Letting a Principal Ask the Portfolio a Question Without Getting a Hallucination Back

A principal at a single-family office in Abu Dhabi wants to ask one question, in plain English, on a Sunday evening: “How much of my wealth is in private credit right now, and what did it return last quarter?” He does not want a 60-page consolidated pack. He wants the number, and he wants it to be right.

That request sounds trivial. It is not. Behind it sits a private equity fund whose NAV arrived 60 days after quarter-end, a private credit sleeve reported in a different format by a different administrator, listed positions sitting at a Swiss custodian and a Gulf one, a real-asset holding last valued nine months ago, and a couple of co-investments tracked in a spreadsheet someone updates by hand. The honest answer to his question is not one number. It’s several numbers, captured at different moments, denominated in different currencies, defined under different conventions. Put a chatbot on top of that mess and ask it to answer in natural language, and you have built a very confident way to mislead the person whose money it is.

This is the specific problem GCC family offices are walking into in 2026, and it deserves more care than the demos suggest.

Why this office is harder than a corporate ledger

GCC single-family offices have grown into serious allocators. Across the region nearly 80% of family-office investors plan to increase private-equity exposure, and close to half already hold more than 20% of AUM in PE alone, with real assets and private credit stacked on top. That allocation mix is exactly what makes natural-language querying dangerous.

A corporate finance team mostly reconciles ledgers that update daily. A family office reconciles reporting streams that update on wildly different clocks. PE NAVs commonly land 45 to 75 days after quarter-end, and many investors carry an extra quarter of lag before that figure reaches their own books, so a “current” allocation can quietly be a six-month-old picture. Roughly two-thirds of family offices still aggregate by hand. Private-market data is repeatedly cited as the single biggest technology problem these offices face. None of that is a model problem. It is a data problem, and a model placed on top of it inherits every gap.

So when the principal asks for his private-credit weighting, the real question is: as of when, valued how, in which currency, net of which fees? A language model does not know to ask that. It pattern-matches to a plausible-looking percentage and hands it over with the same calm tone it uses for everything.

RAG over PDFs is the wrong foundation

The tempting build is to take every quarterly statement, capital account, and custodian report, drop them into a vector store, and let retrieval-augmented generation answer questions over the documents. It demos beautifully. It fails quietly.

Three reasons. First, these documents are tables, and language models read tables badly. Capital accounts, NAV roll-forwards, and fee schedules are dense grids where the meaning lives in the relationship between cells, and that is precisely where extraction breaks. Second, RAG retrieves and summarizes; it does not compute. Ask for a blended return across four funds and the model has to actually do weighted math, and an LLM doing arithmetic across retrieved snippets is guessing, not calculating. Third, RAG has no concept of “as of.” It will happily blend a Q1 NAV with a Q3 market price into a single answer because both fragments mention the same fund, and nothing in the retrieval step flags that you have just compared two different points in time.

The result is an answer that is fluent, sourced-looking, and wrong in a way that takes an analyst an hour to unpick. For a principal making an allocation decision, that hour is the whole point of asking.

The order that actually works: reconcile, then ask

The fix is not a smarter prompt. It is putting the reasoning and the math in different places, and putting a reconciled model between them.

First, aggregate and reconcile into one model. Pull every source, custodian feeds, fund administrator statements, capital accounts, the spreadsheets, into a single structure where each position has an explicit valuation date, currency, asset class, ownership chain, and a link back to the document it came from. Where a NAV is stale, the model knows it is stale and carries that flag, rather than pretending the number is fresh. This is the same single-source-of-truth discipline a multi-entity group needs, and we make the consolidation case in detail in one reconciled model across 40 SPVs.

Second, run the math deterministically. When the principal asks for his private-credit weighting or a blended quarterly return, the calculation runs through a fixed engine, not the language model. Weighted returns, currency conversion at a defined rate on a defined date, allocation percentages against a defined total: these are formulas, and formulas should be executed, not predicted. The language model’s job is to understand the question and phrase the answer. It never touches the arithmetic.

Third, make every answer traceable. The reply to “how much is in private credit” should expand into the positions that make it up, each with its valuation date and source document. If a number rests on a NAV that is two quarters old, the answer says so. That single habit, showing the asof date and the lineage, is what separates a tool a principal can act on from a toy that produces a vibe. It is also the discipline the rest of GCC finance is being held to, which we cover in FY2025 audited statements and AI traceability.

Where natural language genuinely helps, and where it doesn’t

I want to be fair about the limits. Natural-language querying is a real upgrade for one thing: access. A principal who would never open a BI dashboard will type a question into a chat box, and giving them direct, conversational reach into their own wealth is genuinely valuable. The interface is not the problem.

What it cannot do is invent certainty the underlying data doesn’t have. If a fund hasn’t reported its Q4 NAV yet, no model can tell you the true current value of that position; it can only tell you the last known value and how old it is. A reconciled layer is honest about that. A chatbot over PDFs papers over it. The difference shows up the day a principal makes a real decision on a number that turned out to be six months stale.

And it cannot fix definitions you haven’t agreed. “Return” net of which fees, “wealth” including or excluding the operating business, “liquid” on what horizon? Those are governance choices the family office has to make once and encode in the model. The deterministic engine then applies them the same way every time, which is the whole point: ask the same question twice and get the same answer twice.

The takeaway

A principal asking a question in plain English is a good idea built on a fragile assumption, that the numbers underneath are already trustworthy. In a GCC family office spanning PE, private credit, real assets, and listed holdings across jurisdictions, that assumption almost never holds by default. The work is the reconciliation, not the chat interface. Build the single reconciled model first, route every calculation through a deterministic engine, and make every answer trace back to a dated source, and natural-language access becomes safe. Skip that, and you have given the most important person in the building a fluent way to be confidently wrong.

If you want to see what asking a portfolio a question looks like when every figure ties back to source, book a demo and bring your messiest fund statement.