Greg Tselikis, director of technology advisory at SC&H, and Nick Scott, director of data analytics at SC&H, discussed these issues in the August 12, 2026 FO Pro webinar, “How Leading Family Offices Are Turning AI Investment into Real Impact,” sponsored by SC&H. Watch the replay.
More than half of family offices globally have invested in generative AI. Fewer than 15% are using it day to day.
The distance between those two numbers is where most offices now find themselves: licenses purchased, a few enthusiastic users, and no institutional result anyone can point to. The offices that have closed that gap did not buy better software than everyone else. They did the work that comes before the software.
“AI is only as powerful as what it knows,” says Nick Scott, director of data analytics at SC&H. Off-the-shelf tools “need to learn about your family office. You have to train it on context as to why you approach customers a certain way, how you operate, what your industry standards are.”
Individual wins that never become institutional
Ask a family office where AI lives today and the answer is usually a list of people. Someone uses Copilot to work through Outlook. Someone else drops two contracts into a chat window and asks where the language diverges. A third person has quietly become the office’s most capable prompt writer.
None of that is wasted effort. But very little of it survives contact with the org chart.
“The rest of the organization’s not benefiting from that individual use case,” Scott says. “If that person were to leave or move on or switch departments, that knowledge kind of gets stuck.”
This is a different adoption pattern than family offices are used to managing. Cloud, mobile and enterprise resource planning all arrived from the top: the organization bought the technology, then worked to get individuals to use it. AI has arrived from the bottom.
“It’s a little bit of an inverse,” says Greg Tselikis, director of technology advisory at SC&H. “We do see that individual layer of people in the organization really understanding that this is a tool they need to adopt, and then bring in and surface to the rest of the organization.”
Leadership’s job, in other words, is not to persuade anyone to try AI. It is to build enough structure that what individuals discover becomes something the office owns — shared workspaces and projects rather than private chat histories, documented practice rather than personal habit, and governance that keeps the whole enterprise inside the fence.
Assess before you automate
The first stage is the one most offices want to skip, and the argument for it is financial rather than procedural. Skip the assessment and you get the project that is a quarter built before someone discovers the underlying data was never captured.
“It’s going to save you time and money,” Scott says, “versus when you start going down the road of trying to roll something out or build something that can get about 25% of the way there — and then you figure out we didn’t have the right things in place.”
Assessment covers three things: data, processes and people. On an engagement with a national nonprofit operating in several states, Scott and Tselikis spent the early site visits barely discussing AI at all. They watched how the work actually happened, asked where staff were putting in extra hours and went behind the scenes of the proprietary platforms running operations to learn whether the data could even be reached. They also looked hard at whether the people in a given department were positioned to adopt something new.
Roughly 20 candidate use cases surfaced. The short list was considerably shorter, drawn on the basis of clean data and a willing group of stakeholders.
Notably, the first projects were not the highest-return options available. Those depended on data or on a people structure that did not yet exist.
Both advisors are emphatic that the ambient sense of being behind is mostly imaginary. “What we’re really finding is, for organizations that we work with, they’re no further ahead or behind than anyone else,” Tselikis says. “There is still plenty of time to approach this with a measured crawl, walk, run approach.”
Context is the whole game
At a desk, context is whatever files a user drags into the window. At institutional scale, context means the source systems — the general ledger, the investment records, whatever runs operations. Which is why the data conversation has to precede the AI conversation, and why one of the most frequently requested projects is among the worst places to start.
“The first ask we get is, I want to get rid of all these dashboards and reports, I just want to be able to ask a question of anything about my data and have the answer come back to me,” Scott says. Without a foundation structured for a language model to read, that portal will produce confident, wrong answers. “It’s very easy to stand something up that can talk to your data, but it’s not going to give you very good answers.”
His advice on giving an executive team or family members open-ended query access across every data set in the office: do not go there first.
The more encouraging finding is that AI is genuinely useful for preparing data for AI. It can de-duplicate vendor, customer and holdings lists, and it can install validation checks at the point of capture so records arrive clean going forward. On one engagement cleaning up the records of a newly acquired manufacturer, Tselikis estimates the work ran about 60% more efficiently than it would have without AI in the mix.
Small, measurable and contained
SC&H sorts candidate projects into three tiers, which map to crawl, walk and run. First-tier work passes three tests: the benefit is measurable, the scope is contained to a single department or objective, and the pilot is affordable. Third-tier work is the connected, cross-system, agentic activity that dominates vendor marketing.
The counsel most likely to save an office money is that there is no obligation to move up.
“It’s okay to stay at tier one for as long as your organization needs to,” Tselikis says. Most offices, once assessed, have far more first-tier opportunity than capacity to deliver it — and that work is both the easiest and the cheapest to solve. “It’s not a race to the finish line.”
Asked what family offices most often try to automate too early, Tselikis names two things. The first is work that turns on human judgment. Finance is full of good repetitive use cases, but tasks requiring genuine decision-making — in-depth financial statement review, for instance — disappoint, because the tool loses the context that matters. The second is anything built on unhealthy data.
Scott adds a third caution: most family offices are not software development shops. A rough internal interface is easy to stand up and hard to finish. “How are you going to secure that from a login standpoint? How are you going to make sure people can’t then access data that they shouldn’t be accessing?” he says. “When it breaks, who’s going to maintain or own it?”
A policy is not a control
For an office holding a family’s financial life, governance is the stage with the least room for improvisation — and an acceptable-use policy, on its own, is not governance.
“Don’t underestimate the amount of shadow AI that may be happening within your organization,” Tselikis says. A policy listing approved and banned tools is necessary but insufficient. Leadership should expect IT to answer concretely which AI tools are in use, what data is going into them and where it is leaking. That is not surveillance, he says, but visibility: tools available today will flag an unapproved model or a document marked confidential being pasted into a chat window.
He also punctures a comfortable assumption. Paying for an AI subscription does not by itself protect anything.
“There’s been common guidance that if you’re paying for an AI app, it must be secure and they’re not using your data. That is not necessarily true,” Tselikis says. On the major platforms, chats on personal and consumer plans can feed model improvement unless the setting is switched off, while team and enterprise plans are contractually excluded from training. For an office handling family financial information, that distinction is the one that matters.
Alongside it belongs a real vendor evaluation process — SOC 2 certification, where the data physically resides and how it is protected. Most SC&H clients run off-the-shelf tools on enterprise plans rather than private models. At SC&H itself the controls are hard-wired: staff laptops cannot reach ChatGPT at all, and the approved stack is Copilot and Claude Enterprise.
One more guardrail belongs in every policy, and it has nothing to do with software. An employee who sends AI-drafted work owns that work. “It doesn’t lift responsibility of what we do in our day-to-day to make sure that we’re thorough and accurate,” Scott says.
Training is where the return shows up
Tselikis has walked into engagements where a chief financial officer has a full year of AI spend and nothing to attribute it to. The cause is almost always the same.
“The organization bought an AI tool, told people where to go log into it, and the right structured, ongoing, repeatable training wasn’t in place to foster the use,” he says. Turning that training on has reversed the picture in a number of those situations.
The word both advisors use is fluency — whether an office has enough AI in its DNA that people instinctively ask whether the tool could help with the task in front of them. In practice that means a core group of power users, training on a platform’s shared features rather than only its chat box, and guidance on matching the model to the job. Running the most capable and most expensive model on every routine request is a reliable source of costs no one can justify later.
What it looks like when it works
Two results suggest the scale of what is available at the low-risk end.
A manufacturer needed a total addressable market analysis for its board, pulling together Dun & Bradstreet data, outside market research and its own revenue and product mix. Internally, the team estimated six weeks. Using an enterprise AI license and a shared project to hold the accumulated context, the analysis came together in hours. The detail that matters for an office weighing a custom build: this was off-the-shelf software.
In the second case, an organization placing students in foreign exchange programs had advisors buried under a 48-hour inbox backlog, spending hours researching complex placement questions. Response time dropped to minutes, freeing a lean team to serve more students.
Neither was a moonshot. Both were measurable, contained and affordable — which is, more or less, the point.