AI in Fund Administration: The Right Use Cases, Risks, and Questions to Ask

By now, most fund administrators have integrated AI into their workflows. But before you automatically assume AI functionality will benefit your fund, it’s important to take a step back to understand why and where AI was implemented within the fund administration operational structure. And, maybe even more importantly, how? If AI integration was rushed primarily for competitive reasons and the marketing power behind the ‘AI-enabled’ buzzword, there may be hidden risks to your fund.

First, Let’s Look at How AI Works

Modern GenAI large language models (LLMs) are built on pattern recognition. We anthropomorphize LLMs and assign them an unrealistic level of reasoning, but they actually work in predictable ways. They study data, extrapolate what should come next in a sequence, and apply it to future situations. However, if something occurs outside of the known dataset, LLMs will use existing patterns to hallucinate a solution that may be illogical, inefficient, or just plain wrong for the assignment at hand.

Unlike humans, who learn deductively by internalizing a set of rules and solving problems based on those rules, AI learns inductively by studying problem sets and building rules from that data. Human intelligence is generally defined by how efficiently we learn new skills. Contrast that with AI, which needs to be fed millions upon millions of data sets (more than a human being could ever possibly absorb) to master a skill like playing chess.

Harvard Business Review studied the impact of AI usage on mental fatigue and burnout. It found that if how AI is being used requires a lot of oversight, then mental effort is increased, not reduced, as paradoxical as it seems. The more workers use it, the more stress it causes. This can lead to higher decision fatigue, poorer decision making, and increased employee turnover.

On the other hand, if AI is used to take over repetitive tasks, it reduces mental load. For example, using an AI system to automatically update routine documentation is a way to save employees from manually inputting repetitive information into the system.

Best Uses for AI in Fund Administration

In fund administration, it’s important to find the correct use cases for AI. Certain work, especially in financial services, is deterministic. If you write rules or formulas, it will always result in the same answer. AI does not work like that. It will often generate different answers for the same question. AI is also probably not the best choice in situations where something must be learned quickly and efficiently. It works better when applied to situations where statistically significant outliers aren’t relevant, such as summarizing report data for publication to manager and investor portals.

For example, calculating NAVs is not a task that should be outsourced to AI. Why? Because established formulas for fee computations, allocations, and profit and loss work are perfectly reliable and are easy to write, understand, and debug. These formulas can also adapt to situations that happen infrequently, whereas AI might not know how to handle outliers.

There can be vastly different outcomes based on the type of work offloaded to AI. Ideally, your fund administrator deploys AI to free people from monotonous work so they can focus on judgement, relationships, and accountability — things that AI cannot.

What to Look for in Fund Administrator AI Adoption

• AI goals: Your administrator should have clear, value-driven goals when implementing an AI system, and these initiatives should produce noticeable benefits for you as a client. An administrator should take their time with implementation, not rushing to slap AI processes on functions that are already working well.
• Human accountability: How accurate is the information provided by your admin’s AI, and will they take full accountability if it’s wrong? If a chatbot provides incorrect data, your administrator may claim it’s ultimately your responsibility to review it. That means your administrator expects you to act as an oversight agent for AI-produced data. If your provider won’t take full accountability for the veracity of its AI-produced outputs, you may need to ask yourself how useful it is if it requires constant review.
• AI literacy: You want the team working on an AI-enabled project to understand how it works. Employees should be able to answer detailed questions about the process, understand why AI was introduced, and demonstrate how it’s improved the process.
• Decision-making: Decisions should be made by people, not AI. AI is ideal for automated repetitive tasks, not creative work or work involving a lot of review and oversight, and definitely not for making critical decisions. Make sure your administrator understands the distinction.
• Machine learning versus LLMs: How sophisticated are the AI processes being incorporated by your service provider? Is your administrator creating custom models, using robotic process automation, or incorporating machine learning-backed pattern recognition? Or, are they bolting on existing LLMs like Claude or ChatGPT and reselling them under their own branding?

Potential Risks to Clients with Fund Administrator AI Adoption

• Cybersecurity: AI-generated code increases the surface area of attacks, especially if engineers are not as familiar with how the code works. This is even more critical if the person writing the code is “vibe coding” and does not have a technical background. When Meta’s management pushed the use of AI coding earlier this year, the company saw a 40% spike in service disruptions and potential data leaks over the year before, and a 70% spike in the time employees spent on damage control according to Reuters.
• Accuracy: Hallucinations are a real concern with AI. Since AI functions by finding patterns in data and extrapolating what’s next in a sequence, hallucinations are currently a known consequence of how the technology works and may continue to be a major problem for some time into the future. However, if AI hallucinations cause a fund administrator to produce inaccurate NAV calculations even 1% of the time, that would be a major problem for you and your investors. Your fund administrator should have built-in human checks and backups to ensure hallucinations don’t make their way into your monthly reporting.
• Downstream decision-making risk: A related point is the risk of incorrect data muddling your decision-making process. If a staff member asks your fund admin’s chatbot a question, and the chatbot provides incorrect information, that may inadvertently be used as a building block in your decision-making. You have to trust the quality of your admin’s outputs, and if those outputs are wrong, then be ready for the risk that your downstream decision-making may be affected by faulty data. Otherwise, you’ll have to be prepared to verify every piece of data produced by AI.
• Data confidentiality: If the administrator is using a third-party vendor for AI implementation, they must ensure that your investors’ private data is properly protected. Some third-party vendors retain data for training purposes, so you may need to explicitly spell out in your contract that this data will be protected. You might also ask if the administrator has conducted a data privacy and security audit on the third-party vendor.
• Cost: Is your administrator asking you to absorb the cost of third-party vendors or upgraded technology? For example, AI output is verbose, so larger codebases require more future maintenance. At Meta, coders increased the amount of code produced by 220% when using AI, but its actual features only increased by 36%. This is a hidden cost you could be absorbing without realizing it. Another question: how does usage work? LLMs often now charge by usage as well, so ask your vendor if you’ll have to absorb the increase if token costs rise.
• Accountability: If AI is being used to replace previously human-led decision making or other important processes, the people currently in charge should understand how the underlying AI tech works. If there’s a gap or issue with the AI results, it should be clear who’s in charge of the process and how the error will be resolved.

Conclusion

AI is impressive technology, but despite the current hype it’s not a magic bullet that should be implemented everywhere. Some parts of fund administration, like NAV calculations, don’t lend themselves to AI processes. Other parts, like document extraction, can be improved and automated by AI. The key, when assessing how it’s been implemented by your fund administrator, is to make sure that the benefits are clearly defined, you understand the inherent risks you’re taking on, and you know who’s accountable when something goes wrong.

It’s clear that if incorporated well, AI can be an incredibly useful tool. Generally, companies who adopt something new in a thoughtful, considered way can avoid the risks and losses of those who rush in for no other reason than to say they were ‘first.’

At NAV Fund Services, we have built custom proprietary models to enhance certain data and document extraction processes. AI is not used in automated decision-making, and its results are always reviewed by a human. We are focused on ensuring data privacy and security and obtain enterprise-grade licenses when available. When using AI, accuracy is paramount, as generative AI still suffers from a degree of fact hallucination. To address these concerns, we have constructed strict validation processes on all AI-enabled functions.