not, what's working, what's not. This would give financial institutions a lot more confidence on what they can and can't deploy.

How would that clarification change how you develop and deploy AI

A really good example is customer-facing AI. Consumer LLMs are generally designed to be helpful and engaging, but when it comes to something like financial guidance, you don't necessarily want it to be always optimistic. We need systems that are balanced, transparent and aligned to regulatory expectations. You want to be realistic as much as possible.

If I’m using an AI agent’s financial guidance, there's a question of liability and accountability. And that is not clear from a regulatory standpoint. As AI ecosystems become more complex, there needs to be greater clarity around accountability across model developers, technology providers and regulated firms. Will model providers assume any responsibility if the model provides a wrong recommendation? Probably not. But that is not clear, and hence any customer-facing AI deployment – fraud, personal advice, investment guidance, etc. – comes under scrutiny and jeopardy because of that lack of clarity.

Is fixing that done by reining in the type of guidance ChatGPT can offer or expanding banks’ ability to offer AI-generated guidance?

I think it’s expansion. AI has the potential to democratize access to financial information and guidance. The important thing is ensuring consumers receive safe and appropriate outcomes while allowing regulated firms to innovate responsibly.

How much faster could adoption go if your recommendations were all implemented?

There are other industries adopting AI faster than financial services. Financial services operates within a highly regulated environment where governance, risk management and consumer protection are critical. Regulatory uncertainty can therefore slow adoption, particularly for more advanced and customer-facing use cases. If we can provide greater regulatory clarity, it will certainly help unlock future innovation. However, my view is that this is not the primary constraint on adoption over the next couple of years.

Most of the banks I speak to are still working to scale proven use cases that can already be deployed within existing regulatory frameworks. There is significant value available today without requiring regulatory change. However, regulation needs to be ready for what comes next, because AI will increasingly reshape how customers interact with financial services and how firms deliver products and services.

Ashutosh Chaudhari is now executive director of the AI center of excellence at Wells Fargo. He was most recently senior director of process excellence and intelligent automation at Cisive, a firm which does employee background screening. Earlier in his career, he was VP of process transformation at Barclays.

Brent Reston joined BMO as chief digital officer of U.S. banking, where he’ll “play a key role in helping us accelerate our digital strategy,” his new boss Aron Levine, president of BMO U.S., wrote.

Teresa Heitsenrether, JPMorganChase’s outgoing chief data and analytics officer, lined up her first post-banking gig: She’ll be joining the board of CVS Health on Nov. 18. “We look forward to benefiting from her deep knowledge in data, analytics, artificial intelligence, operational transformation and decades of financial expertise,” said CVS CEO David Joyner.

Jonathan Lewis joined wealth management firm LPL Financial as chief technology and information officer. He was previously head of digital and trading technology for wealth management at Wells Fargo. Earlier in his career, he was head of asset management technology at JPMorganChase.

Dipendra Singh Mal joined JPMC as applied AI/ML lead. He was previously a principal data scientist for applied AI at Capital One. In the new role, he’ll be “building agentic workflows and GenAI solutions for Asset & Wealth Management,” he wrote.

Margaret Glover joined Morgan Stanley’s wealth management arm as VP of AI product oversight and governance. Glover previously founded Product Better, which she described as “a fractional product and operations consultancy.”

Khaled Alizai will become head of AI model assessment at Danske Bank next month. He’s been with the bank since 2023.