AI in fund reporting: who writes the numbers?

Every investment fund now hears that AI will transform reporting. The demonstrations are seductive: a quarter's numbers in, a polished letter out, hours collapsing into minutes. And somewhere in the room, the person who signs that letter feels a correct unease, because the question that matters is not how fast the letter appeared. It is who wrote the numbers.

Two different jobs wearing one name

Reporting is really two activities. The first is assembly: gathering positions and movements from administrators, custodians and internal records, reconciling them, laying them against last quarter and against the mandate. Mechanical, unforgiving, endlessly repeated. The second is the account: explaining what happened and why, in the fund's voice, with the judgement of the people responsible. Investors pay fees for the second. Teams lose their weeks to the first.

AI's honest role splits along exactly that line. Assembly is structured, rule-bound and checkable, precisely the work current systems do well and people do resentfully. A system that watches the portfolio continuously, reconciles as data arrives and flags what moved outside expectation does not replace anyone's judgement; it delivers the quarter pre-assembled so that judgement starts on day one instead of day nine.

The account is different. A model can draft fluent commentary, and fluent is the danger. A paragraph that reads perfectly and misattributes performance by a subtle error is worse than a rough draft that is right, because polish disarms review. The rule that keeps the account honest: systems draft from the firm's verified numbers and the firm's own prior letters, every claim in the draft traceable to its source, and a person owns every sentence that goes out. The letter's signature is a human promise, and no fund should let software make promises.

The continuous portfolio

The deeper change is not faster quarters. It is that reporting stops being an event. A portfolio watched continuously can report its own movement as it happens: the position that breached a limit this morning, the concentration drifting toward a threshold, the number that will be awkward in the LP call and the explanation already prepared. The quarterly letter becomes a summary of things already known rather than an archaeology project. Funds that reach this state do not merely save time; they stop being surprised by their own portfolio.

The trust questions, which are the real questions

Fund data is mandate strategy in numerical form, and LPs increasingly ask their own diligence questions about how managers use AI. Four answers a fund should be able to give. Our data is dedicated: no pooling across customers of any supplier, nothing training shared models. Every automated step is logged with its reasoning and reviewable, so an auditor or an LP can reconstruct any number's path. A person approves everything that leaves the firm. And our exit is contractual: everything returned readable, everything deleted at the supplier.

A fund that can give those four answers has not weakened its controls by adopting AI. It has usually strengthened them, because the log of an assembled, traced quarter is better evidence than the mailbox archaeology it replaced.

Where we stand

We build this as one Brain per fund: the portfolio watched continuously against the mandate, movement surfaced with its reasoning, reporting assembled from verified numbers and drafted in the fund's own voice for people to approve. Nothing pooled, everything logged, shadow before live. The shape of it for investment funds is on the page for MANDATE, the Fund Brain.