This article is part of a series in our special report on technology in family offices. To see all the articles, click here.
In the competition between artificial intelligence platforms such as ChatGTP, Google’s Gemini, Grok by X and China’s upstart DeepSeek, the term “large language models,” or LLMs, is often tossed around almost as an afterthought. But exciting things are happening under the hood of AI platforms, where LLMs operate, and their rapid evolution is capturing the imagination of many family offices, who are intrigued by the possibilities of using AI in natural language to create valuable insights more efficiently.
There is a flipside, however, which suggests that LLMs should be adopted and deployed with caution. Poorly implemented or improperly used, they can expose sensitive data, reinforce user biases—and even offer outright lies instead of insights.
Zeros and ones, pluses and minuses
To get an idea of how LLMs work, let’s consider perhaps the buzziest of buzzy AI offerings, DeepSeek, which grabbed headlines earlier this year. It was hailed as a game-changer not only because users found value in its insights, but also because it appeared to achieve those outcomes using fewer resources than its competitors, says Sidney Shapiro, assistant professor of business analytics at the Dhillon School of Business, University of Lethbridge.
“Think of each word we use as something printed on thousands of fridge magnets,” he says. “You could arrange them in any order you like, but statistically you have a very high chance of arranging the words in a very particular pattern so they make sense. LLMs tokenize these fridge magnets so they’re represented by numbers, and they look at the statistical patterns of how words have been used before, so they fit together like a jigsaw puzzle.”
The challenge, however, is that as users demand more from LLMs, the computing cost of processing all those words rises exponentially. While other platforms have largely relied on massive hardware scale-ups through data centre construction to process more requests, DeepSeek appears to have found a workaround: software acceleration that allows a greater capacity for understanding natural language queries and offering better answers.
Better answers, but still not necessarily accurate answers, says Vered Shwartz, assistant professor of computer science at the University of British Columbia and a Canadian Institute for Advanced Research (CIAR) AI chair at the Vector Institute. LLMs don’t lie, she says—they just do their best to answer natural language queries, even if those answers sometimes fail to meet any reasonable (human) standards of accuracy.
“LLMs can ‘hallucinate’ by confidently delivering inaccurate information,” Shwartz says. “As our queries become more complex, it becomes harder and harder to spot mistakes in the answers. I recently saw a query where a person asked AI to assess a short story they’d written, and it came back with what looked like some valuable insights because it was asked to. Ultimately, when asked the proper question, the AI admitted it could not open the PDF it was asked to analyze.”
Large, but limited (for now)
Shwartz points out that although LLMs are increasingly presented as human-equivalent, they’re still limited in understanding context and the implied meaning of both queries and answers that people instinctively understand. The language of family office business culture may be even more alien to them. They may also provide answers that are unconsciously biased by the source material with which they’re trained by the programmers. Even the way queries are worded can influence the answer.
Experts have noted, for example, that LLMs generally provide positive information above accurate information because that’s what they find in source material. They’re also likely to produce abbreviated lists that quit at two or three points, because short lists are valued above long ones in training material. Explicit wording can help to overcome these limitations. For example, “Analyze this proposal” might give an LLM too much leeway to indulge its weaknesses; on the other hand, “Tell me what is wrong with this proposal” or “Search the Internet for every negative comment about the company in which I’m considering investing” may produce more accurate and useful answers.
Financial management software that leverages LLMs may also fail to understand what’s important to a family office. Shapiro says that LLMs can be trained to learn special patterns of language and understand which associations in family office data, for example, should receive higher priority than others.
Is it nice to share?
And then there’s security. If third-party LLM software exists in the cloud, family offices should be mindful of the likelihood that natural language queries incorporating proprietary data can be uploaded, potentially exposing the information.
“AI platforms in general say that they won’t use or share your data, but you have to trust that that’s the case,” Shapiro says. “It really comes down to whether you believe that data is actually being sequestered in the way that they say it will be. More companies are now looking at building internal systems where they host a server and they’re running their own LLM. That can be expensive because you have to have a lot more IT support and hardware infrastructure.”
Naturally, where there’s a need, a service provider will arise. Among them in the private LLM space is Toronto-based FutureVault. It’s an AI-powered digital vault platform that automates, aggregates and centralizes client, advisor and enterprise documentation and extracts critical data from secure documents via private LLMs.
“Our AI-powered tools leverage foundational private LLMs to interpret and synthesize vast amounts of data into meaningful, actionable insights specific for the modern family office,” says Scott Paterson, co-founder and executive chairman of FutureVault. “Unlike generic AI solutions, FutureVault’s technology is purpose-built to meet the unique needs of wealth management enterprises and family offices, focusing on precision, privacy and scalability.”
The private LLMs leveraged by FutureVault are trained, he says, to be family office-savvy, helping to ensure that natural language queries result in secure, accurate and contextually driven insights, aligned with the unique needs of family offices and their clients.
“We fine-tune our private LLMs with domain-specific data and expertise, while also ensuring we add confidence intervals behind every use of AI across the platform,” Paterson says. “Our models are also designed to incorporate human feedback, ensuring that the insights provided align with real-world use cases and that they are relevant in all scenarios. By combining advanced AI with human oversight, we create a system that delivers meaningful insights at both a technical and emotional level, helping advisors foster trust and deeper relationships with their clients.”
Separating the reality from the hype
So, should family offices embrace LLMs as a solution? Some experts are skeptical. Buyers of any products or services advertising their “LLM” credentials, they say, should first ensure that the feature can actually deliver the promised real-world benefits to their enterprise. There is, after all, plenty of hype around AI generally and LLMs in particular—a fact that is reflected in the soaring valuations of LLM-driven startups.
Dan Conner, a CPA and general partner at Ascend Venture Capital, an early-stage VC firm in St. Louis, says his firm has been “cautious” about investing in companies that leverage LLMs as a core part of their strategy.
“We’re so clearly in the upswing of the hype cycle that the revenue models haven’t yet proven sustainable, and valuations, absent directly measurable criteria, are being stacked higher and higher by comparisons to recent AI deal history,” he says. “At this point, no quantitative analysis can explain the market dynamics at hand. AI start-ups are raising stacks of cheques at nosebleed valuations with little more than a keyword-stuffed pitch deck—sometimes even less.”
Shapiro says that products offering LLMs will continue to improve and develop. At this point, most of the advances in LLMs continue to be made in frontier data centres where training new models can cost $100 million. The future, however, likely belongs to more dedicated AI and associated LLMs that are designed for specific applications—like managing a family office.
“Companies are now realizing they could use AI chips in our phones, our cars and our microwaves that are trained to do very simple jobs like turn on an air conditioner or bake a potato in six minutes,” Shapiro says. “They don’t need to know about the habitats of penguins and all the other accumulation of human knowledge. Those chips don’t yet exist, so in the meantime, hardware acceleration is likely going to be much slower than software development in meeting the demand for LLM-enhanced services that we currently see.”
By embracing natural language, LLMs give AI an appealingly human face. Yet family offices considering adopting these platforms would be wise to look beyond the hype to learn more about the costs, the benefits, what they can and can’t do, and the risks of improperly using them—before they find out the hard way.
The Canadian Family Offices newsletter comes out on Sundays and Wednesdays. If you are interested in stories about Canadian enterprising families, family offices and the professionals who work with them, but like your content aggregated, you can sign up for our free newsletter here.
Please visit here to see information about our standards of journalistic excellence.