Lean teams. A volume of documents and reporting wildly out of proportion to headcount. Portfolios spanning public markets, private funds and several jurisdictions, each generating paper somebody has to read. By any reasonable technology assessment, the family office is close to an ideal early adopter of generative AI. Citi Institute’s May 2026 research puts adoption for operational tasks or investment analysis at 22%.
The interesting part is not the shortfall. It is that the hesitation is largely deliberate, and rests on constraints the technology has not yet answered.
Rising, but not along the institutional path
The 22% is up from around 13% in 2024, and specific functions are moving faster still: investment performance reporting has reached 16% and portfolio construction 13%, both more than doubling over the same period. The direction is not in question. The route is.
Citi Institute’s parallel work on institutional managers describes a three-stage progression — middle- and back-office automation through 2023 and 2024, front-office exploration with research assistants and investment co-pilots through 2024 and 2025, and now agentic acceleration, where systems plan and execute multi-step research and monitoring with limited human involvement. Family offices are not travelling that road slowly. They are on a different one, shaped by pressures institutions do not share to the same degree: acute sensitivity around family privacy, decision-making that is principal-driven rather than committee-driven, a preference for owning infrastructure outright, and a reluctance to add headcount even where the technology would justify a specialist.
One of those pressures is decisive. Lack of internal expertise is cited by 57% of offices as the largest barrier to adoption — more often than any other constraint. An organisation with no one whose job is to evaluate a tool, configure it safely and take responsibility for its output has a rational reason to defer, whatever the tool can do in a demonstration.
Efficiency, not alpha
What offices actually use AI for makes the intent plain. The prime goal is leanness rather than outperformance. Document summarisation sits at the mature end: condensing hundred-page investment memoranda, standardising private placement memoranda into comparable summaries for manager selection, extracting key terms from legal and regulatory papers. One principal interviewed for the research put the boundary as clearly as anyone has: “AI is an assistant, not a decision-maker... Human legal review remains essential.”
Meeting transcription has followed the same path — hours of manager meetings converted into notes with an audit trail — as has email and communication management. Reporting automation has proved more stubborn. Monthly adviser reports are still frequently consolidated by hand, driven as much by doubt about the accuracy of generated output as by privacy. That is a substantive obstacle rather than conservatism: reporting is the one output a family office cannot afford to be approximately right about.
Privacy is not one barrier among several
Family office data sits at an intersection few institutional contexts occupy: highly sensitive financial information combined with deeply personal family information, held by an organisation whose value to the family rests partly on discretion. The prospect that a third-party tool might retain or expose that combination shapes almost every technology decision an office makes, and 28% cite cybersecurity or privacy concerns as a barrier to adopting technology at all.
The research raises one point that deserves more attention than it usually gets: a number of offices may already have AI exposure through the back door, via SaaS products and everyday devices with features switched on by default. Exposure nobody chose is harder to govern than exposure somebody selected. The pattern the research associates with better practice is sequencing rather than abstention — beginning with on-premise or locally run models to build internal confidence, insisting on properly configured enterprise versions where external tools are used, with data residency, access controls and audit trails, and classifying information clearly enough that staff know what may never leave the building. None of it is exotic. It is ordinary information-security discipline applied to an unfamiliar category of tool.
Adoption is arriving from below
Across Citi Institute’s interviews, the impetus frequently came from junior analysts and younger family members — often the third generation — who experimented, demonstrated value and were subsequently formalised by senior leadership. Where that enthusiasm meets indifference, the research identifies a retention risk worth taking seriously: professionals who regard these tools as basic equipment tend to move to organisations that treat them that way.
The same generational split can be managed rather than merely tolerated. Older family members typically prefer meetings, calls and detailed written reports; younger ones expect digital-first access and analysis on demand. Several offices in the research now automate traditional reporting for one group while providing interactive analytics to the other, drawing on identical underlying data. It is a modest use of the technology, and a good example of what it is currently best at: presenting the same truth in more than one form without multiplying the work.
Buy, build, or wait
Implementation shows a genuine split rather than a consensus. Most offices rely on consumer or enterprise versions of major large language models — ChatGPT was mentioned most often, with Microsoft Copilot where the office already sits in that ecosystem — typically standardising on one model for budgetary reasons, though interest in comparing frontier models is growing. A smaller group prefers open-source small language models run on-premise, accepting reduced capability in exchange for lower cost and less exposure to dominant providers. Agentic systems remain, by the research’s own account, nascent in family offices; where teams of specialised agents are being tested, the technology is described as unreliable rather than production-ready.
The more disciplined offices apply a high bar before committing resources, with several describing an efficiency threshold above 80% before a bespoke build is considered worthwhile, preferring to wait for a mature product than fund something that falls short. Declining to buy every new capability is not caution for its own sake. In a sector where the cost of a failed system includes the data it was fed, it is a defensible allocation of scarce internal attention.
The line the sector has drawn
Running through every finding is a boundary between AI as operational infrastructure and AI as investment decision-maker. Offices are using it for deal sourcing and diligence support — surfacing opportunities in thinly covered private markets, screening manager research, flagging anomalies in financial statements — and insisting on a human in the loop wherever a decision follows. The insistence reflects unresolved uncertainty about whether generated investment research is reliable enough to be decision-ready, as opposed to a useful first pass that still requires verification.
Worth stating plainly: nothing in this research supports family offices delegating investment judgement to AI systems, now or in any near-term horizon it describes. What it does support is a firmer case for AI as the operational backbone beneath a lean team — summarising, transcribing, extracting, flagging — so that the people who remain accountable spend more of their time on the judgements only they can be held to. Whether an office reaches that through frontier models, on-premise open-source or a specialist partnership matters considerably less than whether it can explain, to the family, exactly where its information goes.
"Nothing in this research supports family offices delegating investment judgement to AI systems, now or in any near-term horizon it describes."