Hey, I’m Matt. I’m a former Bloomberg News reporter, and you’re reading AI Street, where I report on how Wall Street uses AI. Every Thursday, I cover the week’s biggest AI-in-finance stories with original reporting and analysis.
OCC Focuses Bank AI Exams on Agent Permissions
U.S. regulators have been pretty mum on how to supervise AI in financial services, but this week the OCC released guidance for examiners on how to assess institutions’ use of AI, including agents.
The OCC isn’t creating a new risk category for AI. It’s treating AI as something that affects risks banks already manage, like operational and compliance risk. And it’s not saying here’s how you should use XYZ model. It’s broader than that: What can an AI access? What can an AI agent do on its own? And what could go wrong if it does? The guide isn’t binding like a rule, but banks tend to pay attention to what their examiners are looking for.
The point is that the risk depends less on how sophisticated the technology is than on what it’s allowed to do. A simple agent authorized to initiate transactions could pose more risk than a complicated model used to help draft internal documents.
The SEC’s Mark Uyeda made a related point this week. Even when firms rely on AI, people decide how to deploy and supervise it, he said. “Automation does not break the chain of accountability.”
SPONSORED BY BRICKROAD
The Next Great Dataset May Be Hiding Inside an Ordinary Company
Brickroad helps hedge funds and AI companies find, evaluate and obtain proprietary data before it reaches the catalogs everyone else searches.
The next source of investment edge may already exist inside a retailer, logistics provider, mobile application or struggling public company. The problem is that the company may not realize it is sitting on valuable data, while potential buyers may not know it exists.
Brickroad uses AI agents to connect the two.
How it works
A hedge fund or AI company describes the information it needs. Brickroad’s agents continuously search established sources as well as companies that may produce data through their ordinary operations. They assess what information those companies are likely to possess, look for indications that the owners may be willing to sell it and begin obtaining access.
When a source looks promising, Brickroad helps secure a sample and evaluate whether the data could be useful for the buyer’s task. The buyer can then decide which opportunities to pursue without having to conduct the initial search or establish a separate connection with every potential supplier. People remain involved as an opportunity moves through diligence, negotiation and licensing.
CEO and co-founder Freeman Lewin points to a dashcam company as an example. Its cameras could provide information about traffic patterns and storefront advertising. A change in its terms of service indicated it owned the data it collected. Signs that the company was preparing to raise money suggested it might be interested in licensing that data.
Brickroad is built to uncover less obvious opportunities, including data held by companies that have never considered themselves data suppliers.
The hard part
But finding a source does not make its information usable. A buyer has to:
document its provenance
establish the seller’s licensing rights
identify issues involving personally identifiable information, material nonpublic information and other legal or compliance risks
determine the data’s history, coverage and freshness
test whether it adds anything to an investment strategy or AI system
Lewin built Brickroad around a problem he had seen firsthand while licensing data for his own companies and then for others. A former attorney with experience in data licensing, he saw deals get stuck between companies that did not know how to sell their information and buyers facing months of outreach, diligence, legal review and technical integration.
The founding team combines that market experience with expertise in AI and data infrastructure. Co-founder and CPO Michael Gordon is a mathematician and repeat founder who has built production AI systems. Co-founder and CTO Luis Oala helped create Croissant, a standard for describing machine-learning datasets.
The bigger picture
Brickroad calls its larger vision a “data multiplexer,” a common connection between data suppliers and the funds, AI companies and software agents seeking their information. Instead of requiring a separate search, contract and technical integration for every dataset, Brickroad aims to give buyers one route through which they can discover, evaluate, license and receive data from many sources.
For buyers, that could mean spending less time searching catalogs and negotiating one-off integrations. For companies, it creates a way to turn overlooked information into a new source of revenue.
And for an investment firm, it offers the possibility of reaching a valuable dataset before everyone else knows it exists.
Preparing for AI Agents to Move Money
The industry is preparing for a world where agents are moving money. Schwab recently invested in Nevis, a startup building AI agents for financial advisors. Nevis CEO Mark Swan says he expects agents moving money between accounts for advisors to become common over the next 12 months.
Schwab isn’t giving agents that authority yet. “Today, we do not allow agents to act on any information at Schwab,” Alison Dooher, its head of AI for advisor services, told InvestmentNews. For now, they can read data, but taking action on an account is a line Schwab has yet to cross.
CFTC Asks Who’s Responsible for AI Trading Agents
The CFTC is examining a related question: if an AI agent executes a trade for a customer, who should regulators oversee?
“Does the developer of that agent become subject to registration? Who is liable if the agent goes rogue?” Chairman Michael Selig asked at the agency’s Oct. 28 Frontier Forum. He suggested regulators may need a version of “know-your-customer” for agents: “know-your-agent.” Selig didn’t propose a specific rule.
Participants at the forum suggested keeping risk controls and audit logs in systems agents cannot alter, and setting limits on how much risk agents can take, both individually and collectively when acting for the same customer. Those safeguards could help firms control agents, but leave open Selig’s question of who is legally responsible.
Goldman Limits AI Coding Tools. Will Others Follow?
Goldman Sachs has stopped most traders and bankers from using Claude Code and GitHub Copilot, limiting the coding tools to engineers after a trial, Bloomberg reported Monday.
Goldman didn’t say whether costs drove the decision. But I suspect it won’t be the last large firm to narrow access to expensive AI tools. When I asked Rob May, CEO of AI inference company Neurometric, whether others would follow, he said via email: “Large firms, yes. Cost is a big issue.”
Morgan Stanley, meanwhile, is letting wealth advisors experiment with building agents. Its wealth advisors have made nearly 50,000 attempts to build AI agents, American Banker reported Monday, though most don’t get past early testing.
I asked Morgan Stanley whether all advisors can still build agents and whether cost factors into who gets access. A spokesperson hasn’t responded.
Few Financial Firms Have Scaled AI Agents: Survey
59% of financial firms are piloting or deploying systems in which multiple agents work together. Just 10% say they can coordinate those agents across the business, according to a report published Thursday by French technology consultancy Capgemini.
Cost is shaping how they build the systems, too. Nine in ten firms are using or plan to use model routing, which sends a task to a model based on its cost and performance.
ROUNDUP
What Else I’m Reading
Hudson River Trading Joins Wall Street’s Prediction Market Shift | BBG
FINRA Warns AI Is Making Investment Fraud Harder to Detect | Traders Magazine
The ‘Shorter’ AI Workweek: 9 to 5:45ish | BBG
US SEC’s top cop says agency is monitoring AI corporate disclosures | Reuters
Wells Fargo wants customers to prove they've revoked AI agents | American Banker
Cboe, Citadel Securities-backed group sues US regulator over single-stock perpetual futures | Reuters
Finpilot Raises $10 Million Series A | Press Release
Morgan Stanley Selects MDOTM AI-Powered Portfolios | Press Release
Sponsor AI Street
November sponsorships are SOLD OUT. Limited December placements are available starting Dec. 3.
Reach more than 5,000 institutional investors and decision makers across finance, AI, and law. To discuss available dates and sponsorship fit, email sponsors@ai-street.co or fill out this form here.
This Week in AI Street
Citadel Securities Develops AI Agents with Clients
Citadel Securities is helping clients build AI agents, a sign of how market makers could expand AI adoption by providing hands-on support to institutional investors.
Speaking to clients in New York Oct. 20, Citadel Securities’ technology COO Jeff Maurone said: “We have teams working with your teams building agents together.”
In his talk, Maurone said the firm is helping clients build secure sandboxes around agents, organize data for models to access, simplify trading platforms to support agents and scale quantitative research. The market maker has spent time with more than 100 client tech teams over the last year across AI adoption, agent development and quantitative research.
CALENDAR
Upcoming AI + Finance Conferences
Asset & Wealth Management Summit — November 10–11 • Boston, MA
Nicsa summit on alts and AI transforming asset/wealth management, with speakers from Anthropic, OpenAI, Morgan Stanley, KKR and Microsoft.21st SFI Annual Meeting — November 12 • Zurich
Swiss Finance Institute’s annual meeting on AI’s impact on finance, with keynotes from UBS, ETH AI Center and Oxford on AI’s risks and opportunities in banking and trading.ICAIF ‘26 — November 14–17 • Milan
7th ACM International Conference on AI in Finance, bringing together researchers and practitioners on generative AI, trading, risk, fraud detection and responsible AI, hosted at Bocconi University.AI in Financial Services Conference USA — November 18 • New York
GlobalData conference on generative and agentic AI deployment across banking, insurance and wealth management.
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
Manage how often you receive AI Street
AI Street is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.