Rob Mallamee

Mallamee also warned about the dangers of family offices bypassing vendors and attempting to build AI systems from the ground up by “vibe coding.”

A do-it-yourself approach can work when assembling a relatively simple AI application such as a power point presentation, Mallamee said. But a more complex platform like a CRM and other systems of record  are “hard to recreate from the ground up,” he said.

A homegrown CRM may perform routine tasks perfectly well, but is likely to stumble when confronted with “edge cases” that a CRM from Salesforce or another legacy vendor has encountered for decades. He referred conference attendees to a recent tech post from consulting firm Bain & Company on agentic software development; it reported that while most companies have AI tools, “few have built the engineering systems that make those tools effective.”

Mallamee also highlighted security concerns regarding in-house AI applications. While they “may get functionality right,” vibe coded programs could also incorporate “sensitive information” by mistake, he warned.

Breaking down complex tasks
The most compelling use of AI for family offices is in ontology, “a plain English wrapper over all your underlying databases and software platforms that also grasps how your office works,” said Circulus CEO Joe Reilly . “That is a game changer. It will finally allow families to map their actual workflows and find real efficiencies in what is typically a messy flowchart and lots of tacit knowledge locked up in the family’s longtime accountant.”

Joe Reilly

Another AI advance is the notion of “recursive decomposition,” or breaking down complex tasks and “decomposing” them into subtasks, said Brian Malhstredt, chief technology officer at the AI firm Metis.

Following the release last month of Anthropic’s Claude Opus 5.5, which has a higher benchmark performance, faster speeds, and lower pricing than its predecessor, intelligence is no longer an AI bottleneck, Malhstredt, a former SpaceX engineer, noted.

After breaking down complex tasks into subtasks, AI can then execute them in sequence or in parallel and reassemble into a final result. “We don’t think like that because it would involve lots of tedious work or very high-end software development,” Reilly said. “But now you can just prompt it in a chat window.”

“There will be some accidents”
The conference also featured Wall Street veteran Peter Krause, former CEO of AllianceBernstein who is now CEO of Aperture Investors.

Krause boosted equities as his preferred growth investment, but cautioned that the most pressing macroeconomic problem was global inflation and corresponding market volatility. He was bullish on US stocks, despite high interest rates and persistent inflation. The US has avoided stagflation by being the fastest growing economy in the world, he said, predicting that there was “still much growth to be had.”

Despite fears of too much concentration in AI stocks and comparisons to the Internet bubble of the late 1990s and early 2000s, Krause noted that companies like Anthropic already have hundreds of millions of dollars in revenue, while early dot-com companies had none.

One of the AI hyperscalers may be overvalued, Krause acknowledged, but he “didn’t know which one.” Krause also speculated that a huge tech company like Oracle may have found itself with an overextended balance sheet that could implode with systemic repercussions. “At some point there will be some accidents,” he said.