The Iceberg Problem: Why AI Can’t Build You a Reliable Family Office List

There is a question that surfaces in almost every conversation about AI and private markets intelligence: “Couldn’t we just use ChatGPT for this?” It is a reasonable question. The answer, it turns out, depends entirely on what you mean by “this” — and the distinction matters a great deal more than most people realise.

The Iceberg Problem in Family Office Data

Imagine the global family office universe as an iceberg. The visible portion above the waterline — offices that have given interviews, co-sponsored conferences, published investment theses, or appeared in deal announcements — is real, but it represents a fraction of what is actually there. The majority sits below: offices that have always operated quietly, offices that were once visible but have since retreated from public view, and offices whose most important characteristics (investment appetite, ticket size, decision-maker relationships) were never published anywhere at all.

AI systems are excellent tools for mapping the part of the iceberg above the waterline. They are, by design, incapable of seeing the rest. This is not a limitation that better models or more compute will overcome — it is a consequence of what those systems are trained on and how information flows (or doesn’t) in the family office world.

The practical consequences of this distinction become very concrete when you try to use AI to build a comprehensive family office list from scratch.

We Modelled the Attempt. Here Is the Arithmetic.

Say the goal is a list of 100 verified single family offices in the United States — with investment mandates, approximate AUM, and enough detail to be useful for origination or relationship development. How would you actually do this systematically with an AI API? And what would it cost?

There is no shortcut to a reliable answer. You would need to work through at least three major data streams, plus a structured profiling phase for every candidate you surface.

Step 1 — Work the Wealth Lists

The Forbes 400 is the logical entry point. Individuals with personal wealth above $1 billion almost always establish a family office; the question is what kind. Confirming whether a specific billionaire runs a single family office — as opposed to a multi-family arrangement or a private banking mandate — requires targeted research for each name: searching news and filings for entity names, checking EDGAR for registered adviser records, verifying structure, extracting investment focus, and identifying key people. That runs to around 7 queries per person.

400 names × 7 queries = 2,800 queries. Expected yield: ~150 genuine SFO leads.

Step 2 — Scan Transaction Data

Family offices are among the most active — and least visible — buyers in commercial real estate. Identifying them requires monitoring deal flow and working backwards from buyer identity. MSCI Real Assets records roughly 8,000–10,000 U.S. commercial property transactions above $10 million annually.^[1] Each requires at minimum two queries to check for family office involvement.

~9,000 transactions × 2 queries = 18,000 queries. Incremental yield: ~80 additional SFO leads.

A similar logic applies to VC and PE deal flow. PitchBook records approximately 20,000 U.S. private market transactions per year, of which roughly 3,000 show documented or probable family office participation.^[2] Verifying investor identity for each runs to three queries on average.

3,000 deals × 3 queries = 9,000 queries. Incremental yield: ~60 additional SFO leads.

Step 3 — Profile Every Candidate

At this point you have approximately 200 unique candidates after deduplication — more than you need, but with significant noise. Each requires structured profiling before it becomes usable: confirmation of single-family status, AUM estimate, sector and geographic focus, portfolio positions, personnel, and contact information. Around 8 queries per candidate.

200 candidates × 8 queries = 1,600 queries.

31,400 queries. Each one carries a token payload — context window, disambiguation instructions, prior findings, structured output requirements. A realistic per-query average: 1,500 input tokens and 800 output tokens. Across the full run: 47.1 million input tokens and 25.1 million output tokens, for a combined total of roughly 72 million tokens.

The Bill, By Model

API pricing at April 2025 standard rates: Anthropic, OpenAI, Google AI Studio. No batch or caching discounts applied.

Those figures assume 100% accuracy on every query — no hallucinated entities, no conflation of holding companies with family offices, no SFO/MFO misclassification. In practice, LLMs working from public data achieve something closer to 70–80% reliability on this kind of structured entity research. Adjusting for that, effective cost per usable entry climbs to $4.50–$7.40 across the three models.

There is also an environmental dimension worth noting: at a conservative 0.003 kWh per query (per Luccioni et al., 2023^[3]), the full research run consumes roughly 94 kWh and generates approximately 36 kg of CO₂ against the U.S. grid average — equivalent to driving 150 miles in a standard petrol vehicle, in exchange for a dataset that requires further manual verification.^[4]

What the Model Cannot See, and Why That Is Permanent

The cost argument is actually the weaker of the two cases against AI-generated family office lists at scale. The stronger one is about knowledge — specifically, about what kind of knowledge exists in the world and where it lives.

A meaningful share of family office intelligence was never digitised in a form that would end up in a training corpus. Some of it was online once and is no longer: offices that appeared in a 2016 conference programme, were mentioned in a regional business newspaper that has since shuttered, or were listed in a directory that no longer exists. Crawling the internet today returns nothing — but those offices are real, active, and investable.

A further layer of intelligence was never published at all. An office’s actual risk appetite, its minimum cheque size, its preference for club deals over fund commitments, the fact that decisions run through one specific family member rather than the official CIO — none of this appears in any source an AI can access. It emerges from conversations, from co-investments, from introductions that led to follow-up calls that led to relationships. It lives in networks, not databases. And it is precisely the intelligence that makes the difference between a cold list and a warm one.

The Real Comparison

When positioned against curated, research-backed alternatives, the AI-generated approach loses on both dimensions that actually matter: cost per reliable entry and depth of coverage. Specialist providers of family office databases — such as the global single family office lists available from dedicated research platforms — typically price a U.S.-focused list of 500+ verified SFOs at around $800.

Adjusted figures reflect ~75% AI accuracy rate on entity verification. Curated database figure assumes $800 / 500 entries.

So When Does AI Actually Make Sense Here?

This is not an argument that AI has no role in family office research. It clearly does — in the right context.

The useful heuristic is scope. For a mandate that requires five to fifteen highly specific names — a particular geography, a defined sector, a minimum AUM threshold — AI-assisted research can return a reasonable shortlist quickly and cheaply, with acceptable error tolerance. The same applies to point-in-time enrichment: adding a specific data field to an existing list, or cross-checking a name against a particular event or announcement.

Where it breaks down is precisely where the investment in curated data pays off: when you need comprehensive market coverage, when the completeness of the list is material to the outcome, and when the intelligence you need has never been, or is no longer, publicly available.