What investors should ask their fund administrator about AI

Where does our data actually go?

Operational due diligence on an investment manager increasingly includes questions about its use of artificial intelligence — what it is used for, how it is governed, and what happens when it goes wrong. The same questions are just as relevant, and asked far less often, of the fund administrator sitting behind that manager. The questions below are offered in that spirit: not as marketing for Berkley, but as a reasonable checklist for any promoter, manager or investor assessing any administrator.

The first and most important question is the simplest: when an AI tool processes investor data, where does that data go, who can see it, and under what contractual protection? An administrator should be able to answer this specifically — naming the tool, its processing location, and the agreement governing it — rather than in general reassurances about taking data protection seriously. An answer that cannot specify where the data goes is itself an answer worth noting.

Who reviews what the tool produces?

Ask what happens between an AI tool producing a figure or a draft and that figure reaching an investor statement or a regulatory filing. A credible answer names a specific review step, performed by a specific person or role, with a record of that review kept. An administrator that cannot describe this chain clearly is relying on the tool more than it may realise.

What was vetted, and by whom?

Ask how a particular AI tool came to be adopted: who decided to use it, what alternative approaches were considered, and whether the decision and its rationale are documented anywhere an auditor or regulator could later review. This is less about the specific tool and more about whether the administrator treats AI adoption as a considered decision rather than something that crept in through individual staff experimenting on their own initiative.

What happens when something goes wrong?

Finally, ask what the administrator's process is when an AI-assisted figure turns out to be wrong — how it would be caught, who would be told, and what would change afterwards. An administrator with a mature process will have a ready answer involving exception monitoring, client notification and a documented remediation step; one without a clear answer has likely not thought the scenario through.

None of these questions require a promoter or investor to understand the underlying technology. They require only that the administrator can explain, in plain terms, where the data goes, who checks the work, and what happens if something slips through. An administrator confident in its own governance will welcome the questions rather than deflect them — Berkley's team included.