Who checks the machine? Why AI analysis needs analyzing

A fund can reconcile perfectly, comply with every internal workflow and produce a polished investor report, yet still arrive at the wrong answer.

As AI becomes embedded in fund administration, tax and compliance processes, the risk is no longer that a calculation fails. It is that an incorrect calculation appears entirely correct. A missed side letter, an overlooked waterfall provision or a misunderstood tax election may not trigger an obvious exception. The output can look complete, consistent and ready to use.

The question for asset managers is therefore no longer whether AI should be used. It is who is accountable for validating the answer before it reaches investors, regulators or auditors.

Ocorian's recent Global Asset Monitor found that 99% of alternative asset managers now use AI within their investment processes, yet only 5% have a formal policy governing its use. The gap suggests that AI adoption is advancing much faster than the controls needed to oversee it.
 

When a clean answer conceals an error

Some of the hardest AI errors to spot look entirely plausible. Take a tool that reconciles a fund's NAV or capital account data against its limited partnership agreement. The figures may appear to reconcile perfectly, yet the calculation can still miss a subsequent closing and equalization adjustment, apply the wrong catch-up tier in a waterfall calculation or overlook a management fee step-down triggered by the end of the investment period. A side letter or MFN provision elsewhere in the documentation may also change the treatment for a particular investor.

If the tool misses that document, links it to the wrong investor, applies the wrong tier or interprets a provision incorrectly, it can still produce a polished and consistent result. The calculation may be internally accurate based on the information it has seen. The final answer, however, can still be wrong.

The same risk exists in tax and compliance. An AI tool may produce a technically sound result based on general tax principles but miss critical facts unique to the engagement, such as a Section 754 election, disguised sale considerations, UBTI implications for tax-exempt investors, state-specific modifications, or complex allocation provisions in the partnership agreement. The resulting error may only surface after it has flowed into capital accounts, investor reporting, tax returns, or regulatory submissions.

That is the problem with 95% accuracy in fund operations: the missing 5% can be the part that changes the answer. AI's greatest operational risk is often not inaccuracy but the illusion of certainty. When an answer looks complete, scrutiny naturally falls away.
 

The experience gap

AI has exposed a gap between technical skill and subject-matter experience. The people closest to the tools often know how to set up systems and write prompts. However, they may have less experience of the accounting treatments, fund documents and regulatory duties behind the result.

Senior professionals face a different issue. They can spot a questionable assumption but cannot always see how the system reached its answer. They may not know which documents it reviewed or what fell outside its scope.

A sound review needs both perspectives. Reviewers must be able to see the source material and instructions behind the output. They must also have the professional experience to challenge the conclusion.
 

Responsibility across the operating chain

AI-assisted analysis often passes through several hands before it reaches an investor. A tax adviser develops an interpretation or filing position, which is then reflected in the fund's books and records by the administrator and subsequently relied upon by the GP for investor reporting, financial reporting, tax filings, and regulatory submissions. A parallel risk sits on the operational side: an AI tool used for a capital call or distribution calculation can reconcile cleanly against internal figures while missing a timing or notice requirement set out in the credit facility agreement or the LPA itself.

Problems arise when each party assumes that the prior participant validated the AI-generated output rather than independently reviewing the underlying analysis. If the original analysis is wrong or incomplete, the error can pass through the chain without being questioned.

Each party therefore needs a defined role. The adviser producing the analysis must stand behind its quality, while the administrator using it must be satisfied that it fits the fund’s records and governing documents.

Fund administrators are well placed to provide this check. They work closely with the fund’s accounts, investor data and key documents. This helps them spot when an external conclusion does not match what they know about the fund. Specialist tax, legal or compliance advice remains with the relevant adviser. The fund administrator provides an additional control before AI-assisted analysis enters fund records or passes to investors, checking that the answer makes sense and that nothing material has been missed.
 

Building a review process that works

For fund administrators, providing this check requires a clear review process.

The level of review should reflect the risk attached to the output – using AI to summarize an internal meeting carries less risk than using it to support a NAV calculation, for instance. The risk is also higher when AI is used to interpret an investor entitlement or prepare a regulatory submission.

For this higher-risk work, the review should begin with the information the AI has used. In fund administration, this means checking that the relevant LPA, amendments, side letters, MFN elections and investor records are complete and up to date.

An experienced professional must then test the output. This may involve recalculating a material figure, tracing an investor treatment back to the relevant provision or checking how the system handled a known exception.

Firms should record what was reviewed, who reviewed it and which checks were carried out. If reviewers find an error, the firm must trace its cause, check whether other outputs are affected and decide whether the review process needs to change.
 

Making responsibilities clear

Once the review process has been defined, the roles within it must be set out clearly. A formal AI policy should specify which tools are approved and what data can be entered. It should also explain which tasks AI may support, and the review needed before an output is used. Compliance and regulatory reporting remain the area where AI is used most, and where formal policies remain rare. This is where a defined review process, covering filings such as Form PF and Annex IV, has the most immediate impact.

The policy should cover work that passes from one party to another. If a tax adviser's analysis is incorporated into the books and records, the manager must know who reviewed the underlying work and what validation procedures were performed. Without clear accountability, there is a risk that each participant in the reporting chain assumes someone else has already verified the analysis. It must also be clear who compared it with the fund’s own information and who approved the final output.

This gives tax advisers, fund administrators and GPs a clear view of their respective roles. It reduces the risk that each party assumes someone else checked the work before it enters a calculation, filing or investor communication.
 

Judgement as the differentiator

AI is already widely used across fund management and administration, including by administrators themselves in reconciliation and document intake. That proximity to the data raises rather than lowers the case for a defined review layer. Firms will increasingly stand out through the quality of their reviews and the way they assign responsibility, particularly when AI-assisted work affects fund records, regulatory obligations or investor outcomes.

AI is changing who performs the analysis. It is not changing who is accountable for the outcome.

As the technology becomes more sophisticated, the differentiator will not be which firms use AI. Nearly everyone will. The differentiator will be which firms can demonstrate that important decisions, calculations and filings were subjected to informed human challenge before they were relied upon.

In an industry built on trust, judgement is becoming the most valuable control.