Is Wall Street's AI risk analysis right for RIA portfolios?

Risk management is getting an AI makeover as Anthropic inked a new partnership with Millennium to build a digital risk analyst for the hedge fund. 

Anthropic, the developer of Claude, will design its AI risk analyst to alert Millennium’s staff of risk insights across asset classes. Anthropic’s latest move represents a shift in AI’s place in finance, says Jake Miller, chief solutions officer at Opto Investments, an alternatives marketplace for RIAs and wealth managers.  

“Most AI tools advisors touch today are request-response. You ask a question, you get an answer, and the value depends on knowing what to ask,” Miller told InvestmentNews. “What Millennium and Anthropic describe is different in kind: a system embedded in the risk stack that carries context across interactions and produces recommendations without waiting to be prompted. That moves AI from answering questions to noticing things.” 

Does continuous risk monitoring suit advisors?

The advance from Millennium and Anthropic follows JPMorgan’s development of “long-running autonomous agents” that analyze markets and client holdings for hours without human intervention, an executive at the bank said in June. But while hedge funds like Millennium find value in continuous risk monitoring, the new AI tool could be less fitting for advisors in the RIA space, Miller says. 

“Continuous position-level risk monitoring with daily attribution is Millennium's core business. For an advisor managing client portfolios, the highest-value work happens up front: setting the risk budget and building a portfolio the client's goals and liquidity can actually support,” Miller says. “Get that right and daily market moves should not whip the portfolio around. A tool that surfaces recommendations every day invites exactly the behavior good advisors spend their careers talking clients out of.” 

Millennium said in its announcement with Anthropic that the hedge fund will keep “human judgment at the center of decision-making" while using Claude to reason through risk positions and explain daily changes in risk exposure. 

“Where continuous monitoring could genuinely help advisors is watching the plan rather than the market: portfolio drift and creeping concentration, plus early warning on real tail risk. Alerts tied to the client's plan are useful. Alerts tied to the tape mostly generate activity,” added Miller. 

LLMs versus deterministic engines

Millennium joins several other financial firms to reach formal partnerships with Anthropic, including Rockefeller Capital Management, LPL Financial, and Orion. Toby Wade, founder and CEO of DeepVest, an AI platform for investment advisers, is skeptical of Anthropic’s AI being used to perform underlying financial calculations to assess risk. That sort of math involves determining portfolio values, measuring options risk, and estimating potential losses. 

“What's unclear from the announcement is whether this is a pure LLM-based reasoning system or if it uses deterministic calculation engines under the hood. For institutional risk management, you need deterministic accuracy. The AI should orchestrate workflows and identify insights, but deterministic tools must do the actual math,” said Wade. “An LLM forming opinions on risk exposure is fundamentally different from an LLM querying verified calculation engines and presenting the results.” 

For continuous risk monitoring, DeepVest has launched MonitorLab, which evaluates client portfolios for RIAs against each client's documented objectives. The software triggers personalized alerts on portfolio drift beyond tolerance bands, concentration changes, risk threshold breaches, and rebalancing opportunities. 

Potential hallucinations in financial calculations is the biggest risk Wade sees in using large-language models or frontier models from AI giants like Anthropic.  

“General-purpose LLMs, even frontier models, are trained to predict the next token, not to perform deterministic financial analysis,” said Wade. “They can reason about portfolio strategy, explain concepts, and identify patterns, but they're not calculators. When you ask an LLM to compute a Sharpe ratio, analyze cost basis, or evaluate tax-loss harvesting opportunities, you're relying on probabilistic text generation, not verified computation.”