Tracking AI Agents Across Finance
The FSB wants firms to track, test and control AI agents, plus Magnetar’s AI-powered fund, the pushback on job-loss fears and AI’s investing wunderkind.
Hey, it’s Matt. I’m a former Bloomberg News reporter, and you’re reading AI Street, where I report on how Wall Street uses AI.
While I know it feels like we’ve been in AI times forever now, the reality is we’re in something like 1997 days when software updates arrived on CDs in the mail.
That’s because there’s no user manual for AI. We lack definitions of common AI terms: What makes an agent, an agent? What’s the difference between a chatbot and a copilot?
Yet the technology is live. You can have your agent trade for you, invest for you, and apparently attend meetings you don’t want to attend for you.
Along the way, it’s gonna break. Regulators are always worried about things breaking but you kind of have to feel for them because who can really keep up with all this?
This week, the Financial Stability Board, an international group that monitors risks to the global financial system, is trying to get out in front by floating what look like pretty sensible proposals:
Maintain an inventory of every AI use case, including its purpose, owner, underlying models, data sources, risks, dependencies and approved or prohibited uses.
Give AI agents individual identities and limited permissions. Banks should record which databases, APIs and tools each agent can access and certify agents to operate only within defined boundaries.
Monitor how agents complete tasks, not just their final answers. Firms should log agents’ intermediate decisions, tool use and database queries to detect errors, misuse or attempts to expand beyond their approved role.
Install kill switches that can stop an AI system entirely or reduce its autonomy and return the task to human operators.
Test AI against existing systems before deployment. Banks should check whether a new model actually outperforms the incumbent, remains stable over time and continues working under new market conditions.
Actively try to break AI systems. Staff should attempt to make agents ignore instructions, exceed their permissions or access financial accounts before and after deployment.
Prepare for vendor failures and model changes. Banks should monitor third-party updates, require notification of material changes and maintain alternative providers or manual processes.
This is still a draft. The FSB is seeking industry feedback through July 22.
A NOTE FROM OUR SPONSOR
Your AI stack is only as good as the data underneath it.
Carbon Arc Lenses puts an MCP Server with 210+ institutional data assets directly into your AI workflow. Card spend, payroll signals, foot traffic, medical claims, and more, queryable in plain English, with no SQL, no data team, and no annual contract.
Built for AI-native research workflows at institutions that can’t wait weeks for data to clear procurement. Now live in the official ChatGPT App Directory, Perplexity Pro, and as a custom connector in Claude.
Start querying with Lenses using code AISTREET30 for 50% off your first month.
USE CASE
It’s hard to tell where exactly we are in terms of live AI in financial services. I often hear firms say they “do” AI, but then I talk to someone who works there and the reality is more limited. In the FSB report, the group highlighted specific use cases firms have deployed. Here are a few highlights:
Fraud Detection With Agentic AI: Built in three months, a bank’s agent proposes fraud rules for human approval. Built on AI monitoring over 80 million daily signals, it contributed to three-quarters of card-fraud rules and helped reduce losses by over 20%. Separately, a digital bank uses facial and background-image analysis to improve detection of mule accounts, deepfakes, fake identities, and account takeovers. Report page 9.
Scaling Relationship Management With AI: A G-SIB’s AI platform combines internal and external information to prepare tailored client proposals. It halved preparation time, tripled client dialogue, and was expected to reach approximately 3,500 relationship managers in FY2026. Report page 10.
Operational Efficiency With AI: An insurer eliminated approximately 400,000 manual processing instances in 2025. Its AI underwriting system handles roughly 4,000 monthly requests, cutting turnaround from up to three days to about 45 seconds. An AI coding tool reached 80% developer adoption and wrote six million lines of code. Report page 11.
Developmental Testing: A bank postponed deploying an ML trading model after it failed to consistently outperform the existing model, showed stability issues, and used unrepresentative training data. Report page 39.
CIOs managing billions read AI Street. Paid subscribers access original reporting, data analysis, and interviews with the executives and researchers leading Wall Street’s AI buildout.
Magnetar Builds AI-Only Analyst Team
From Bloomberg:
Magnetar Capital, the $18 billion hedge fund firm, will shun human analysts for its newest offering and instead deploy hundreds of AI bots to research stocks.
The firm’s AI technology seeks to replicate the depth of research and analysis usually provided by fleets of humans, according to people familiar with the matter, who declined to be identified because the information is confidential. The bots will scour the investing universe for ideas, analyze stocks, make recommendations and forecast trends, the people said. Humans will make the final decision on any trades.
This is the first firm I know of that is relying on AI agents to conduct research for a fund.
Trevor Mottl, Magnetar’s head of AI Quant, built the system after managing portfolios at Walleye, Lazard and Man Group and overseeing long-short equity risk at Balyasny.
This Week in AI Street
Former Millennium-Backed Traders Start AI Commodities Hedge Fund
Les Finemore is building an AI-driven hedge fund for global commodity markets he says have been slow to adopt new technology.
AI isn’t Coming For Your Job
The current data don’t support the idea that AI will inevitably take your job, something I keep banging the drum about. I’ve written many posts pushing back on this assumption (see here, here and here). So, it was nice to see several articles make the same case over the past week.
A reality check on the AI jobs hysteria MIT
The Future of Work and AI WSJ
AI Is Upending One of Finance’s Cushiest Jobs BBG
ICYMI
AI’s Boy Wonder
The haters came for this WSJ story (check out the comment section) on Leopold Aschenbrenner, a 24-year-old investor who’s up 1,000% in fewer than two years. Prior to starting his fund, Situational Awareness, he had no professional investing experience. He now manages $20 billion and counts Jane Street as a backer, per WSJ.
These returns are ridiculous, and two years is far too short a track record, of course, but I think there are too many knee-jerk reactions dismissing this as a bubble rather than thinking through whether he might be right.
Granted, you will need a lot of time to think these things through. You can listen to this 4.5-hour podcast(!) with Aschenbrenner from two years ago or you can read his 45,000-word thesis.
ROUNDUP
Quick Hits
Citigroup Is Rolling Out Tokenized Shares of Private Companies WSJ
AI Isn’t Replacing Credit Hedge Fund Traders Yet, Barclays Says BBG
How finance teams use Codex OpenAI
Goldman CEO Solomon on Running a Bank in the Age of AI Odd Lots
Inside Hudson River Trading’s Blistering Token Burn Odd Lots
SPONSORSHIPS
Reach Wall Street’s AI Decision-Makers
AI Street reaches institutional investors, C-suite executives and Big Law attorneys at firms including JPMorgan, Citadel, BlackRock, Skadden, McKinsey, and more. Sponsorships are reserved for companies in AI, markets, and finance. Email sponsors@ai-street.co for more details.
CALENDAR
Upcoming AI + Finance Conferences
AI in Capital Markets Summit London – June 17 • London
Capital-markets AI conference covering data, trading, compliance, and operational use cases.AWS Summit NYC / Anthropic at AWS Summit – June 17 • NYC
Enterprise AI and cloud event with Anthropic participation at AWS Summit New York.
KNIME Data Summit New York – June 18 • NYC
Data/analytics summit centered on KNIME workflows, tooling, and applied use cases.
IAQF + LSE: AI/ML in Finance – June 23 • London
Finance-focused AI/ML conference hosted by IAQF with LSE involvement.
Thanks for reading!
I’m always happy to receive comments, questions, and feedback.
connect with me on LinkedIn, or
send an email to matt [at] ai-street.co
We have reached the point where an insurer eliminated 400000 manual processes and a coding tool wrote six million lines, yet the headline grabbing story is still a 24 year old turning a thesis into 20 billion dollars. Apparently incremental efficiency gains do not trend, but a young guy with a great track record always will.