Dutra contrasts this approach with some of how generative AI is being tacked on to legacy planning systems. He argues that in many other cases, AI is being used as a marketing buzzword rather than a value add. Generative AI can get tacked on to an existing system to add qualitative analysis and commentary on manually inputted data, too, but Dutra and Neami argue this is a less ideal use of the tech. By putting AI at the data input part of the process, though, he believes his firm can leverage what AI does well and then use hard-coded automation to power the calculations and ensure that the mathematical outputs are correct.
Just as they try to use AI for its capacity to understand and act on plain language, the Friedmann team doesn’t use AI where it’s weakest: external numerical data gathering. Because of the capacity for hallucination in many large language models, they have instead used hard-coded tools and a knowledge base of accurate financial data that the AI can pull from. Neami explains that AI hallucinations often occur when the LLM loses context for the situation, which can be a product of the AI agent relying on its long-term memory. Their agent lacks long-term memory, and will only pull the tools it needs based on specific queries, resulting in a situation where the LLM is protected against hallucinations. In addition, they give advisors full insight into the thinking that the AI model employs, allowing them to double-check the work done.
“Say an advisor is asking for a retirement plan, how do they accumulate from RRSPs, TFSA, et cetera. It will literally show the entire map, exactly what it’s thinking, where it’s connecting,” Dutra says. “The advisor can follow essentially the script if they wanted to, to make sure that end result makes sense.”
Dutra explained how he now plans to grow this platform, noting that while FriedmannAI is currently marketed towards advisors it began as a direct to consumer idea. They may still launch a consumer version, either on their own or in partnership with one of the robo-advice platforms. Dutra, himself a former advisor, elected to pursue the advisor market first in part to test his platform on a more knowledgeable client set and because he saw a huge demand for this platform among advisors. He noted that many advisors are currently using public, un-gated LLMs to achieve similar functions, risking data security and potentially driving poor quality output.
The platform, he explains, is also limited in terms of what clients can access. The platform can’t open up TFSAs or RRSPs for clients, it can’t sell a life insurance policy or draw up a will, it can’t incorporate a client or invest in an ETF on their behalf. All that still must go through the advisor. He argues that this platform can serve as a “connector” between advisors and their clients. Because the platform makes no recommendations as to portfolio structure or securities selection, it stays compliant with KYC and KYP regulations. Dutra argues that the ease of data input in this platform can also help with that process.
