Anthropic released Claude for Financial Advisors, an artificial-intelligence overlay that connects to existing advisor systems and automates back-office tasks. The work can occupy up to 80% of advisor time at some firms, the company said.

The tool raises questions about how far AI overlays can go in wealth management before they cross into fiduciary liability, according to Dan Eyre, chief operating officer at DeepVest. Large language models are probabilistic systems that guess answers based on patterns in training data, he said.

Hallucination rates for financial calculations reach the mid-to-high 80s in multiple studies, Eyre said. That applies to Monte Carlo analysis, maximum drawdown calculations, portfolio optimization and rolling correlations. The models are statistical text networks and are not built to perform mathematical operations, he said.

A tool that saves 80% of working hours but produces wrong recommendations one time in 100 would be unacceptable for a fiduciary, Eyre said. A tool wrong 88 times in 100 is completely untenable, he said. The only reason anyone is pretending otherwise is the hype in the AI space, he said.

The two most prominent AI labs are planning multi-trillion-dollar initial public offerings in the coming months, Eyre said. Michael Burry has argued the labs have used the threat of civilizational collapse as a marketing tactic, he said. The labs are not profitable and open-weight models are very close to parity, he said. Their spending is massive and they are deeply entrenched, creating a systemic risk if one collapses overnight, he said.

AI overlays inherit all the limitations and flaws of the foundation systems they are built on, Eyre said. Advisors do not want to become prompt engineers, and open-ended prompting creates a blank-page problem for professionals who want technology to simplify their work, he said. Hallucination rates limit what an AI overlay can do, he said.

Fiduciaries can use AI for orchestration, interpreting intent, qualitative summarization and adjacent areas, Eyre said. Probabilistic calculations are a feature of large language models, not a bug, so the systems will always hallucinate no matter how sophisticated they become, he said. To keep AI agents reliable, firms need a rigorous use-case and process framework, deterministic tools, a consistent governance structure and contextual skills, he said. For the first time in technology history, a demo alone cannot address whether a tool can do what an advisor needs, he said.