Building AI Inside BlackRock
Dhagash Mehta on translating trader problems into AI, making agents auditable and why he wants them to challenge one another.
BlackRock is best known as the world’s largest asset manager, overseeing more than $15 trillion. But CEO Larry Fink casts the company as part money manager, part global technology company. Its Aladdin-led technology-services and subscription business generated about $2 billion in revenue last year, nearly as much as Cloudflare.
BlackRock began using Aladdin, its in-house portfolio- and risk-management system, in the late 1980s and started selling it to clients in the 1990s. Originally built for institutional investors, Aladdin has since expanded into a broader platform spanning wealth management, trading and private markets, with BlackRock acquiring eFront and Preqin along the way.
Aladdin now has more than 130,000 users and accounts for most of BlackRock’s $2 billion technology-services and subscription business.
More recently, BlackRock has started adding AI to its products and investment process. An Aladdin Wealth feature used by Morgan Stanley draws on portfolio holdings, client preferences, risk data and the bank’s market outlook to draft commentary for financial advisers. Internally, BlackRock’s Asimov system uses AI agents to scan research notes, company filings and emails for information that could affect portfolios managed by its fundamental-equities teams.
BlackRock is also rolling out RockAI, an internal platform that allows employees to create specialized agents without writing code.
To understand how BlackRock is approaching the next phase of that work, I recently spoke with Dhagash Mehta, one of the researchers behind its experiments with AI agents. I’ve previously covered his research here and AI developments at BlackRock:
Mehta runs applied AI for investment management at RockAI. He spent the previous three years in Aladdin Financial Engineering, the firm’s centralized quant group, working on trading, liquidity and portfolio optimization. His job is to take problems from portfolio managers and traders and turn them into models that could eventually run on Aladdin — with the access controls, guardrails and audit trail required to use them.
This interview has been edited for clarity and length.
What makes building an AI platform for a regulated industry different?
We have to have many guardrails, and access management is a big thing because we always have to make sure that certain data is not accessible to certain people. And tracing as well—tracing the whole thing and making everything auditable. Plus, there are different developers who all might have different rights.
How do you select the research topics?
Previously, I was in academia for many years—some would argue too long. There, whenever I used to choose a problem, it was basically from my head. I would create a toy model, create a problem out of thin air, and try to solve it. Then I switched to industrial research, and I am realizing that real-world problems are even more challenging and even more interesting.
My strategy is—I am not running a completely academic AI lab here. My eventual goal is still to deploy models for us, and they might have some very good use for our portfolio managers and traders, eventually on Aladdin. The problems here come from the actual PMs and traders. Most of the time, they start talking about some problems in their language, and then I translate it to the machine learning and AI language. Most of the time, no off-the-shelf method would solve this problem as is. I always have to innovate something, even some small thing. Then I realize that maybe no one else has thought about this in the literature, so that is my opportunity to write a paper.
How do those problems usually reach you?
Because I am embedded in the business, we always have frequent meetings with PMs and traders. We also have planning for the next year, the next quarter, and so on. That is when we go through many discussions, and that is where all these problems come up. PMs might say, “This is our burning problem right now. Can you help solve this for us?” That is how it starts.
Your agent paper used three agents to assess companies. How did that work?