Allan Roth wanted to test how well artificial intelligence handled portfolio allocations.
So the founder of Wealth Logic in Colorado Springs, Colorado, put it through its paces and analyzed the results.
In a recent piece for Advisor Perspectives, Roth explained how he fed Anthropic's Claude two Vanguard statements, one consolidated Fidelity statement, and one statement for an employer 401(k) held at Fidelity.
He instructed Claude to target asset allocation at 68% equities with two-thirds in domestic stocks and the remainder in international. Roth asked Claude to summarize the portfolio and make recommendations.
Roth was mostly pleased with the results, with one glaring problem area: math. That's because while large language models are good at predicting text, they can't be trusted with calculations.
Roth told ThinkAdvisor the solution Claude provided in its response to these concerns was to create Python code to check its numbers alongside its overall portfolio allocations.
"Financial planners that embrace it and learn to use it should still always be checking it," he said. "I don't know whether we'll ever get to the point where [portfolio allocations] can be done without checking."
Tyler Grube, a consultant at Sapling Financial Consultants in Toronto, said he has conducted similar experiments in Claude and has encountered similar calculation errors.
"LLMs hallucinate and in the finance business, where millions to billions of dollars are on the line, such mistakes cannot afford to occur," he said. "The real problem is false confidence in the LLM's output and not double-checking its results."
Still, this sort of capability has the potential to change the role of the financial advisor from forming these portfolios to using AI to create them, Grube said.
At that point, he said, it is on the advisor to review and catch any errors in the output.
"Advisors who embrace this tool will be able to manage more clients and win more business for their firms," he said.
Likewise, Jon Ulin, managing principal and private wealth advisor at Ulin & Co. Wealth Management in Boca Raton, Florida, has been using AI, including ChatGPT, to analyze portfolios. He said this has significantly improved speed and structure when reviewing complex client balance sheets across multiple custodians.
Ulin said AI performs well at organizing information and identifying patterns. It can summarize holdings across statements, flag concentration risk, identify overlapping exposures, estimate tax sensitivity ranges and outline restructuring options, he said.
"In effect, AI functions like a fast analytical associate," he said.
Where AI is weaker is precision math, cost basis verification and suitability judgment, said Ulin. These models, he said, are trained to recognize language patterns, not replace portfolio accounting systems or planning software.
"Advisors should continue relying on custodial data, rebalancing tools and planning systems for numerical accuracy," he said.
The practical approach, he said, is to use AI early in the process to frame options, narrow solutions and identify risks.
"The advisor still applies experience, fiduciary judgment and client-specific considerations before implementing recommendations," he said.
© Arc, All Rights Reserved. Request academic re-use from www.copyright.com. All other uses, submit a request to TMSalesOperations@arc-network.com. For more information visit Asset & Logo Licensing.