By Hannah Arrighi.
An accessible solution for many families is a tech stack that integrates cutting-edge AI models with existing technology service providers. This allows the family to query and access information across platforms in real time.
Novum Partners, a $10B AUM multi-family office based in Geneva, Switzerland, is evolving into an AI-enabled office using this approach. To begin, they consolidated data from various sources – Excel files, Bloomberg data, Reuters data, their CRM, their portfolio analytics platform, Guardian, and others – and created a database from scratch.
Their AI layer, Nova, sits on top of that unified database. While OpenAI was the engine that built Nova, the data itself is stored in a private, secure, internal cloud environment; data queries use API tokens with strict protections baked into the OpenAI contract, not the consumer model most people are accustomed to. Only relevant slivers of data get processed by the LLM. None of the data is used to train external AI models.
Addepar, widely regarded as best-in-class portfolio management software, is using a similar structure for its AI, Addison: it stores data on a secure cloud-based platform (using Databricks and AWS) and overlays its AI platform on top.
Like Nova, Addison can query the portfolio to show a client’s performance, exposure, asset class distribution, cash position, risk profile, and more. It can connect news articles to potential impacts on a portfolio. And, compellingly, Addison now integrates a “show your work” feature.
According to Addepar’s Head of AI, Kunal Gosar, “The AI is going to give you the nice high-level summarized approach, and the show-your-work functionality provides the underlying rationale for how it got there. We’ve seen a ton of interest and excitement around that feature because that’s what makes the AI trustworthy.”
Like Novum, a family office could integrate its own AI on top of tech software like Addepar across performance reporting, accounting, governance, and more. One of Novum’s senior bankers, Maria Mavridoglou, reflects that this significantly reduces the time needed for reconciliation and customized report generation. “In the past, the process would have been completely ad hoc and required six or seven people pulling spreadsheets together and doing manual SQL programming,” she states. “Now, it is instantly ready for review.”
However, as Gabriele Gallotti, the founder and CEO of Novum notes, “AI won’t be the miracle or hand coming down from the sky to save your business. That’s probably the hope of many, but I think that’s just a big illusion.” AI is only as good as the humans who run it. Novum has spent years crafting Nova to produce quality outputs, and it took ample dedication and resources to build a system that works.
Salash Motiani, a next-generation family office principal, has also put considerable thought and energy into his family’s use of AI. In his case, he used multiple LLMs to build his system; he deployed Claude, Cursor’s Composer, and others to create his AI workflows.
Most of his data is secured in the cloud, where the bulk of his AI system operates. However, he requires massive computational power for AI agent stock and algorithmic trading, which exceeds what the cloud can efficiently provide. Motiani thus has a dual cloud and on-premise system.
He switches between developer models opportunistically; if DeepSeek has a discount, he uses DeepSeek instead of Claude or OpenAI. For live queries, he tries to pass data relationships to the LLMs rather than the data itself – metadata rather than data. In addition, he’s architected the system so no single LLM has a monopoly on his queries. This means that none have enough information to piece together his big picture, leaving the LLMs “very disabled” by design.
He likens proper use of AI to baking – “You can use Claude, you can use Composer, you can use a DeepSeek. They all basically pick up the ingredients you give them and start cooking. So let’s say you want a cake. What we’ve done is we’ve basically given AI the direction that we need to make a cake. We’ve gone ahead and given it the ingredients also. We’ve given it the saucepans also. So, we try to give it enough things, and we simply say, now that you have everything, can you put it into a mix and get on with it? When you give it that clarity, what’s the worst it can do?
Well, even with careful instructions, someone could forget to turn the stove off and accidentally set the whole house on fire. But the better the supervision in the kitchen, the less likely that will happen.
With AI agents in particular, families need a dedicated CTO or tech team to oversee them effectively. Otherwise, the system is at risk – data could be exposed when agents interface with counterparties, or data could simply be compromised internally.
-
Hannah Arrighi is the Founder and Managing Partner at strategic communications consultancy Élever Partners. Before founding Élever Partners, she led marketing on the Investor Relations team at a global growth-stage private equity firm, where her investor materials helped raise over $1 billion.
Note: this article was originally part of a longer piece titled AI in the Family Office.
