I recently came across an interesting example of an AI-generated family office report. And, I think it’s pretty cool to see how AI can generate consolidated financial reports in the same class as the incumbent software providers (at a fraction of their cost). This development also demonstrates how quickly the reporting landscape is changing, and also of why I don’t think professional family offices are ready to abandon dedicated reporting systems just yet.
The three-page quarterly summary generated by AI is impressively straightforward. According to the document, it was created by asking Claude a plain-English question: essentially, “summarize the family’s quarter”. Through a connector that allowed the AI to read the general ledger directly, entity by entity and transaction by transaction. The report describes the process as requiring “no exports, no reformatting.”
A basic report was generated that asked the basic questions that may be asked.
Its also getting easier to do this.
AI is becoming very good at creating the report
Once an AI tool has access to reliable underlying data, producing a polished summary is no longer particularly difficult.
The example I reviewed includes a concise narrative of the quarter, a bridge showing the change in family net worth, a comparison of assets between periods, an asset allocation summary, partnership capital information, cash-flow information and several items flagged for the principals’ attention.
The net-worth table, for example, compares cash, public investments, private investments, real estate, art, a private operating company and liabilities between a quarterly timeframe.
This is exactly the sort of work where generative AI shines. Give it structured financial information and it can:
- identify material changes;
- translate accounting data into plain English;
- identify unusual transactions;
- produce tables and charts;
- write an executive summary; and
- answer follow-up questions about the underlying records.
For an individual family office, this can make ad hoc reporting dramatically easier.
But there is an important distinction between generating a report and operating a reporting system.
The difficult part is still consistency and auditability
Imagine putting an AI reporting system into practice and receiving one of these reports every quarter for five years.
A decision maker doesn’t simply want an intelligent summary of what happened this quarter. They want to be able to compare this quarter with last quarter, last year and the original investment plan.
Reporting also needs to respond to input from readers. Things like “show me this” and “show it to me this way instead of that way”. Staff can configure the current performance reporting software to meet customization requests. The reports get better over time, and the company providing the software can make improvements over time too, while rolling out those improvements to all users.
This can still happen with AI reporting, but the burden is no longer on a software provider. The burden falls on the family office itself.
I’ve heard a lot of people ask me recently, “when will you dump Addepar and just use AI”? And, while it is technically possible now. Producing a comparable output using AI at the same standard as our current software providers shifts the development and maintenance burden to our team.
Simply because something is possible, doesn’t mean it’s the best choice. Our family office needs to consider more than just output. We need to consider doing our work at consistency and scale.
- Was private credit classified in the same way last quarter?
- Was a particular investment included in private equity or real assets?
- Was performance calculated using the same methodology?
- Did the benchmark change?
- Did the definition of net worth change?
- Are entity ownership structures being treated consistently?
AI can follow templates, and it will undoubtedly improve its consistency. But professional reporting software has an important advantage: the reporting methodology is embedded in the system itself.
A platform such as Addepar does more than create attractive reports. It warehouses information in a structured format. Securities, entities, accounts, ownership structures, asset classifications, benchmarks, transactions and valuations are stored within a consistent data model.
A report is an output of a controlled dataset rather than a new interpretation of source material every time somebody asks a question.
That distinction matters.
AI still needs a data warehouse
None of this means a family office must buy commercial reporting software.
It is entirely possible to build your own reporting infrastructure using AI, you just have to do the work and create the logic yourself. And, you have to maintain it. The family office needs to document that knowledge to create resiliency for the families they work with.
A family office could maintain its information in a relational database. It could use Excel workbooks, CSV files or structured text files. It could maintain a collection of standardized spreadsheets in Google Drive and give an AI system access to them.
With enough discipline, that architecture can work very well.
But somebody must still decide:
- what the authoritative data source is;
- how investments are classified;
- how benchmarks are assigned;
- how historical information is preserved;
- how corrections are handled;
- how valuations are documented;
- how performance is calculated; and
- who is responsible for validating the information.
In other words, eliminating the reporting platform doesn’t eliminate the reporting infrastructure required.
In fact, it simply transfers more responsibility for building and maintaining that infrastructure to the family office itself.
That can be perfectly reasonable for a sophisticated organization with lots of staff (and overhead) or one with relatively simple reporting requirements. But it shouldn’t be confused with simply connecting AI to a collection of financial records.
The scale problem
There are also practical operational challenges for the people actually doing the work.
Producing an intelligent ad hoc report for one family is one thing.
Producing many reports every month or quarter—consistently, on schedule, using the same methodology—is another.
A single-family office might reasonably experiment with an AI-generated quarterly briefing. If something looks strange, somebody who understands the family’s affairs can investigate it.
A multi-family office serving 10 families has a different problem.
Now there may be many entities, lots of accounts, enormous volumes of transactions and many different people relying upon the reports.
At that point, repeatability becomes extremely important.
Professional reporting software provides structure around ingestion, reconciliation, classifications, calculations and presentation. AI can sit on top of that structure and make it substantially more useful. But asking AI itself to become both the database, calculation engine, quality-control system and presentation layer creates a much greater operational burden.
What is missing from an AI sample?
The sample AI report I recently reviewed is effective as an executive briefing. It tells the principals what changed and highlights several issues worth discussing. Its relatively easy to produce this as a “one-off”.
But it isn’t yet what I would consider a complete investment reporting package.
For example, the report is missing investment performance information, including attribution.
If the portfolio gained $37.5 million, how much of that resulted from market performance, foreign exchange, asset allocation, manager selection or additional capital?
Where are risk measurements?
Volatility, drawdowns and other risk metrics are much more useful when they are calculated using exactly the same methodology every reporting period. The same applies to benchmarks and performance calculations.
An AI can certainly calculate these things. The issue is not whether the math is possible. The issue is whether the methodology is controlled, documented and repeatable.
Is the report audit-ready?
Interestingly, the sample AI report I reviewed itself says that every figure ties to the full reporting package and the general ledger. Great.
That is an important control, and its much stronger than simply asking an AI system to interpret a pile of statements.
But more questions are still unanswered.
If an accountant, auditor, trustee, investment committee member or family member challenges a number two years later, can we reconstruct exactly where it came from?
- Can we identify the source transaction?
- Can we reproduce the calculation?
- Can we determine which valuation was being used at the reporting date?
- Can we identify who changed a classification and when?
- Can we demonstrate that the methodology being used today is the same methodology that was used historically?
Those are data-governance questions more than AI questions.
A beautiful report is the visible end-product. The less visible infrastructure underneath it is what makes the report dependable.
Where I think this is going
Our team uses AI extensively.
We use it to help interpret information, review work, identify inconsistencies, answer questions and improve internal processes.
But today, we don’t rely on AI alone as the front-end performance reporting system for our families.
We continue using professional reporting software because we value having a structured, persistent and consistently maintained dataset underneath the reporting process.
That may change.
I can easily imagine a future in which the traditional reporting interface largely disappears. But something more robust will take its place, and someone still needs to develop and deploy it at a professional scale.
Rather than logging into a reporting platform and navigating through predefined reports, a family member might simply ask:
- How did we perform last quarter?
- Why did our Canadian equity portfolio underperform?
- How much liquidity do we have available for the next 24 months of capital calls?
- Which private investments have materially changed in value?
The AI will create the report dynamically.
But even in that world (which is coming, no doubt), something underneath the AI still needs to maintain the accounting records, security master, classifications, benchmarks, valuations and historical data.
The interface may become conversational.
The need for reliable financial infrastructure remains.
That, to me, is the important lesson from examples like the AI reports we are not reviewing.
AI is rapidly commoditizing the presentation layer of family office reporting.
The harder problem, and ultimately the more valuable one for decision makers, is maintaining organized, consistent and well-governed data underneath it.
