Family offices do not need another abstract lecture about the transformative power of artificial intelligence. They need a practical answer to a simpler question: what should the technology actually do on Monday morning?
That was the strength of the discussion at the 2026 Nassau Street Partners Family Office Summit in London. Peter Danenberg, a senior software engineer at Google DeepMind, did not treat large language models as a universal solution. He described a collection of narrower applications, each with clear benefits and equally clear limitations. The resulting picture was more useful than a grand forecast. It resembled a working playbook for private capital.
The first lesson was to begin where language and unstructured information create a bottleneck. Family offices handle private placement memoranda, presentations, credit agreements, legal documents, consultant reports, and internal investment papers. Much of the work lies not in obtaining information but in reading, comparing and questioning it.
Large language models are well suited to that environment. Danenberg discussed sentiment analysis of earnings calls, where newer systems can extract signals that older models may miss. In a private-market context, the same principle can be applied carefully to management interviews, customer conversations, board materials and recurring updates from portfolio companies. The aim is not to declare whether an executive is truthful based on tone. It is to identify changes, inconsistencies and areas that deserve follow-up.
Compliance is another natural starting point. A family office may consider it less exciting than finding a new source of alpha, but the economics are compelling. A model that helps review policies, investment restrictions, and transaction documents can allow a team to identify issues earlier. Qualified professionals must make the final determination, but earlier visibility can save time and reduce avoidable risk.
The second lesson was to use AI as a challenger rather than an authority. One attendee described comparing investment memoranda against investment-committee guidelines and using the model to test whether it had answered important questions. This is a strong design principle. The system should be asked to find omissions, contradictions, and weak assumptions, not to issue a final investment recommendation.
The distinction matters because models often produce too many answers. Danenberg compared the process to asking an AI system to find bugs in a vast software codebase. It may generate a thousand potential issues, of which only a small number are meaningful. The family office therefore needs a process for ranking the output. Which concern is supported by evidence? Which one would change the investment case? Which one should be tested through management diligence, legal review or a revised downside scenario?
The third lesson was that data preparation is more important than the model name. The summit repeatedly returned to the difficulty of grounding a system in the right information. A long document can overwhelm the model, with material at the beginning and end receiving more attention than the middle. A collection of files may need to be cleaned, divided, tagged and placed into a searchable database before the model can use it reliably.
This is not glamorous work, but it is where much of the value is created. A family office with organized investment memoranda, past decisions, portfolio reports and sector research can build a knowledge system that reflects its own standards. A competitor using the same model without that foundation will not achieve the same result.
The fourth lesson was to avoid dependence on one model. An attendee explained that his organization used several systems and an additional AI tool to decide which model to use for the next quarter. Danenberg observed that the age of one model fitting every task may be ending. This is likely to be particularly true in finance, where one model may perform well on document analysis, another on structured reasoning and another on coding or data extraction.
The fifth lesson was to protect the human last mile. One participant described developing an AI version of himself to work faster and train junior staff. He was equally clear that the system would not be allowed to advise a client independently. That boundary is sensible. AI can help capture institutional knowledge, demonstrate how an experienced professional frames a problem, and provide junior colleagues with a more responsive training resource. It should not obscure who is accountable for advice.
This practical approach aligns with the growing importance of family offices in private markets. As more capital is deployed through direct investments, co-investments, private credit and specialist funds, family offices need institutional-grade processes without necessarily building the headcount of a large asset manager. AI can help close that gap. It can let a compact team review more opportunities, maintain better internal records, and answer routine portfolio questions more efficiently.
The risk is that efficiency becomes an excuse to cede judgment. Danenberg described a young engineer whose working day involved managing five to ten AI agents, constantly switching between tasks. The scene was impressive but also unsettling. If the operator becomes overwhelmed by the agents, the technology has not reduced complexity. It has merely changed its form.
Nassau Street Partners struck the right tone by allowing the event to explore both sides of the argument. The firm positioned itself not as a promoter of technological fashion, but as a convener between private capital and technical expertise. That is a constructive role. Family offices need access to new ideas, but they also need environments where those ideas can be challenged before they are embedded in investment processes.
A sensible first-year AI program for a family office would therefore be modest. Choose a small number of repetitive, document-heavy tasks. Establish a clear human reviewer. Test the system against prior cases. Record where it fails. Separate confidential information appropriately. Measure whether the workflow improves the quality or speed of a decision rather than merely producing more text.
The opportunity is significant because private markets remain information-intensive and relationship-driven. AI can make the information easier to work through. It cannot replace the relationships, the context, or the judgment that make family offices valuable counterparties. The summit’s most useful contribution was clarifying that boundary. For private investors, progress will come not from handing the process to machines, but from designing a better process around them.
