I generally prefer not to write articles that are too closely tied to current events. Articles tend to last longer when they are not anchored to a particular moment in time.

This time, however, there was something I wanted to preserve.

Since re-establishing a fund administration business, I have occasionally organised seminars with the rather presumptuous intention of introducing interesting topics and interesting professionals in the fund industry to a wider audience.

The seminar that prompted this article was about generative AI.

The actual content of the seminar belongs to those who attended. We recorded the sessions as well, but the recordings are available only to those who registered.

So I am not going to reproduce what the speakers shared.

What I do want to explore is a question that emerged from the discussion and that, I think, matters well beyond AI.

Let us start somewhere slightly unexpected.

Let Me Ask Again: What Is the Business of Running a Fund?

What does fund operations actually do?

And as I often say, I mean fund operations, not investment management, although the same way of thinking could equally be applied to the investment process.

When someone first joins a company without knowing their way around, the easiest thing to do is whatever the manager says, or whatever has always been done.

That is not unique to funds.

Replace “manager” with “investor”, or even “CEO”, and the pattern remains the same.

The person keeping you busy may simply be doing something because someone else told that person to do it.

So perhaps nobody is particularly different from anyone else.

But is that really what a company, an organisation, or a fund back office is supposed to be?

Just doing what we are told?

If nobody, from the CEO to the most junior member of staff, stops to ask that question, the discussion ends rather quickly:

“Well, that’s just how fund operations works.”

Thank you very much. We can all go home.

But let us instead step back and ask what a fund is trying to achieve, and therefore what actually needs to happen.

At its most respectable, a fund invests capital to increase the value of assets and thereby contribute to society.

Underneath that respectable description, of course, people may also want to buy low, sell high, make money and perhaps earn enough carried interest to retire early.

Let us not say that part too loudly.

Don’t worry. Whatever you say, everyone probably knows what is going on in the back of your mind anyway.

In order to invest, the fund first needs investment opportunities.

On the other side, it needs capital.

That capital can come from investors, banks providing acquisition or bridge financing, private credit funds and other providers of leverage.

The investment strategy, the underlying companies or real estate, and whatever else is available must be explained convincingly enough for those capital providers to commit their money.

When the investment is ready, capital is drawn or borrowed and used to acquire the investment.

We therefore need a neutral vehicle to connect those who provide capital with those who need it.

That vehicle is the fund.

By the time the first investment is made, the vehicle has been established, capital has been raised and money has gone into an investment.

At that point, it certainly feels like:

“Ah. We are running a fund now.”

Unfortunately, that is only the entrance.

Investors and lenders are waiting to get their money back.

The manager therefore needs to report on the progress of investments and explain how eventual repayment or distributions may occur.

The portfolio company or asset needs active work to increase its value.

Potential buyers need to be identified.

Exit negotiations need to happen.

Eventually, if things go reasonably well, an investment is sold, cash is recovered, capital is returned to investors and lenders, and the investment reaches its end.

Only when all investments have completed that journey can the fund finally fulfil its purpose, wind itself up and end its life.

Now step back and look at that entire picture.

What actually exists inside a fund?

There are contracts defining obligations between investors and the fund: commitments, capital calls, reporting obligations and distributions.

Those contracts cause movements of cash.

On the asset side, there are subscription agreements, purchase agreements and eventual sale agreements.

Those also cause movements of cash.

Viewed purely as economic activity, much of a fund therefore consists of evidence-backed transactions executed according to agreed terms.

Those transactions are recorded through accounting.

The accumulated results become financial statements.

Auditors then examine those statements against supporting evidence and applicable accounting principles.

None of this is particularly exotic.

Businesses have been doing versions of it for centuries.

I once experienced an interesting example myself in the audit of a Cayman fund.

As part of discussions surrounding management representations, I was asked about the use of generative AI and ended up discussing the reasoning behind the question with the Cayman auditor.

AI usage has changed considerably since then, so I would not assume that the same wording or confirmation is still requested today.

But the underlying issue remains important.

The real question is not whether AI happened to be used.

The question is whether the facts underlying the financial statements and the process that produced them can be explained, and whether somebody ultimately accepts responsibility for the output.

---

Does Fund Operations Have a Way of “Tidying Up” Information?

So why do we expect generative AI to make operational processes more efficient?

Perhaps because a fund is, among other things, an information business.

Again and again, we:

obtain information, organise it, and use it.

And at every stage there are tedious, expensive and time-consuming tasks.

Take a transaction agreement.

Buried inside pages of unstructured language may be the information required to settle an investment:

• payment amount;
• payment date;
• recipient;
• settlement instructions.

One could, in principle, simply send the entire agreement to the investment committee.

If the committee asks when the money is required and how much is needed, one could say:

“It’s in the agreement. Go and find it yourself.”

After all, I had to struggle to find it. Why shouldn’t everyone else?

This is technically possible.

It is also a wonderful way to make an investment committee slightly less functional.

At least put a sticky note on the relevant page.

Once we think beyond the immediate decision, the same information becomes much more valuable.

Suppose the fund needs to call capital not only for the investment price, but also for transaction expenses.

Invoices arrive from lawyers and accountants for due diligence and transaction work.

If the relevant information from all those documents is extracted into the same structured format:

• when;
• payable to whom;
• how much;
• for what purpose;

then even something as simple as an Excel sheet can become the basis for multiple subsequent processes.

The information can support:

• the investment committee summary;
• the capital call notice;
• future cash-flow forecasting;
• payment instructions for online banking;
• accounting;
• later investor reporting;
• eventually, performance calculations based on cash flows.

A small amount of organisation today can reduce quite a lot of work for your future self and everyone else.

Possibly even some overtime.

The difficulty is that these activities have traditionally depended heavily on people.

Someone must read the document.

Someone must understand the language.

Someone must imagine how the information may be needed later.

Consider something slightly less familiar.

Suppose someone suddenly hands a Japanese fund professional an invoice written in Swahili.

How many people could immediately understand it and process the payment correctly?

And even when everyone can read the language, another problem remains.

One person stores a document here.

Another stores a copy somewhere else.

So where is the information?

Which version is correct?

If both are correct, where did each copy come from?

We trace where our wagyu beef, tomatoes and other food came from.

Yet in fund operations, we often struggle to trace where a number or piece of information came from.

Interesting, isn’t it?

The way people tidy up a home reflects the shape of the home, the habits of the family and, perhaps, the family’s history.

If nobody famous has said that before, I shall claim to be the first.

The same idea applies to organisations.

If information management follows the personal habits of individual employees, the system may make perfect sense to the person who created it and absolutely no sense to anybody else.

If that person leaves, everyone else starts wandering around like a confused robot vacuum in a house whose layout it never learned.

Multiply that behaviour across several employees and the result becomes even more interesting.

Multiple copies of identical information occupy limited corporate storage.

The copies quietly undermine one another’s credibility:

“Are you the real version?”

Two corporate doppelgängers staring silently at each other.

There is also another temptation.

Information asymmetry can create the illusion of power.

Someone may believe that possessing information nobody else has creates personal importance.

So that person keeps important information privately, shares it with nobody, and quietly enjoys being “the person who knows”.

Usually, nothing particularly dramatic follows.

The information ages.

Its usefulness declines.

Its supposed strategic advantage disappears.

Eventually somebody asks:

“What was this again?”

Information tends to create more value when it is shared with the people who are entitled to use it.

The broader lesson is simple.

When information is obtained, stored and used based only on local optimisation, the organisation can easily end up very far away from overall optimisation.

---

Will Generative AI Solve This?

Perhaps AI can help with parts of the problem.

AI can be extremely useful when obtaining and extracting information.

A PDF may technically be a PDF while containing nothing more than scanned images.

AI-enabled tools can read the images, convert them into usable text and extract structured information far faster than a person could manually retype everything.

Return to the Swahili invoice.

Today it is entirely conceivable that a tool can identify:

• the amount;
• the due date;
• bank details;
• the nature of the invoice;

and translate the contents into a language the user understands.

That is genuinely useful.

It can produce very significant efficiency at an individual level.

It can also make us feel rather competent.

But if the process ends with:

“Great, I understood the invoice.”

then we have solved only a personal problem.

A fund is not run by one person.

Fund investing involves a wider operating ecosystem that may include a GP sponsor, directors of the GP vehicle, corporate secretarial functions, administrators, auditors and others.

Knowing something personally, even as the CIO or CEO, does not move the fund forward.

So what?

The information needs to proceed to the next step.

Is the invoice actually related to the fund?

Has it already been paid?

Is the payment authorised?

What accounting treatment is required?

What information must be retained?

Who else will need the data later?

Once those questions are considered, we arrive at a different design problem.

How do we make the information available as a trusted Single Source of Truth (SSoT) so that subsequent processes can use it consistently?

And perhaps that concept eventually needs to extend beyond the boundary of the organisation as well.

Why?

Because everyone in the operating chain has versions of the same problem.

---

Is This Really an AI Problem?

Probably not.

Before asking what AI should do, we may first need to look across the entire workflow.

How is information obtained?

How will it be used later?

How should it be structured?

Where should the authoritative version be stored?

Which unnecessary movements and copies can be removed?

In other words, perhaps the problem is much closer to something we learned at nursery:

Decide where things belong, use them, and put them back where everyone knows where to find them.

Only after that do we ask:

• What should people do?
• What should traditional software or macros do automatically?
• Where might AI genuinely help?

And one principle matters throughout:

Facts must remain facts.

Auditors and investors do not want a plausible-sounding story that nobody can trace back to evidence.

Financial statements need support. Investor reports also need support.

Capital call notices requesting investors’ money need support.

Distribution notices returning investors’ capital need support.

IRR, TVPI and DPI calculations need reliable underlying cash flows.

Carried interest allocations generated through waterfall calculations need reliable source data and reproducible logic.

If AI can generate correct code implementing a waterfall calculation, wonderful.

Once the calculation logic has been turned into deterministic code, however, the calculation itself does not need a generative AI model to run it.

That is a feature, not a problem.

We can use AI where AI adds value while removing worries about probabilistic language generation from processes that require accuracy, consistency and reproducibility.

Human review does not disappear either.

We have always needed to review work against evidence.

At a minimum, we still need the equivalent of the sticky note saying:

“Look here. This is the evidence.”

Whether the work was performed by a person or by AI does not change the organisation’s responsibility for the output.

We would not normally accept:

“The member of staff made a mistake.”

as a sufficient explanation.

So why would:

“AI produced it.”

be any better?

Optimising processes therefore creates two separate responsibilities.

First, someone must decide what should be done by people, by conventional systems and by AI.

Second, the organisation needs controls allowing it to accept responsibility for the result regardless of who, or what, produced it.

---

Now, Let Us Start Again from Nothing

Consider a thought experiment.

Suppose you are starting a company from scratch.

Congratulations. You are finally the boss.

What exists inside the company on the first morning?

Almost nothing.

For better or worse, you have the most troublesome form of freedom imaginable: nothing has been decided yet.

Unfortunately, most people who join you will assume that rules exist.

Even in a tiny company, employees often expect policies, procedures and consistent treatment similar to those provided by much larger organisations.

Someone will eventually say:

“But my previous company did it this way.”

Of course they will.

You started from zero.

So what should you do?

Decide which principles actually matter to the organisation.

Define rules for the outputs you need.

Select tools that support those rules.

Allow people to accumulate experience by repeatedly producing reliable, reproducible outputs.

That repetition produces business consistency.

Consistency supports stable performance.

And, over time, both the company and its people can grow.

Funds are no different.

A fund manager should not design an operating model merely because the Cayman regulator issued a governance code.

Nor because a regulator asks about operational arrangements.

Nor because an investor asks a question during operational due diligence.

Those external requirements matter.

But good operations should exist because the organisation itself has decided how it intends to operate.

The fund should have a rulebook.

Inputs such as the investment strategy, underlying investments and investor requirements should be understood.

Outputs should be defined.

Processes should be operated consistently.

Where something does not work, the model should be adjusted.

Investments, capital movements, accounting, reporting and all the other economic activities around the fund then operate within that framework.

Information is obtained.

It is stored so that people authorised to access it can use and transform it.

It is shared appropriately.

And only then does the AI question become particularly useful.

If the same final output must be produced regardless of whether a person performs a step, conventional software performs it, or AI performs part of it, then perhaps the best design is the one reached by working backwards:

Which combination produces the required outcome most efficiently, most consistently and with the lowest likelihood of error?

Ultimately, that choice says something about the culture of the organisation.

Are we building a business as a team?

Or are we optimising for each individual to work in whichever way happens to suit that person best?

---

Wait. Wasn’t This Supposed to Be an Article About an AI Seminar?

Yes.

The seminar was about generative AI.

But the question I brought home was not really about AI.

It was this:

How well do we actually understand our own work?

In an odd way, it felt like returning to something we learnt at nursery.

Generative AI may automate work.

Before it can do so effectively, however, we need to be able to explain the work.

What is fact?

What is judgement?

What is explanation?

What is coordination?

Organisations that can make those distinctions will be better placed to use AI.

They will also be better placed to let people do what people do best.

And perhaps, more importantly, to give people room to think and grow.

So maybe the first step in adopting AI is not adopting another AI tool.

It is something much less fashionable:

Understand your own work well enough to explain it.

AI can come afterwards.