Governance and oversight: keeping humans in the loop

Where AI tools get things wrong

A reasonable question follows from the efficiency gains described elsewhere in this series: if AI tools are getting better at reconciliation, reporting and document review, does the human review step eventually become unnecessary? The answer, for the parts of fund administration that matter to investors and regulators, is no — and understanding why is as important as understanding what the tools can do.

The errors that matter in fund administration are rarely dramatic. They are a misread clause in a limited partnership agreement that changes how a fee or a waterfall is calculated, a subtle misclassification in a document extraction that goes unnoticed because it looks plausible, or an anomaly in a reconciliation that the tool resolves quietly in the wrong direction. These are exactly the kind of errors that a confident, fluent tool is prone to producing — and exactly the kind that cause real harm if they reach an investor statement or a regulatory filing unchecked.

The industry commentary on this point is consistent: AI tools can misinterpret complex legal language, and an error in a fee or waterfall calculation can lead directly to an investor dispute or regulatory scrutiny. That risk does not diminish as the tools improve generally — it simply moves to rarer, harder-to-spot cases, which is arguably a worse outcome if the review step has been relaxed in the meantime.

What proper oversight looks like

The standard that holds up is not avoiding the tools, but keeping a person accountable for every figure that reaches an investor or a regulator. In practice that means several concrete things:

• exceptions flagged by a tool are reviewed by a person who understands the fund, not rubber-stamped;
• a clear record exists of what was reviewed, by whom, and when; and
• no calculation — a NAV, a fee, a distribution — is released without a sign-off step that a person, not a system, performs.

Explainability matters here as much as accuracy. An operations team needs to understand why an item was flagged in order to act on it sensibly, which means a tool that simply produces a number without showing its working is harder to govern than one that shows its reasoning, however imperfect that reasoning might be.

Oversight as an advantage, not a constraint

Properly applied, this discipline is not a drag on the efficiency gains described earlier in this series — it is what makes them usable. An administrator that can show a clear, documented chain from a flagged exception to a reviewed and approved figure is in a stronger position with both investors and regulators than one relying on manual process alone, because the audit trail itself becomes evidence of a well-run operation.

Continuous, AI-assisted monitoring of transactions and investor activity, reviewed by people rather than acted on automatically, gives a fund's auditors and its promoters earlier visibility into anything unusual — a genuine improvement on periodic, manual review, provided the review step at the end is never skipped.

Berkley’s standard

Berkley Fund Services applies this principle as a matter of course: AI-assisted work narrows the volume of routine matching and drafting, but every figure that reaches a client or an investor is reviewed and signed off by a person who understands the fund in question, with a documented record of that review. Promoters and managers are welcome to ask what that review process looks like for their own fund's administration.